wildminder/awesome-ltx2

wildminder/awesome-ltx2 has 602 GitHub stars and 43 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-12 23:18 UTC. Last push 2026-11-04.

Cómo se miden estas cifras

All available LTX-2 models, encoders, workflows, LoRAs for ComfyUI

602 estrellas 43 forks

Crecimiento de estrellas medido

in 24h: 0
in 7 days: Recopilando historial
in 15 days: Recopilando historial
in 1 month: Recopilando historial

Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.

README.md

Original

Awesome LTX-2

A curated list of models, text encoders, and tools for the LTX-2 video generation suite.

ltx-logo

Table of Contents

Intro

▓ Apps & Tools

LTX2.3-Multifunctional

LTX2.3-Multifunctional is a desktop-optimized version of LTX that lowers GPU requirements and simplifies usage. It integrates all features including image-to-video, text-to-video, start/end frames, lip-sync, video enhancement, and image generation into a single application.

Key Features:

Lower GPU Requirements: Only needs 24GB VRAM (vs 32GB for standard desktop version)
All-in-One Interface: No complex ComfyUI workflows or error-prone nodes
Features: T2V, I2V, start/end frames, lip-sync, video enhancement, image generation, LoRA support
Multi-Frame Insertion: Two modes for generating long videos
Easy Setup: No third-party software required, just install LTX desktop

Downloads & Resources:

HuggingFace | GitHub | ComfyUI Node | Tutorial

O2noor LTX-2.5 Int4 Tile-Train (Beta)

O2noor LTX-2.5 Int4 Tile-Train is a ComfyUI node pack for LoRA training on LTX-2.5 22B distilled using tiled training across multiple GPUs with as little as 18 GB VRAM per card. The base is fully self-contained quantized weights (bnb-NF4 DiT, Gemma-4-12b text encoder — run in 8-bit LLM.int8 spread over 2 GPUs during captioning only, freed for training; works on 12 GB cards).

Downloads & Resources:

Node pack (GitHub) | Model weights (HF) — includes ltx-2.5-22b-distilled-bnb-nf4 (10.45 GB), embeddings_processor_bf16 (6.34 GB), gemma4-12b-with-proj-ltx-2.5-bf16 (26.26 GB), video + audio VAEs, plus int2 experimental variants

elismasilva LTX 2 Image Custom Blocks (Modular Diffusers)

Custom Modular Diffusers blocks that extend LTX 2 Image (elismasilva's image-only LTX-2.3) with image-to-image, plus a unified AutoBlocks pipeline that folds t2i and i2i into one — the workflow is chosen automatically from the inputs you pass (prompt → text-to-image; prompt + image → image-to-image with optional strength). Code-only repo (trust_remote_code); loads components from the ltx2.3-image-base weights repo. Apache-2.0.

Downloads & Resources:

Custom blocks (HF) | Image-only checkpoints: ltx2.3-image-comfyui (ComfyUI) · ltx2.3-image-base (diffusers)

WanGP LTX-2 Model Pack (DeepBeepMeep)

The complete set of LTX-2 / 2.3 / 2.5 weights used by WanGP — DeepBeepMeep's low-VRAM video app (down to ~6 GB VRAM, old-GPU friendly, auto-downloads the model variant matching your architecture). 169 files, ~984 GB total, pre-packaged so no manual ComfyUI file layout is needed.

What's inside:

Main transformers (repo root) — LTX-2.5 22B dev / distilled in bf16 (38.0 GB), int8-convrot (19.5 GB) and nvfp4 (14.7 GB); LTX-2.3 22B dev / distilled / distilled-1.1 bf16 (38.0 GB) plus quanto bf16-int8 (19.5 GB) and nvfp4 (13.5 GB); LTX-2 19B dev / distilled full (43.3 GB), fp8 (27.1 GB), fp4 (20.0 GB) and diffusion-model variants; Q4_K_M / Q6_K / Q8_0 "light" GGUFs (13.0–20.6 GB)
Third-party audio models — JoyAI-Echo (bf16 + quanto), Scenema, DramaBox, ltx23_echoVid-ltxAud_surgical_fp8, plus Kokoro TTS, Seed-VC, Whisper, HuBERT, BigVGAN, Sherpa
Shared / offloadable components — video + audio embeddings connectors (bf16 / int8-convrot / nvfp4), text embedding projection, video + audio VAEs, vocoder, spatial and temporal upscalers x2
Text encoders — both gemma4-12b-ltx-v1 for LTX-2.5 (bf16 23.8 GB / int8-convrot 12.9 GB) and gemma-3-12b-it-qat-q4_0 for LTX-2 / 2.3 (24.4 GB + quanto 13.2 GB)
LoRAs — LTX-2.5 IC-LoRAs (refine-details, ingredients, SDR-To-HDR + its scene-emb, deblur, decompression, pixel-spatial-upscaler), LTX-2.3 IC-LoRAs (ingredients, in-outpainting, outpaint, refocus, uncompress, ungrade, HDR + scene-emb, union-control, pixel-spatial-upscaler, detailer), LTX-2 IC-LoRAs (detailer, union-control, canny/depth/pose control), distilled-lora-450 / -384, celebvhq ID-LoRAs (LTX-2 and 2.3), Edit-Anything reference, Licon-MSR (2.3 V1/V2, 2.5 V1 + slot embeddings), VBVR-I2V, OmniNFT RL-LoRAs
Manifest — LTX-2.5-MANIFEST.md documents the LTX-2.5 runtime set, incl. the note that Dev/Distilled connectors are shared only after equality verification and the official NVFP4 transformer needs its own BF16 connector pair

Downloads & Resources:

WanGP (GitHub) | Model pack (HF)

LTX-2 Marking (VideoSeal / PixelSeal / AudioSeal)

The weights and configs for the opt-in, post-hoc invisible marking used by LTX-2: PixelSeal (VideoSeal family) for video frames and AudioSeal for stereo PCM. MIT licensed, pytorch library. This is not a generator model - there is nothing to run or train against it, only to download.

File	Purpose	Size
`pixelseal.pth`	PixelSeal embedder + extractor weights (256-bit payload, 256 px processing size)	308.95 MB
`audioseal_generator.pth`	AudioSeal 16-bit generator (embed)	58.81 MB
`audioseal_detector.pth`	AudioSeal 16-bit detector	34.67 MB

How it fits together - the LTX-2 loader (create_markers) requires all five model files (the three above plus pixelseal.yaml, the PixelSeal model card describing the architecture, and videoseal_attenuation.yaml, the JND attenuation channel map) in one directory; the code that embeds and detects lives in packages/ltx-pipelines/src/ltx_pipelines/marking in the LTX-2 repo. The .pth files are PyTorch pickles - verify the SHA-256 sums printed on the model card after download.

Enable it with --marker-root-path path/to/marking on the generation, editing and text-to-audio CLIs; omit the flag to leave pixels and PCM unchanged. Video generation and the HDR paths mark video while the text-to-audio path marks only audio, and detection returns MarkDetectResult(detected, bit_accuracy, p_value, payload) - video marks are judged on every 8th decoded frame at log10(p_value) < -6, audio over the full stereo track.

◦ The marks are a robust invisible signal, not a cryptographic signature: a detected mark means the clip came from an LTX-2 pipeline with marking enabled, while the absence of a mark does not prove a clip was not generated.

▓ Models

LTX-2 models are available in various formats including full weights, transformers-only, and GGUF quantizations for efficient inference.

▣ Checkpoints

Lightricks/LTX-2 - Official repository.
Lightricks/LTX-2.3 - Official repository (latest version).
Drbaph - Quantization
unsloth - Quantized dev / distilled DiTs (FP8, INT8, INT8-ConvRot) + GGUF quants

Ver	Name	Size
2.5	`dev`	42.02 GB
2.5	`dev`	21.50 GB
2.5	`distilled`	42.02 GB
2.5	`distilled`	21.50 GB
2.5	`distilled`	18.72 GB
2.5	`pt` (pre-trained)	46.2 GB
2.5	`distilled`	22.1 GB
2.5	`dev`	12.52 GB
2.5	`distilled`	12.52 GB
2.5	`distilled`	21.03 GB
2.5	`distilled`	18.7 GB
2.5	`dev`	15.4 GB
2.5	`distilled`	15.4 GB
2.5	`distilled`	23.5 GB
2.5	`dev`	21.64 GB
2.5	`dev`	13.57 GB
2.5	`distilled`	21.64 GB
2.5	`distilled`	13.57 GB
2.5	`distilled`	25.11 GB
2.5	`distilled`	19.06 GB
2.5	`distilled`	25.11 GB

2.3	`dev`	46.1 GB
2.3	`dev`	29.1 GB
2.3	`dev`	29.9 GB
2.3	`dev`	29.2 GB
2.3	`dev`	21.7 GB
2.3	`dev`	29.1 GB
2.3	`distilled`	46.1 GB
2.3	`distilled`	29.5 GB
2.3	`distilled`	29.9 GB
2.3	`distilled`	29.1 GB
2.3	`distilled`	17.6 GB
2.3	`distilled`	29.7 GB
2.3	`distilled 1.1`	46.1 GB
2.3	`distilled 1.1`	16.65 GB
2.3	`distilled 1.1`	15.37 GB
2.3	`ltx23_srx fp8_e4m3 experimental`	23.1 GB
2.3	`distilled`	19.06 GB
2.3	`distilled`	19.32 GB

2	`ltx-2-19b dev`	43.3 GB
2	`ltx-2-19b dev`	27.1 GB
2	`ltx-2-19b dev`	20 GB
2	`ltx-2-19b distilled`	43.3 GB
2	`ltx-2-19b distilled`	27.1 GB
2	`ltx-2-19b distilled`	20 GB
2	`ltx-2-19b dev`	18.91 GB
2	`ltx-2-19b dev`	19.09 GB

◦ unsloth quant rows — pre-quantized DiTs built from the diffusers-format releases (torchao layout; loaded strictly by diffusers / Unsloth Studio — not ComfyUI single-file diffusion_models/ weights). Each file is bit-identical to quantizing the dense bf16 weights on the fly, so it skips the dense load + quantize step. Repos: LTX-2-FP8 (dev + pre-cast fp8 text encoder) · LTX-2.3-FP8 (distilled + bf16 originals) · LTX-2.5-FP8 (distilled; INT8-ConvRot recommended).

Quantized to fp8_e5m2 to support older Triton with older Pytorch on 30 series GPUs. For WangGP in Pinokio

Ver	Name	Precision	Size	Download
2	`ltx-2-19b dev`		27.1 GB

· · · · · · · · · · · · · ·

❖ Image-only LTX-2.3 (text-to-image)

Image-only LTX-2.3 checkpoints by elismasilva — the 13B LTX-2.3 family pruned to the still-image generation path (video/audio weights removed), packaged as unified ComfyUI safetensors. Inherits the LTX Video 2 Open Source License.

Ver	Name	Size
2.3	`image dev`	33.61 GB
2.3	`image distilled`	33.61 GB
2.3	`image dev`	21.81 GB
2.3	`image distilled`	21.81 GB

· · · · · · · · · · · · · ·

❖ silveroxides Quantizations (mxfp8)

Note: The mxfp8mixed quantization requires a custom fork of ComfyUI-Kitchen with mxfp8 support. Standard ComfyUI installations may not support this quantization format.

Model	Quant	Size	Download
`ltx-2.3-22b-dev`		29.2 GB
`ltx-2.3-22b-distilled`		29.1 GB
`ltx-2.3-22b-distilled`		29.2 GB
`ltx-2.3-22b-distilled`		29.7 GB

· · · · · · · · · · · · · ·

❖ Distilled LoRA

Ver	Rank	Size	Download
2.5	`450`	8.90 GB
2.5	`450`	5.06 GB
2.5	`450`	2.66 GB
2.5	`450`	8.90 GB
2.5	`384`	7.61 GB
2.5	`256`	5.10 GB
2.5	`256`	5.10 GB
2.5	`128`	2.58 GB
2.5	`64`	1.32 GB
2.5	`128`	2.31 GB
2.5	`72`	1.38 GB
2.5	`256`	4.01 GB
2.5	`96`	1.80 GB
2.5	`32`	0.66 GB

2.3	`384`	7.61 GB	┊
2.3	`208`	4.97 GB
2.3	`159`	3.83 GB
2.3	`111`	2.74 GB	┊
2.3	`105`	2.59 GB

2	`384`	7.67 GB
2	`242`	4.88 GB
2	`175`	3.58 GB
2	`175`	1.79 GB

◦ Rank-reduced 2.5 builds - the canonical source is the official Lightricks/LTX-2.5/loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors (8.90 GB, rank 450). The ibyteohdear reduced variants rerank that file with a factored SVD (rerank_method: factored_svd, source lora_rank: 450) and then cap the result at 32 / 96 / 256 - the ceil suffix is that ceiling and fro90 the singular-value floor (same scheme as pyros-vault).

· · · · · · · · · · · · · ·

❖ TenStrip Distilled LoRA Experiments

Experimental distilled LoRAs optimized for finetunes and I2V workflows. These LoRAs avoid the issues of the massive rank 384 official LoRA which can be counterproductive with conditioned inputs and finetunes.

Name	Rank	Mode	Size
`distilled v1.1`	`32`	condsafe	380 MB
`distilled v1.1`	`36`	—	739 MB
`distilled v1.1`	`52`	condsafe	487 MB
`distilled v1.1`	`72`	condsafe	662 MB
`distilled`	`72`	—	1.4 GB
`distilled v1.1`	`72`	energy	1.72 GB
`distilled v1.1`	`96`	energy	2.28 GB

Notes:

Lower rank LoRAs (72 and below) can be used at 1.0 strength safely for I2V first pass, with upscale pass at 0.4-0.5 strength
_ceil suffix indicates the dynamic ceiling during reranking
_condsafe suffix indicates cross-attention and other conditioning layers have been zeroed for better I2V compatibility
The official rank 384 LoRA can actively dampen conditioning signals in I2V workflows; cond_safe versions work much better

Download All LoRAs

· · · · · · · · · · · · · ·

❖ Spatial Upscaler

Required for current two-stage pipeline implementations in this repository. Download to COMFYUI_ROOT_FOLDER/models/latent_upscale_models folder.

Ver	Name	Size	Download
2.5	`spatial-upscaler x2 1.0`	995.78 MB	┊
2.3	`spatial-upscaler x2 1.0`	996 MB
2.3	`spatial-upscaler x2 1.1`	996 MB	┊
2.3	`spatial-upscaler x1.5 1.0`	1.09 GB

2	`spatial-upscaler x2 1.0`	995.77 MB

⚠️ The official ltx-2.3-spatial-upscaler-x2-1.0.safetensors was deleted upstream from Lightricks/LTX-2.3; only the ChrisColeTech mirror still carries the 1.0 build. The upstream successor is x2 1.1 (linked above) — the repo now ships x1.5 1.0 and x2 1.1 only.

· · · · · · · · · · · · · ·

❖ Temporal Upscaler

Required for current two-stage pipeline implementations in this repository. Download to COMFYUI_ROOT_FOLDER/models/latent_upscale_models folder.

Ver	Name	Size	Download
2.5	`temporal-upscaler x2 1.0`	0.26 GB	┊
2.3	`temporal-upscaler x2 1.0`	262 MB

2	`temporal-upscaler x2 1.0`	262 MB

══════════════════════════════════

▣ Merges

Custom merged models combining multiple control signals or specialized configurations.

Ver	Name	Description	Download
2.3	`ltx-2.3-22b-distilled-1.1-fused-union-control`	Merged model combining Canny, Depth, and Pose control signals for unified control
2.5	`Alpha-Gen VideoOnly48 NVFP4`	Official Alpha-Gen IC-LoRA merged into the base weights at strength 1.0, then structurally pruned to 48 blocks and quantized to NVFP4 with dynamic activation scales. Ships no LoRA — the matting behaviour is in the checkpoint. Ships a `provenance.json` with base/alpha SHA-256 and a full quality-vs-teacher report.

◦ mskim8584 caveats — read provenance.json before trusting it. The producer's own audit reports "quality_accepted": false and "accepted": false: whole-frame MAE 0.0059, PSNR 30.8 dB, SSIM 0.972, gradient MAE 0.0042, temporal flicker MAE 0.0034, temporal-boundary MAE 0.0416. catastrophic_inversion_frame_count is 0, so nothing blew up, but both the BF16 intermediate and the NVFP4 pass were explicitly marked as not accepted quality. The BF16 merge also required cached_final_empty_prompt_required conditioning and was fused on CPU. Treat as an experiment, not a drop-in replacement for the LoRA.

══════════════════════════════════

▣ Finetunes

Community finetuned models based on LTX-2.3 with specialized improvements and optimizations. Each finetune family may include a backbone checkpoint, low-VRAM component splits, GGUF quants, and merged or extracted LoRAs. Variant cell links go directly to the resolve/main safetensors/gguf file when a single canonical asset covers the row.

❖ DaSiWa

High-performance LoRA-integrated checkpoint family based on LTX 2.3. Includes distilled (4-step) and non-distilled (20-30 step) variants. Recommended sampler: Euler + Simple/Normal/Linear_Quadratic.

Ver	Build	Name	Size
2.3	Distilled	`Treasurechest V1`	21.02 GB
2.3	Distilled	`Solsticecoin V2`	30.13 GB
2.3	Distilled	`Dragonleap V4`	18.38 GB
2.3	Distilled	`Dragonleap V4`	27.64 GB
2.3	Non-Distilled	`GoldenLace V3`	29.16 GB
2.3	Non-Distilled	`GoldenLace V3`	21.73 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	8.51 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	10.6 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	13.33 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	15.9 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	18.63 GB
2.3	Non-Distilled GGUF	`GoldenLace V3`	23.61 GB

DaSiWa extracted LoRA:

Build	Name	LoRA Rank	Size	Download
DMD v2 audio	`LTX2.3_DMD_v2_avgrank86_audio160_L80-D20`	86 (audio 160)	2.27 GB

· · · · · · · · · · · · · ·

❖ 10Eros

I2V-optimised merge using layer scaled merges at different steps. Not a straight weight merge — behaves much nicer than standard LoRA loading and respects prompts.

Ver	Build	Name	Size
2.3	Full	`10Eros v1`	46.14 GB
2.3	Full	`10Eros v1`	29.16 GB
2.3	transformer-only	`10Eros v1`	29.56 GB
2.3	Full	`10Eros v1.2`	46.14 GB
2.3	Full	`10Eros v1.2`	34.33 GB
2.3	Full	`10Eros v1.3`	46.14 GB
2.3	Full	`10Eros v1.3`	29.16 GB
2.3	Full	`10Eros v1.4`	46.14 GB
2.3	Full	`10Eros v1.4`	29.16 GB
2.3	Full	`10Eros v1.4 DMD int8 ConvRot`	29.16 GB
2.3	Full	`10Eros v1.5`	46.14 GB
2.3	Full	`10Eros v1.5 DMD int8`	29.16 GB
2.3	transformer-only	`10Eros v1.5 DMD int8`	25.03 GB
2.3	transformer-only	`10Eros v1.5 int8`	25.03 GB
2.3	Full	`10Eros v1.5 fp8mixed (experimental, learned)`	30.24 GB
2.3	Full	`10Eros v1 INT8 ConvRot`	23.51 GB

◦ The 10Eros Max line is a different architecture — see ❖ 10Eros Max (MiniMax-H3 grafts) below. The gated TenStrip/LTX2.3-10Eros_Version-Testing repo despite its name holds MiniMax-H3 (Hailuo 3) grafts, not LTX-2.3 DiTs. Everything in the table above is a genuine LTX-2.3 AVTransformer3DModel checkpoint (model.diffusion_model + vocoder + vae + audio_vae, 23.07 B params).

◦ 10Eros v1.5 companions — the int8 and fp8mixed v1.5 builds live in third-party repos: CornLogic/10EROS-INT8 (full int8 checkpoint + two transformer-only variants) and LokkenJP/10EROS_1.5_fp8_exp_learned (fp8mixed, experimental, with a .quantization.json + .sha256 sidecar).

◦ 10Eros v1 INT8 ConvRot mirror — the original bertbobson/LTX2.3-10Eros-INT8-ConvRot repo is no longer public (the API returns Invalid username or password and it is absent from the author's model list). The row above points at the public milo01/LTX2.3-10Eros-INT8-ConvRot mirror, which carries the identically-named file at the same byte size. Other public mirrors: Winnougan/10Eros-INT8-Convrot (v1.2), CornLogic/10EROS_1.4_Int8_ConvRot and chfm/10EROS_1.4_Int8_ConvRot (v1.4, both).

◦ 10Eros GGUF — vantagewithai low-VRAM quants

vantagewithai/LTX2.3-10Eros-GGUF

vantagewithai/LTX2.3-10Eros-GGUF — v1

Quant	Size	Download
	11.13 GB
	10.34 GB
	12.98 GB
	13.9 GB
	14.3 GB
	13.2 GB
	15.26 GB
	16.18 GB
	16.14 GB
	15.04 GB
	17.77 GB
	22.76 GB

vantagewithai/LTX2.3-10Eros-1.2-GGUF

vantagewithai/LTX2.3-10Eros-1.2-GGUF — v1.2 — Note: files are named 10Eros_v1.210Eros_v1.2-…gguf (upstream double-stamp glitch); we link verbatim.

Quant	Size	Download
	11.13 GB
	10.34 GB
	12.98 GB
	13.9 GB
	14.3 GB
	13.2 GB
	15.26 GB
	16.18 GB
	16.14 GB
	15.04 GB
	17.77 GB
	22.76 GB

vantagewithai/LTX2.3-10Eros-1.3-GGUF

vantagewithai/LTX2.3-10Eros-1.3-GGUF — v1.3

Quant	Size	Download
	11.13 GB
	10.34 GB
	12.98 GB
	13.9 GB
	14.3 GB
	13.2 GB
	15.26 GB
	16.18 GB
	16.14 GB
	15.04 GB
	17.77 GB
	22.76 GB

vantagewithai/LTX2.3-10Eros-1.4-GGUF

vantagewithai/LTX2.3-10Eros-1.4-GGUF — v1.4 (latest)

Quant	Size	Download
	11.13 GB
	10.34 GB
	12.98 GB
	13.9 GB
	14.3 GB
	13.2 GB
	15.26 GB
	16.18 GB
	16.14 GB
	15.04 GB
	17.77 GB
	22.76 GB

vantagewithai/LTX2.3-10Eros-1.5-GGUF

vantagewithai/LTX2.3-10Eros-1.5-GGUF — v1.5 (latest)

Quant	Size	Download
	11.13 GB
	10.34 GB
	12.98 GB
	13.90 GB
	14.30 GB
	13.20 GB
	15.26 GB
	16.18 GB
	16.14 GB
	15.04 GB
	17.77 GB
	22.76 GB

◦ 10Eros Splits

per-version component split

vantagewithai/LTX2.3-10Eros-Split — v1

Component	Precision	Size
Model		42.01 GB
Model		28.28 GB
VAE	—	1.45 GB
Audio VAE	—	364.86 MB
Text encoder	—	2.31 GB

vantagewithai/LTX2.3-10Eros-1.2-Split — v1.2

Component	Precision	Size
Model		42.01 GB
Model		30.2 GB
VAE	—	1.45 GB
Audio VAE	—	364.86 MB
LoRA (bundled)	—	662.07 MB

vantagewithai/LTX2.3-10Eros-1.3-Split — v1.3

Component	Precision	Size
Model		42.01 GB
Model		25.03 GB
VAE	—	1.45 GB
Audio VAE	—	364.86 MB
Text encoder	—	2.31 GB

vantagewithai/LTX2.3-10Eros-1.4-Split — v1.4 (latest)

Component	Precision	Size
Model		42.01 GB
Model		25.03 GB
VAE	—	1.45 GB
Audio VAE	—	364.86 MB
Text encoder	—	2.31 GB

◦ 10Eros Splits — AX1Y2JP transformer-only fork (alternative split for ComfyUI)

Variant	Download
v1.4 transformer_only bf16
v1.4 transformer_only fp8mixed_learned
v1.4 video VAE bf16

◦ 10Eros Extracted LoRA — maximsobolev275 trained LoRAs

Variant	Description	Download
rank-768 family (canonical, v1.4)	Author: maximsobolev275. Training LoRAs directly extracted from the 10Eros v1.4 merge (rank 768); lower-rank rerolls for v1.2 / v1.3 / v1.4 are also published in the same repo.	LTX-10Eros-LoRA-r768

◦ 10Eros distilled-baked collection — ibyteohdear

Massive 10Eros distilled-baked checkpoint collection by ibyteohdear — 20 safetensors of full LTX distilled checkpoints with the 10Eros style baked in. The repo was restructured on 2026-10-26, so the catalogue below was rebuilt from the live tree: the earlier r64 / r128 / r768 / r768_0.85 builds and the bundled JoyAI-Echo_r256 / reasoning I2V LoRAs are gone.

◦ LTX-2.3 — LTX2.3_DISTILLED_BAKED_LTX_SULPHUR_STYLE_IS_10Eros_v{1,12,13,14,15}_r{256,512} — full checkpoints (DiT + vocoder + video/audio vae + text_embedding_projection; 5,947 tensors, all BF16, model_version 2.3.0, config AVTransformer3DModel), baked from the Sulphur-2 base with the 10Eros style merged in.

◦ LTX-2.5 — LTX2.5_DISTILLED_STYLE_10Eros_{1,12,13,14,15}[_SULPHUR]_r512_audio_r256_rest_0 — transformer-only (model.diffusion_model.*; 4,349 tensors, BF16 + 290 F32 scale-shift tables, model_version 2.5.0, gemma_source_checkpoint: gemma4-12b-ltx-v1). Two-pass bake over the LTX-2.5 dev base: PASS 1 = 10Eros style LoRA (rank 512, α 512) at 0.85 on the style layers; PASS 2 = a rank-256 LoRA on the audio branch only (0.85 audio / 0 elsewhere — hence audio_r256_rest_0). The _SULPHUR builds swap the PASS-2 LoRA for Sulphur_r256.

Ver	Build	Name	Size
2.3	distilled-baked 10Eros + Sulphur style	`v1 r256`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v1 r512`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v12 r256`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v12 r512`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v13 r256`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v13 r512`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v14 r256`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v14 r512`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v15 r256`	46.14 GB
2.3	distilled-baked 10Eros + Sulphur style	`v15 r512`	46.14 GB
2.5	distilled-baked 10Eros style (dev 0.85)	`v1 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros style (dev 0.85)	`v12 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros style (dev 0.85)	`v13 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros style (dev 0.85)	`v14 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros style (dev 0.85)	`v15 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros + Sulphur style (dev 0.85)	`v1 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros + Sulphur style (dev 0.85)	`v12 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros + Sulphur style (dev 0.85)	`v13 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros + Sulphur style (dev 0.85)	`v14 r512 + audio r256`	42.02 GB
2.5	distilled-baked 10Eros + Sulphur style (dev 0.85)	`v15 r512 + audio r256`	42.02 GB

· · · · · · · · · · · · · ·

❖ 10Eros Max (MiniMax-H3 grafts)

⚠️ This is a different architecture — these are not LTX-2.3 checkpoints. 10Eros Max is a MiniMax-H3 (Hailuo 3) model line. The weights are a 50-block H3 transformer (token_refiner.blocks, final_layer.video_out / audio_out, video_patch_proj, audio_patch_proj, condition_proj) with character cross-model grafted into the attention layers from LTX-2.3, Wan 2.2 and Krea 2. Verified by safetensors header and model card: base_model: MiniMaxAI/MiniMax-H3, image_model: minimax_h3, converted_from: diffusers MiniMax-H3 transformer. The pruned builds are 20.11 B params; the unpruned beta6_bf16 is 33.12 B (535 tensors, all BF16) and carries a bias set that the stock H3 base does not.

◦ Licence: the standard MiniMax-H3 community licence applies. Because the release carries transferred character from LTX 2.3, Wan 2.2 and Krea 2, those source licences apply too to the portions of character that came from each.

◦ Author's method is documented in h3_graft_methodology.md — an unfuse→graft→refuse toolchain that maps donor attention heads and MLP inner dimensions onto H3's decomposition (Krea 2's 6144 hidden is truncated to H3's 5376; Wan 2.2's biases are discarded because H3 is bias-free).

◦ Sampling: res_multistep / simple at 6–8 steps gives the best motion; LCM / simple at 6–8 steps the best turbo audio; er_sde at 4–6 steps. Avoid SLA sparsity — it destroys these weights. "Turbo" builds bake the 4-step turbo delta in (saves ~4.2 GB of VRAM); a 4-step turbo LoRA can be stacked on a reference build to reach 4 steps. Per the author, beta_3 and beta_4 are corrupted test versions and beta_5 is the first fully functional one.

◦ Do not confuse with the LTX-2.3 10Eros line in the table above — those are true AVTransformer3DModel LTX-2.3 checkpoints.

Public weights — TenStrip/10Eros-Max (17 checkpoints + 1 LoRA):

Ver	Build	Name	Size
H3	beta6 — 8-step TURBO, unpruned	`10Eros_Max_h3_8stepTURBO_beta6_bf16`	66.25 GB
H3	beta6 — 8-step TURBO, pruned	`10Eros_Max_h3_8stepTURBO_beta6_pruned`	40.22 GB
H3	beta6 — 8-step TURBO, pruned	`10Eros_Max_h3_8stepTURBO_beta6_int8`	21.71 GB
H3	beta6 — VSA gates (unstable)	`10Eros_Max_FastH3_VSAgates_UNSTABLE_beta6-int8_HQ`	26.75 GB
H3	beta5 — TURBO hybrid	`10Eros_Max_h3_TURBO-hybrid_beta5`	40.22 GB
H3	beta5 — TURBO hybrid	`10Eros_Max_h3_TURBO-hybrid_beta5_int8`	20.97 GB
H3	beta5 — TURBO hybrid, 14 GB opt	`10Eros_Max_h3_TURBO-hybrid_beta5_w4a8_14gb_optimized`	14.00 GB
H3	beta5 — hybrid (non-turbo)	`10Eros_Max_h3_hybrid_beta5`	40.22 GB
H3	beta5 — hybrid (non-turbo)	`10Eros_Max_h3_hybrid_beta5_int8`	20.97 GB
H3	beta5 — hybrid, 14 GB opt	`10Eros_Max_h3_hybrid_beta5_w4a8_14gb_optimized`	14.00 GB
H3	beta3 — TURBO hybrid (corrupted test)	`10Eros_Max_h3_TURBO-hybrid_beta3`	40.23 GB
H3	beta3 — TURBO hybrid (corrupted test)	`10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot`	20.98 GB
H3	beta4 — TURBO hybrid (corrupted test)	`10Eros_Max_h3_TURBO-hybrid_beta4`	40.22 GB
H3	beta4 — TURBO hybrid (corrupted test)	`10Eros_Max_h3_TURBO-hybrid_beta4_int8_convrot`	20.97 GB
H3	beta2 — ref2va, TURBO	`10Eros_Max_h3_TURBO_ref2va_beta2`	40.23 GB
H3	beta2 — ref2va, pruned	`10Eros_Max_h3_ref2va_beta2_pruned`	40.22 GB
H3	beta2 — fl2va, pruned	`10Eros_Max_h3_fl2va_beta2_pruned`	40.22 GB

Gated beta7 set — TenStrip/LTX2.3-10Eros_Version-Testing (18 weights, 476.55 GB)

TenStrip/LTX2.3-10Eros_Version-Testing is gated manual — request access, it is not instant auto-approve, and the model card itself is locked. The repo was fully replaced on 2026-10-27: every beta6 file the README previously linked to is gone, and the tree now holds only beta7 weights (plus a 626-byte README). File names and byte sizes below come from the public tree API; precision labels are inferred from the file names because a gated checkpoint's header cannot be fetched without approval.

Ver	Build	Name	Size
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1`	40.22 GB
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1-int8`	20.97 GB
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1-int8_HQ`	24.63 GB
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1-int8_learned`	20.97 GB
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1-w4a8_conservative_cicalooo`	19.08 GB
H3	beta7 test1 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test1_activation_learned_HQ`	18.69 GB
H3	beta7 test2 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test2`	40.22 GB
H3	beta7 test2 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test2-int8`	20.97 GB
H3	beta7 test2 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test2-native_int8_HQ`	22.89 GB
H3	beta7 test2 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test2_activation_learned_HQ`	18.69 GB
H3	beta7 test2 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test2_cicaloo_w4a8conservative_gptq_activated`	19.08 GB
H3	beta7 test3 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test3`	40.22 GB
H3	beta7 test3 — 8-step TURBO	`10Eros_Max_h3_8stepTURBO_beta7_test3-native_int8_protected_edges_HQ`	22.89 GB
H3	beta7 test1 — raw (no turbo)	`10Eros_Max_h3_raw_beta7_test1`	40.22 GB
H3	beta7 test1 — raw (no turbo)	`10Eros_Max_h3_raw_beta7_test1-int8`	20.97 GB
H3	beta7 test1 — raw (no turbo)	`10Eros_Max_h3_raw_beta7_test1-int8_HQ`	24.63 GB
H3	beta7 test2 — raw (no turbo)	`10Eros_Max_h3_raw_beta7_test2`	40.22 GB
H3	beta7 test2 — raw (no turbo)	`10Eros_Max_h3_raw_beta7_test2-int8`	20.97 GB

int8 ConvRot mirrors — cicalooo/10Eros-Max-h3-int8-convrot + CornLogic/10EROS-INT8

The int8 ConvRot H3 builds are mirrored publicly. cicalooo's _skip_edges variants keep the first and last blocks in BF16 (see the repo's SKIP_EDGES.md) — use them when ConvRot rotation hurts the edge blocks.

cicalooo/10Eros-Max-h3-int8-convrot (8 files, 172.38 GB):

Ver	Build	Name	Size
H3	beta3 — TURBO hybrid	`10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot`	20.97 GB
H3	beta3 — TURBO hybrid, skip edges	`10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot_skip_edges`	22.51 GB
H3	beta2 — ref2va	`10Eros_Max_h3_TURBO_ref2va_beta2_int8_convrot`	20.97 GB
H3	beta2 — ref2va, skip edges	`10Eros_Max_h3_TURBO_ref2va_beta2_int8_convrot_skip_edges`	22.51 GB
H3	beta2 — fl2va, pruned	`10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot`	20.20 GB
H3	beta2 — fl2va, pruned, skip edges	`10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot_skip_edges`	21.74 GB
H3	beta2 — ref2va, pruned	`10Eros_Max_h3_ref2va_beta2_pruned_int8_convrot`	20.97 GB
H3	beta2 — ref2va, pruned, skip edges	`10Eros_Max_h3_ref2va_beta2_pruned_int8_convrot_skip_edges`	22.51 GB

CornLogic/10EROS-INT8 — H3 int8 files only (the repo's LTX-2.3 v1.5 files are listed in the ❖ 10Eros table above):

Ver	Build	Name	Size
H3	beta1 — fl2va, pruned	`10Eros_Max_h3_fl2va_INT8_beta1_pruned`	20.97 GB
H3	test2 — fl2va, pruned	`10Eros_Max_h3_fl2va_INT8_test2_pruned`	20.97 GB
H3	test4 — fl2va, pruned	`10Eros_Max_h3_fl2va_INT8_test4_pruned`	20.97 GB

H3 LoRAs — these target the MiniMax-H3 transformer, not an LTX DiT, so they will not load into any LTX checkpoint. They are listed here with the H3 line rather than in the LTX LoRA tables of block 60.

LoRA	Ver	Size	Description
`Eros_beta6mix_Lora`	H3	387.63 MB	Rank-32 H3 PEFT LoRA for the 10Eros Max beta6 line — a `dare_ties` merge of 17 satellite LoRAs (ties density 0.3, sign consensus mass)
`H3_krea_style_test1_rank512`	H3	1.10 GB	Krea-2 style LoRA extracted from the Krea-2 branch of the author's attn-graft experiments (rank 512 captures ~52% of the SLERP orthogonal delta). T2V only.
`H3_krea_style_test1_rank128`	H3	275.27 MB	Same as above at rank 128 (~24% of the delta — small but visibly shifts visual style)
`H3_Anima_Tokenizer`	H3	763.37 MB	Anima-style LoRA grafted across the H3 text-encoder layers of token-refiner blocks 0–1, extracted at rank 2048. Strongly affects style in T2V and detail in I2V
`Minimax-h3_Singularity_128spect-fro95_lora`	H3	1.28 GB	Extracted from the base hybrid pruned build with spectral dampening, against WarmBloodAban/Minimax-h3_Singularity. Use at 0.5 strength or lower — higher strengths can drift
`Minimax-h3_Singularity_64-fro95_lora`	H3	655.31 MB	Same as above at rank 64
`H3_Wan2.2_i2v_motion_rank1024_test1`	H3	1.38 GB	Wan-2.2 motion LoRA grafted onto the mid-to-late H3 blocks with peak mapping (rank 1024 captures ~69% of the residual delta). Shifts motion style towards Wan's repetitive motion
`H3_Wan2.2_i2v_motion_rank512_test1`	H3	688.14 MB	Same as above at rank 512 (~41.6% of the residual delta)
`lightx2v_hybrid-4to8step-Turbo_r48`	H3	943.95 MB	H3 PEFT turbo LoRA turning 8-step sampling into 4-step (rank 48)
`lightx2v_hybrid-4to8step-full-fusion_Turbo_pruned`	H3	4.46 GB	Full-fusion H3 turbo build (777 tensors: 750 `diffusion_model.blocks` + 24 `token_refiner` + 3 `final_layer`) — loads standalone instead of as a LoRA

· · · · · · · · · · · · · ·

❖ Sulphur-2-base

Repo: SulphurAI/Sulphur-2-base — uncensored video generation model based on LTX 2.3 with built-in prompt enhancer. Merge base for 10Eros. T2V + I2V native.

◦ Main video model

Ver	Build	Size
2.3	dev	46.14 GB
2.3	dev	29.16 GB
2.3	distil	46.14 GB
2.3	distil	29.16 GB
2.3	distil	19.51 GB

◦ vantagewithai component split — vantagewithai/Sulphur-2-Base-Split

Component	Precision	Size
model (`sulphur_dev_bf16_model`)		42.01 GB
model (`sulphur_dev_model_fp8mixed`)		25.03 GB
model (`sulphur_distil_bf16_model`)		42.01 GB
vae	—	1.45 GB
audio_vae	—	364.86 MB

Abiray GGUF quants

◦ Abiray GGUF quants — Abiray/Sulphur-2-base-GGUF

Build	Precision	Size	Download
sulphur_dev		42.03 GB
sulphur_dev		11.13 GB
sulphur_dev		10.34 GB
sulphur_dev		12.98 GB
sulphur_dev		14.3 GB
sulphur_dev		13.2 GB
sulphur_dev		15.26 GB
sulphur_dev		16.14 GB
sulphur_dev		15.04 GB
sulphur_dev		17.77 GB
sulphur_dev		22.76 GB

◦ Sulphur LoRAs

Ver	Build	Size	Download
2.3	`sulphur_lora_rank_768`	10.27 GB
2.3 (experimental)	`sulphur_experimental_lora_v1`	14.54 GB

◦ Prompt Enhancer

Variant	Size	Download
Censored	921.71 MB	┆
Censored	9.53 GB	┆
Uncensored	921.7 MB	┆
Uncensored	9.79 GB	┆
Uncensored (safetensors)	19.31 GB

◦ The uncensored prompt enhancer also ships as a 4-shard safetensors set (19.31 GB total) next to the Q8_0 GGUF — download model.safetensors.index.json and the four model.safetensors-0000N-of-00004.safetensors shards it names (the tokenizer / processor / chat-template JSONs sit alongside them). The four EllaPriest45/LTX2.3_base copies are byte-identical re-packs of the official files (matching byte counts), listed as second links above.

· · · · · · · · · · · · · ·

❖ JoyAI-Echo Surgical

Surgical finetune based on jdopensource/JoyAI-Echo by joeygambino. Combined "echoVid + ltxAud" surgical variant — jointly fine-tuned for both video generation and audio on top of LTX-2.3. A LoRA extracted from JoyAI-Echo (TenStrip's LTX2.3_JoyAI_Lora_Extracted) is listed separately under ### ▣ Special.

Ver	Build	Name	Size
2.3	unsplit (full DiT)	`echoVid-ltxAud surgical`	46.14 GB
2.3	unsplit (full DiT)	`echoVid-ltxAud surgical`	25.13 GB
2.3	unsplit (full DiT)	`echoVid-ltxAud surgical`	29.16 GB
2.3	transformer-only	`echoVid-ltxAud surgical`	27.34 GB

◦ JoyAI-Echo Surgical — GGUF (low-VRAM quants)

joeygambino/joyai-echo-ltx23-echoVid-ltxAud-surgical-gguf

joeygambino/joyai-echo-ltx23-echoVid-ltxAud-surgical-gguf — Surgical DiT GGUF quants by joeygambino.

Ver	Build	Name	Precision	Size	Download
2.3	GGUF	`echoVid-ltxAud surgical`		16.69 GB
2.3	GGUF	`echoVid-ltxAud surgical`		24.83 GB

· · · · · · · · · · · · · ·

❖ JoyAI-Echo (LTX-2.5)

Surgical merge of jdopensource/JoyAI-Echo into the LTX-2.5 dev transformer by joeygambino — JoyAI-Echo's video attention / feed-forward delta transplanted in, with nothing distilled baked on top. Three builds: dev (bring-your-own distill LoRA), comfy-native (few-step, distill baked at 0.5 — RTX 50), and GGUF (few-step — RTX 30/40). Two doses per build: 070T30 (0.7× delta / 0.3× modulation) and 100T50 (1.0× / 0.5×). Workflows: ComfyUI-JoyLTX25.

Ver	Build	Name	Size
2.5	dev	`echoVid 070T30`	42.02 GB
2.5	dev	`echoVid 070T30`	21.48 GB
2.5	dev	`echoVid 070T30`	21.50 GB
2.5	dev	`echoVid 100T50`	42.02 GB
2.5	dev	`echoVid 100T50`	21.48 GB
2.5	dev	`echoVid 100T50`	21.50 GB
2.5	comfy-native	`echoVid 070T30`	21.50 GB
2.5	comfy-native	`echoVid 070T30`	17.01 GB
2.5	comfy-native	`echoVid 070T30`	12.50 GB
2.5	comfy-native	`echoVid 070T30`	11.24 GB
2.5	comfy-native	`echoVid 070T30`	12.52 GB
2.5	comfy-native	`echoVid 100T50`	21.50 GB
2.5	comfy-native	`echoVid 100T50`	13.81 GB
2.5	comfy-native	`echoVid 100T50`	17.01 GB
2.5	comfy-native	`echoVid 100T50`	12.50 GB
2.5	comfy-native	`echoVid 100T50`	11.24 GB
2.5	comfy-native	`echoVid 100T50`	12.52 GB

Z-Image surface-graft variant (× Z-Image) — same echoVid 070T30 v2 engine with Z-Image's spatial attention profile grafted in as per-block q_norm rescaling (dose 0.5, no retraining, no Z-Image weights at inference). Targets surface/texture rendering: fine detail on foliage, straw, gravel, fabric and skin. Drop-in replacement for any stock v2 workflow — same loaders, settings and seeds.

Ver	Build	Name	Size
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	42.02 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	21.48 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	21.50 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	13.81 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	17.01 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	12.50 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	11.24 GB
2.5	comfy-native	`echoVid 070T30 × Z-Image zgraft05`	12.52 GB

◦ JoyAI-Echo × Z-Image — GGUF (RTX 30/40; K-quants run 4–8× faster than emulated 4-bit there)

joeygambino/joyai-echo-ltx25-x-Z-Image-gguf

joeygambino/joyai-echo-ltx25-x-Z-Image-gguf — Z-Image-grafted echoVid 070T30 v2 GGUF quants by joeygambino (Q4_K_M/Q4_K_S are the RTX 30/40 sweet spot).

Ver	Build	Name	Size
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	7.91 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	10.60 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	14.17 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	12.93 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	15.90 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	17.75 GB
2.5	GGUF	`echoVid 070T30 × Z-Image zgraft05`	22.73 GB

◦ JoyAI-Echo (LTX-2.5) — GGUF (low-VRAM quants)

joeygambino/joyai-echo-ltx25-echoVid-gguf

joeygambino/joyai-echo-ltx25-echoVid-gguf — echoVid GGUF quants by joeygambino (few-step, distill baked at 0.5; RTX 30/40).

Ver	Build	Name	Size
2.5	GGUF	`echoVid 070T30`	7.91 GB
2.5	GGUF	`echoVid 070T30`	10.60 GB
2.5	GGUF	`echoVid 070T30`	14.17 GB
2.5	GGUF	`echoVid 070T30`	12.93 GB
2.5	GGUF	`echoVid 070T30`	15.90 GB
2.5	GGUF	`echoVid 070T30`	17.75 GB
2.5	GGUF	`echoVid 070T30`	22.73 GB
2.5	GGUF	`echoVid 100T50`	7.91 GB
2.5	GGUF	`echoVid 100T50`	10.60 GB
2.5	GGUF	`echoVid 100T50`	14.17 GB
2.5	GGUF	`echoVid 100T50`	12.93 GB
2.5	GGUF	`echoVid 100T50`	15.90 GB
2.5	GGUF	`echoVid 100T50`	17.75 GB
2.5	GGUF	`echoVid 100T50`	22.73 GB

· · · · · · · · · · · · · ·

❖ PinkCherry NSFW

Repo: SexGod1979/PinkCherry_NSFW_LTX23 — uncensored NSFW LTX-2.3 finetune family (NSFW content only — do not use for clean content). Pairs with the official LTX-2.3 distilled LoRA 384.

Variant	Precision	Size	Download
v1.3 dev		46.14 GB
v1.3 dev		27.62 GB
v1.5 dev		46.14 GB
v1.5 dev		27.64 GB
v1.6 dev		46.14 GB
v1.6 dev		27.62 GB
v1.6 dev		27.64 GB
v1.7-alpha dev		46.14 GB
v1.7-alpha dev		27.62 GB
v1.7-alpha dev		27.64 GB
v1.8 dev		46.14 GB
v1.8 dev		27.62 GB
v1.8 dev		27.64 GB

◦ PinkCherry GGUF — low-VRAM quants by SexGod1979

v1.7-alpha / v1.8 GGUF

Quant	Build	Size	Download
	v1.7-alpha	15.93 GB
	v1.7-alpha	17.77 GB
	v1.8	15.93 GB
	v1.8	22.76 GB

❖ Elastic

Elastic is a TensorRT-engine distribution of a LoRA-integrated distilled FP8 T2V variant of LTX-2.3, packaged by TheStageAI as .qlip shard files for H100 GPUs.

Build	Name	Precision	Size	Download
distil + LoRA T2V	`Elastic — H100`	fp8	~19 GB (49 `.qlip` shards)

· · · · · · · · · · · · · ·

❖ ChrisColeTech Uncensored Turbo (LTX-2.3)

Pre-merged LTX-2.3 Uncensored Turbo DiT by ChrisColeTech with three fine-tunes baked into the weights: 10Eros NSFW LoRA (1.0), DMD distilled LoRA (1.0, 4-step capable — 8 steps recommended), and the official LTXV ICLoRA Detailer (0.6). Modes: T2AV / I2AV / V2V / A2V / REF2VA; example workflows bundled in the repo. v1.4 is the build documented in the card; the repo tree has since gained a v1.5 fp8_scaled DiT (21.12 GB) that the card does not describe yet.

Ver	Build	Size
2.3	v1.4 fp8mixed	29.16 GB
2.3	v1.5 fp8_scaled	21.12 GB
2.3	v1.4 GGUF	14.18 GB
2.3	v1.4 GGUF	17.76 GB
2.3	v1.4 GGUF	22.74 GB

◦ Component splits live in the same repo under split/: the uncensored v1.4 projections (ltxv23_uncensored_v1.4_projections.safetensors, 2.31 GB — required, goes in text_encoders/), an abliterated Gemma-3-12b Q4_K_M encoder with bi-projections (7.30 GB), the uncensored v1.4 audio / video VAEs, the older v1.0 uncensored build, a stock LTX-2.3 distilled fp8 under v1.0/distilled_base/, and LTX-2.3 spatial upscalers x2 1.0 / 1.1 — see the repo tree for the full list.

v1.0 builds (older, also in this repo)

Ver	Build	Size
2.3	v1.0 fp8	29.16 GB
2.3	v1.0 int8	29.16 GB
2.3	v1.0 GGUF	14.30 GB
2.3	v1.0 GGUF	17.77 GB
2.3	v1.0 GGUF	22.76 GB
2.3	v1.0 distilled fp8 (stock base)	29.53 GB

· · · · · · · · · · · · · ·

❖ ChrisColeTech Uncensored Turbo v1.1 (LTX-2.5)

Pre-merged LTX-2.5 Extreme Uncensored Turbo v1.1 DiT by ChrisColeTech — "Eros10 NSFW & 4-step LoRA baked-in". Five fine-tunes are baked into the weights: the 10Eros NSFW LoRA (0.7), the DMD distilled LoRA (1.0 — 4 steps work, 8 recommended), the official LTXV IC-LoRA (0.7), the IC-LoRA Detailer (0.45, for V2V control on top of T2V/I2V) and an Img2Vid adapter LoRA (0.45). Base is Lightricks/LTX-2.5; modes T2V / I2V / V2V / A2V / REF2VA.

Because the IC-LoRA is baked in, prompts can use the [VISUAL] / [SPEECH] / [SOUNDS] template blocks for precise control of picture, dialogue and ambience. Recommended settings: 1280×720, 124 frames @ 24 fps, 8 steps (16 max), guidance 1.0.

⚠️ The repo was previously published as ChrisColeTech/LTX-2.5-turbo-GGUF; the old URL now redirects here. The card's own component links are broken — it names ltx25_uncensored_v1.1_audio_vae / ..._video_vae, but the files on disk are ltx25_uncensored_audio_vae / ltx25_uncensored_video_vae (no v1.1 infix).

Ver	Build	Size
2.5	v1.1 fp8	21.02 GB
2.5	v1.1 fp8 (scaled)	21.47 GB
2.5	v1.1 int8	21.47 GB
2.5	v1.1 GGUF	14.18 GB
2.5	v1.1 GGUF	17.76 GB
2.5	v1.1 GGUF	22.33 GB

◦ Precision, verified from the safetensors headers — fp8 = 4,349 tensors (BF16 × 2,577 + F8_E4M3 × 1,772, unscaled); fp8_scaled = 5,789 tensors (BF16 × 2,909 + F8_E4M3 × 1,440 + F32 × 1,440 scales); int8 = 5,789 tensors (BF16 × 2,909 + I8 × 1,440 + F32 × 1,440), metadata quantization: int8_tensorwise — the weights are stored as float8_e4m3fn with per-layer scales and full-precision matrix multiplies. All three are AVTransformer3DModel, model_version 2.5.0, gemma4-12b-ltx-v1.

◦ Component splits in the same repo under split/: the uncensored Gemma-4-12b text encoder with the LTX projections already baked in (gemma4_12b_ltx25_uncensored-int8.safetensors, 13.16 GB — see Gemma-4-12b), its GGUF variants, the uncensored video / audio VAEs (byte-identical to the official Lightricks/LTX-2.5 VAEs) and the LTX-2.5 spatial / temporal latent upscalers x2.

· · · · · · · · · · · · · ·

❖ REDGraft (NSFW)

Uncensored NSFW LTX-2.5 checkpoint by EllaPriest45 ("REDGraft"), shipped as a single INT8 full checkpoint together with a bundled two-stage i2V workflow JSON (ManualSigmas ×2 + LoRA loader).

Ver	Build	Precision	Size	Download
2.5	REDGraft NSFW (full ckpt)		17.03 GB

◦ Bundled workflow: REDGraft (NSFW) INT8 - LTX2.5.json.

· · · · · · · · · · · · · ·

❖ colinb83 NSFW (LTX-2.5 full FT)

A full bf16 fine-tune of the LTX-2.5 22B DiT on a 54,702-clip NSFW corpus (1280×704, 10 s + 5 s clips) by colinb83 — the final checkpoint of the completed 7,000-step run. All 48 transformer blocks were trained (18.56B params, 4,032 tensors) with a streaming dense full fine-tune using stochastic-BF16 Adafactor; connectors and non-block tensors are the original base weights. Single RTX PRO 6000 Blackwell with ring offload (8 blocks swapped / 2-block double buffer), ~38 h total.

Ver	Build	Precision	Size	Download
2.5	NSFW full FT (step 7000)		42.01 GB
2.5	NSFW full FT (step 4500 test merge)		42.01 GB

◦ Base / requirements: LTX-2.5 22B dev (AVTransformer3DModel, model_version: 2.5.0). The card declares an abliterated Gemma 4 12B text encoder (gemma4-12b-ltx-v1) — use that encoder plus the standard LTX-2.5 video/audio VAEs. Load with UNETLoader as a standard LTX-2.5 checkpoint in any ComfyUI workflow.

◦ Training recipe: 2,000 steps @ 1e-5 then 5,000 steps @ 3e-6 (constant per phase), batch 1, gradient checkpointing on. Trained at 1280×704 × 97 frames (~10 s); captions are long-form, camera-first, present-tense descriptions.

◦ ⚠️ Distill LoRAs do not transfer cleanly. Distillation LoRAs made for the unmodified dev model (e.g. ltx-2.5-22b-distilled-lora-450) apply a weight delta ~44% the magnitude of this fine-tune and are calibrated against the original base — use them at reduced strength or retrain one against this checkpoint.

◦ No model card in the repo itself (the tree holds only the weights file + .gitattributes); the documentation above comes from the author's sibling repo colinb83/ltx-2.5-nsfw, which ships the same final checkpoint plus an earlier dense_weights_step_00250.safetensors (28.24 GB) snapshot.

◦ Step-4500 test merge — an intermediate snapshot of the same run, published as a separate public repo colinb83/KimCherry4500test (commit title “LTX-2.5 22B NSFW full fine-tune, step 4500 test merge”). Its header matches the final checkpoint exactly — 4,349 tensors, the same model.diffusion_model.* key set over 48 blocks, AVTransformer3DModel — but keeps 146 F32 scale-shift tables (scale_shift_table, audio_scale_shift_table, audio_prompt_scale_shift_table, scale_shift_table_a2v_ca_audio) where the step-7000 file has only 2, i.e. it is the pre-BF16-cast merge state of the same training run rather than a differently-trained model. The repo has no model card either — the single file is KimCherry4500test.safetensors.

◦ License: subject to the LTX-2.x Community License Agreement (full text embedded in the safetensors metadata); derivatives of LTX-2.x must be distributed under that agreement. Content is NSFW/adult — the repo carries the not-for-all-audiences tag.

· · · · · · · · · · · · · ·

❖ SpatialAV2AV (BingoG)

Publicly-released audio→audio-video training checkpoints from BingoG/LTX-2-SpatialAV2AV-Checkpoints for spatial-audio conditioned video generation. Two curriculum layouts (E4 Core Dual + Dynamic; E6 reserved for Full) at 480p cap, 113 frames at 25 fps.

Curriculum	Step	Size
E4 Core (Dual + Dynamic)	400	37.98 GB
E4 Core (Dual + Dynamic)	1000	37.98 GB
E6 Full	1000	37.98 GB

Note: these are intermediate training checkpoints, not inference-ready models — useful for fine-tuning experiments and reproducibility only.

· · · · · · · · · · · · · ·

❖ One-Step Refiner (Owen777)

One-step LTX-2.5 video refiner by Owen777 — a DMD/GAN-style one-step refinement model (EMA step 835) that sharpens/refines generated video in a single pass. Access is gated — request access on the repo before downloading.

Ver	Build	Precision	Size	Download
2.5	one-step refiner (EMA step 835)		26.25 GB

◦ Reinforcement-learning LoRA also in repo: rl_lora/step_000400.pt (~0.81 GB).

❖ SoL-Refiner — one-step 1080p/2K upscaler (szwagros, int8 ConvRot)

Generic one-step refiner from NVlabs' SoL-Refiner project (SANA / SANA-Video): a cheap low-res draft (SANA-Video, WAN, MiniMax-H3, …) is encoded, 2x latent-upscaled with AdaIN, re-noised and denoised in a single transformer forward, turning the draft into 1080p/2K output. Repackaged by szwagros from the Efficient-Large-Model originals into the image-server ltx_core layout (diffusion_models/ + text_encoders/ + vae/ + latent_upscale_models/) — the transformer and text encoder are int8 ConvRot and expect the image_server_kernels.int8_linear backend, so these are not drop-in ComfyUI checkpoints.

LTX-2.3 One-Step — szwagros/SoL-Refiner-LTX-2.3-One-Step-int8-convrot · 41.93 GB total · re-noised to sigma 0.725

Component	Precision	Size	Download
`sol-refiner-ltx-2.3-one-step-transformer-comfy-int8-convrot` (DiT)		23.51 GB
`gemma3-12b-with-proj-…-comfy-int8-convrot` (text encoder)		15.97 GB
`ltx-2.3-22b_vae` (video VAE)		1.45 GB
`ltx-2.3-spatial-upscaler-x2-1.1` (x2 latent upsampler)		1.00 GB

⚠️ Always condition with frame_rate=16, whatever the input fps (encode the output at the real fps) — the model was trained on 16 fps drafts and a 24 fps value mis-scales the temporal RoPE and ghosts moving objects. Uses the v1.1 x2 upsampler, not v1.0.

LTX-2.5 for MiniMax-H3 — szwagros/SoL-Refiner-LTX-2.5-H3-int8-convrot · 41.35 GB total · re-noised to sigma 0.909375

Component	Precision	Size	Download
`sol-refiner-ltx-2.5-h3-transformer-comfy-int8-convrot` (DiT)		23.51 GB
`gemma4-12b-with-proj-…-comfy-int8-convrot` (text encoder)		15.37 GB
`sol-refiner-ltx-2.5-h3-video-vae-bf16` (VAE + diffusion decoder)		1.47 GB
`ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0` (x2 latent upsampler)		1.00 GB

◦ Conditioning frame_rate = the input fps (H3 drafts are 24 fps). One step, no CFG. Video output only — no audio. The text encoder is the refiner's shipped Gemma 4, i.e. the pre-2026-08-17 Lightricks encoder rather than the current one.

◦ Both are finetunes of LTX-2.x, so the LTX-2.x Community License (Lightricks) applies — check it before commercial use. Source repos: 2.3 One-Step · 2.5 for MiniMax-H3.

· · · · · · · · · · · · · ·

❖ SolarWM (World Model)

Research LTX-2.5 world model by junchaoh-cs (arXiv:2609.02886) — a long-horizon, camera-controllable video world model finetuned from LTX-2.5 22B. The repo re-ships the LTX-2.5 22B base (dev transformer + gemma4-12b text encoder + VAE) plus a finetuned stage, SolarWM-ltx-22B-bid-stage0p5-153f (EMA weights, ema/model.safetensors). Access is gated — request access before downloading; exact artifact formats are unconfirmed (gated).

Ver	Build	Precision	Size	Download
2.5	SolarWM bid-stage0p5-153f (finetuned stage)	—	3.93 GB

◦ Base LTX-2.5 22B (dev transformer + gemma4-12b encoder + VAE) also in repo. Paper: arXiv:2609.02886.

· · · · · · · · · · · · · ·

❖ Stubelius Remix (Stuubs)

LTX-2.5 merged finetune by Stuubs (12 likes) — the distilled LoRA baked in, plus (from v1) OmniNFT RL-finetune merged at its trained strength. Mainly aimed at NSFW but works SFW (name clothing explicitly in SFW prompts or you may get unwanted nudity). Ships three v1 builds plus two beta builds; use v1.

Ver	Build	Size
2.5	`v1` (recommended, ComfyUI int8)	21.50 GB
2.5	`v1` (merging / own conversions)	42.01 GB
2.5	`v1` NVFP4 (RTX 50-series only, experimental)	14.40 GB
2.5	`beta2`	21.50 GB
2.5	`beta2`	42.01 GB
2.5	`beta1`	42.01 GB

◦ Run it: drop the int8 file in ComfyUI/models/diffusion_models, load with Load Diffusion Model (or pick it in the Models node of Stubelius Ultimate LTX 2.5). CFG 1, 8 steps (+3–4 if you run an upscale pass). Do not add a distilled LoRA or OmniNFT on top — both are already baked in. Needs gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors (Lightricks re-uploaded it in mid-August — re-download if yours is older) plus ltx-2.5-video-vae-bf16 and ltx-2.5-audio-vae-bf16.

◦ v1 vs beta2, per the author: +55% sharper fine detail (faces, hair, fabric, backgrounds), +15% more motion, same remix / prompt-following / loudness, dialogue still verbatim. Blind AI-judged A/B of full-res frames: OmniNFT at v1's strength won 29–9 with 10 ties. A few scenes come out less saturated on average. VBVR was tested and rejected — it cost motion without improving prompt following.

◦ NVFP4 is experimental: transformer blocks 0–41 in NVFP4, last 6 blocks + text connectors in int8 (Lightricks' own layer choice). Fast only on RTX 50-series (Blackwell) with ComfyUI ≥ 0.32 and PyTorch built for CUDA 13; older cards run it but slower than int8. Renders haven't been fully compared against int8 yet.

◦ The int8 build needs a recent enough ComfyUI to load Comfy-Org's int8 LTX-2.5 files. LTX-2.5 LoRAs load on it the same way as on stock LTX-2.5. Credits: Lightricks (LTX-2.5 + distilled LoRA), FusionCow / Sulp.

· · · · · · · · · · · · · ·

❖ EllaPriest45 (model mirror)

Note: the EllaPriest45 LTX collection is split across sub-repos, not one monorepo. Older revisions of this list linked a single EllaPriest45/LTX2.5_checkpoints; the same author also mirrors LTX weights under 14 sibling repos.

Sub-repo	Contents	Size
`LTX2.5_checkpoints`	Full checkpoints — REDGraft NSFW int8, plus Joy-LTX / Q3_K_M GGUFs	40.6 GB
`LTX2.5_base`	Base runtime — turbo distill LoRAs (r450/r256/r128), VAEs, upscalers, MSR LoRA, Q4_K_M + Gemma-4 Q5_K_M GGUFs, ComfyUI node zips	45.7 GB
`LTX2.5_actions`	Face-swap (BFS r128/r64), EditAnything v2.0, Alpha Gen, BTS Movie Set, NSFW action LoRAs + bundled workflow JSONs	7.9 GB
`LTX2.5_styles`	3D Animation Style (0.674 GB) — the only weight left in this sub-repo	0.7 GB
`LTX2.3_checkpoints`	10Eros v1.4 Bob int8-ConvRot, PinkCherry v1.8 Q5_KM, DaSiWa DragonLeap Q4_K_M, JoyAI-Echo surgical Q5, INT4-Quality dev, GGUF Companion, plus two `fro99-AVGRank104` averaged LoRA merges (listed below)	122.3 GB
`LTX2.3_base`	LTX-2.3 dev int8-ConvRot, Sulphur dev fp8mixed, Gemma-3 encoder variants, connectors, VAEs, upscalers, ID/OmniNFT/distill LoRAs	187.9 GB
`LTX2.3_Actions`	NSFW action / motion LoRAs, EditAnything v1.1, Best-FaceID, OmniCine	154.9 GB
`LTX2.3_Characters`	Per-celebrity / per-character likeness LoRAs (394 files)	139.3 GB
`LTX2.3_Styles`	Visual-style LoRAs — anime, claymation, cyberpunk, post-apocalyptic, Dispatch, GalaxyAce, furry	25.7 GB
`LTX2.3_Dramabox`	DramaBox audio DiT + audio components (unsplit pair)	8.5 GB
`LTX2.3_workflows`	155 ComfyUI workflow JSONs	0.7 GB
`LTX2.3`	LTX-2.3 LoRAs — lipdub IC-LoRA, NSFW Helper v2.0, NSFW Posing Nude, GalaxyAce, Crisp / Soft Enhance, VBVR video-reasoning, Boudoir, Camera Controls, Stylized PBR Animation, Mystic XXX, Amateur Hour r16, epic_cumshots + helper `.py` scripts and workflow JSONs	12.0 GB
`LTX2_oldstuff`	Legacy LTX-2 (v2.0) LoRAs — 37 weights: NSFW action / pose (Tit-Daddy v2.0, Oral Suite, POV NSFW v3.0, Doggy Style, Sexy Move, Best Breasts v2.0, Pillow Humping, …), face swap (BFS v1 focus head), character (Aerith, Jessica Rabbit, Artemis, Gollum), styles (Dispatch v3, Manga, Anime Flat, GalaxyAce) and IC-LoRA Detailer (+ 3 PDF cards)	29.9 GB
`dramabox-dit-int8`	DramaBox TTS 3.3B DiT — selective INT8 weight-only quant (562 layers) + `config.json` / loader scripts	3.6 GB

The LTX2.5_styles and LTX2.5_actions LoRAs are listed individually under ▣ Styles and ▣ Special; the GGUFs under GGUF Quantized Models; REDGraft under REDGraft (NSFW).

Selected non-LoRA weights worth calling out (all links verified; most are re-packs of models listed from their primary sources elsewhere on this page):

Weights	Precision	Size	Download
`LTX 2.3 dev INT8 ConvRot` (standalone, not in LTX-2.3/Comfy)		23.51 GB
`Diffusion Models - Sulphur_dev_model_fp8mixed`		25.03 GB
`LTX 2.3 INT8 Dev Transformer INT4 Quality`		18.98 GB
`LTX-2.3 GGUF Companion` (transformer-only split for GGUF loading)		9.02 GB
`DramaBox audio DiT v1`		6.58 GB
`DramaBox audio DiT v1 — INT8 selective`		3.62 GB
`DramaBox audio components` (vocoder / Mel-band / TAE bundle)		1.94 GB

◦ Not all mirrors add anything. The 10eros v1.4 Q6_K, DaSiWa DragonLeap v4.0 Q4KM, SexGod PinkCherry v1.8 Q5_KM and JoyAI-Echo × LTX-2.3 Surgical Merge Q5 GGUFs are re-packs of quants already listed from SexGod1979 / DaSiWa / joeygambino / vantagewithai — linked from those sections. LTX2.3_base's Sulphur_dev-Q3_K_M.gguf and LTX-2.3-22b-distilled-Q4_K_M.gguf are likewise re-packs of quants already listed (Abiray / unsloth), so they are not duplicated here.

◦ Re-verified 2026-10-13: every sub-repo above has a single commit (all dated 2026-10-08) — the file counts and totals in this table are unchanged, so nothing new has landed since the audit. The LTX2.5_actions bullet under ▣ Special names its LoRAs individually. 2026-10-27: the three sub-repos added above (LTX2.3, LTX2_oldstuff, dramabox-dit-int8) were previously uncatalogued; their file counts and totals above come from the live tree API.

◦ LTX2.5_base also bundles the ComfyUI node zips it expects (ComfyUI-JoyLTX25, ComfyUI-KJNodes, ComfyUI-LTX2.5-MSR, ComfyUI-Sage-EasyInstall) — the repo is usable as a one-stop 2.5 drop folder.

◦ Not LTX assets. Three files inside LTX2.3_base are not LTX weights and therefore do not belong in the component tables: Diffusion models - MelBandRoformer_fp32.safetensors (0.91 GB, also present byte-for-byte as LTX2.3 MelBandRoformer FP32 - LTX2.3.safetensors) is a generic Mel-band RoFormer audio separator (684 tensors: band_split / layers / mask_estimators, all F32); Model Upscaler nmkdSiaxCX_200k - LTX2.3.safetensors (0.07 GB) is a generic ESRGAN / RRDB pixel upscaler (702 tensors, model.0…8, all F32); and Text encoder - Qwen3.5-sulphur_prompt_enhancer-Q8_0.gguf (9.53 GB) is the Sulphur-2 prompt-enhancer LLM, not an LTX text encoder — it is a re-pack of the official prompt_enhancer/sulphur_prompt_enhancer_model-q8_0.gguf (listed above).

◦ The “Image Upscaler” is a LoRA. LTX Image Upscaler - LTX2.3 - 4psc4l1fy,… (0.81 GB) is — despite the filename — an LTX-2.3 LoRA, not an upscaler model: the header holds 768 BF16 tensors, diffusion_model.transformer_blocks.0…47 with lora_A / lora_B pairs, and the metadata reads modelspec.architecture: LTX2/lora, ss_network_dim: 128, ss_network_alpha: 128, ss_ltx_version: 2.3, base ltx-2.3-22b-dev.safetensors. It is listed under LoRA → Special & Enhancers.

══════════════════════════════════

▣ GGUF Quantized Models

These models are optimized for lower memory usage. Note that in ComfyUI, these are typically loaded as transformer-only models.

QuantStack

QuantStack LTX-2.3

Model	Size	Download
ltx-2.3-22b	12.4 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	14.7 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	14 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	17.8 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	16.7 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	19.4 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	18.5 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	21 GB	dev ┊ distilled ┊ distilled-1.1
ltx-2.3-22b	25.5 GB	dev ┊ distilled ┊ distilled-1.1

QuantStack LTX-2

Model	Quant	Size	Download
LTX-2-dev		8.03 GB
LTX-2-dev		10.3 GB
LTX-2-dev		9.57 GB
LTX-2-dev		13.4 GB
LTX-2-dev		12.3 GB
LTX-2-dev		15 GB
LTX-2-dev		14.2 GB
LTX-2-dev		16.6 GB
LTX-2-dev		21.1 GB

Unsloth LTX-2.5

Unsloth LTX-2.5 GGUF

Model	Size	Download
ltx-2.5-22b	42.03 GB	dev ┊ distilled
ltx-2.5-22b	8.46 GB	dev ┊ distilled
ltx-2.5-22b	11.30 GB	dev ┊ distilled
ltx-2.5-22b	13.58 GB	dev ┊ distilled
ltx-2.5-22b	15.63 GB	dev ┊ distilled
ltx-2.5-22b	17.81 GB	dev ┊ distilled
ltx-2.5-22b	22.78 GB	dev ┊ distilled
ltx-2.5-22b	9.5 GB	dev ┊ distilled
ltx-2.5-22b	12 GB	dev ┊ distilled
ltx-2.5-22b	14 GB	dev ┊ distilled
ltx-2.5-22b	16 GB	dev ┊ distilled

◦ Unsloth Dynamic 2.0 quants: the UD- files are mixed precision (per-layer type from a measured sensitivity scan) and beat the uniform files at the same size — recommended pick is UD-Q4_K_XL (14 GB, clearly better than Q4_K_M, on par with Q5_K_M at 1.6 GB less). BF16 is a lossless repack (bit-identical to the bf16 safetensors through the GGUF path). Same layout as the LTX-2.3 repo: dev at the root, distilled in distilled/. The repo also hosts bf16 dev / distilled transformer safetensors (42.02 GB each) plus a Gemma4 12B text encoder, VAEs and the latent upscaler; text encoder, VAEs and upscaler come from Lightricks/LTX-2.5.

Unsloth LTX-2.3

Unsloth LTX-2.3 GGUF

Model	Size	Download
ltx-2.3-22b	42 GB	dev ┊ distilled
ltx-2.3-22b	42 GB	dev ┊ distilled
ltx-2.3-22b	8.28 GB	dev ┊ distilled
ltx-2.3-22b	10.8 GB	dev ┊ distilled
ltx-2.3-22b	9.95 GB	dev ┊ distilled
ltx-2.3-22b	12.7 GB	dev ┊ distilled
ltx-2.3-22b	13.8 GB	dev ┊ distilled
ltx-2.3-22b	14.3 GB	dev ┊ distilled
ltx-2.3-22b	13.1 GB	dev ┊ distilled
ltx-2.3-22b	15.3 GB	dev ┊ distilled
ltx-2.3-22b	16.3 GB	dev ┊ distilled
ltx-2.3-22b	16.1 GB	dev ┊ distilled
ltx-2.3-22b	15.2 GB	dev ┊ distilled
ltx-2.3-22b	17.8 GB	dev ┊ distilled
ltx-2.3-22b	22.8 GB	dev ┊ distilled
ltx-2.3-22b	9.5 GB	dev ┊ distilled
ltx-2.3-22b	13.5 GB	dev ┊ distilled
ltx-2.3-22b	11.4 GB	dev ┊ distilled
ltx-2.3-22b	16.5 GB	dev ┊ distilled
ltx-2.3-22b	14.2 GB	dev ┊ distilled
ltx-2.3-22b	18.3 GB	dev ┊ distilled
ltx-2.3-22b	16.3 GB	dev ┊ distilled

Unsloth LTX-2.3 GGUF - Distilled 1.1

Model	Size	Download
ltx-2.3-22b	42 GB	distilled-1.1
ltx-2.3-22b	42 GB	distilled-1.1
ltx-2.3-22b	7.94 GB	distilled-1.1
ltx-2.3-22b	10.6 GB	distilled-1.1
ltx-2.3-22b	9.74 GB	distilled-1.1
ltx-2.3-22b	14.2 GB	distilled-1.1
ltx-2.3-22b	13 GB	distilled-1.1
ltx-2.3-22b	15.9 GB	distilled-1.1
ltx-2.3-22b	15 GB	distilled-1.1
ltx-2.3-22b	17.8 GB	distilled-1.1
ltx-2.3-22b	22.8 GB	distilled-1.1
ltx-2.3-22b	10.9 GB	distilled-1.1
ltx-2.3-22b	13.4 GB	distilled-1.1
ltx-2.3-22b	16.4 GB	distilled-1.1
ltx-2.3-22b	18.2 GB	distilled-1.1

⚠️ The distilled-1.1/ folder ships only UD-Q2_K, UD-Q3_K_M, UD-Q4_K_M and UD-Q5_K_M. The UD-Q3_K_S, UD-Q4_K_S and UD-Q5_K_S rungs exist only in the non-1.1 distilled/ folder (listed in the table above).

Unsloth LTX-2

Unsloth LTX-2 GGUF

Model	Quant	Size	Download
ltx-2-19b-dev		37.8 GB
ltx-2-19b-dev		37.8 GB
ltx-2-19b-dev		10.1 GB
ltx-2-19b-dev		11.6 GB
ltx-2-19b-dev		8.1 GB
ltx-2-19b-dev		10.7 GB
ltx-2-19b-dev		10.1 GB
ltx-2-19b-dev		9.47 GB
ltx-2-19b-dev		11.3 GB
ltx-2-19b-dev		12.3 GB
ltx-2-19b-dev		12.8 GB
ltx-2-19b-dev		11.9 GB
ltx-2-19b-dev		13.7 GB
ltx-2-19b-dev		14.6 GB
ltx-2-19b-dev		14.3 GB
ltx-2-19b-dev		13.6 GB
ltx-2-19b-dev		16 GB
ltx-2-19b-dev		20.4 GB

Vantage

Vantage AI GGUFs

Model	Quant	Size	Download
ltx-2-19b-dev		9.96 GB
ltx-2-19b-dev		9.28 GB
ltx-2-19b-dev		11.6 GB
ltx-2-19b-dev		12.4 GB
ltx-2-19b-dev		12.8 GB
ltx-2-19b-dev		11.8 GB
ltx-2-19b-dev		13.6 GB
ltx-2-19b-dev		14.5 GB
ltx-2-19b-dev		14.4 GB
ltx-2-19b-dev		13.5 GB
ltx-2-19b-dev		15.9 GB
ltx-2-19b-dev		20.4 GB
ltx-2-19b-distilled		9.96 GB
ltx-2-19b-distilled		9.28 GB
ltx-2-19b-distilled		11.6 GB
ltx-2-19b-distilled		12.4 GB
ltx-2-19b-distilled		12.8 GB
ltx-2-19b-distilled		11.8 GB
ltx-2-19b-distilled		13.6 GB
ltx-2-19b-distilled		14.5 GB
ltx-2-19b-distilled		14.4 GB
ltx-2-19b-distilled		13.5 GB
ltx-2-19b-distilled		15.9 GB
ltx-2-19b-distilled		20.4 GB

Abiray LTX-2.5 Distilled GGUF

Abiray/LTX-2.5-Distilled-GGUF

Model	Quant	Size	Download
ltx-2.5-22b-distilled		12.92 GB
ltx-2.5-22b-distilled		12.65 GB
ltx-2.5-22b-distilled		15.69 GB
ltx-2.5-22b-distilled		15.33 GB
ltx-2.5-22b-distilled		18.12 GB
ltx-2.5-22b-distilled		18.62 GB
ltx-2.5-22b-distilled		23.60 GB

Abiray LTX-2.5 GGUF

Abiray/LTX-2.5-GGUF

Model	Quant	Size	Download
ltx-2.5-22b		9.89 GB
ltx-2.5-22b		9.07 GB
ltx-2.5-22b		13.21 GB
ltx-2.5-22b		12.07 GB
ltx-2.5-22b		14.83 GB
ltx-2.5-22b		14.01 GB
ltx-2.5-22b		16.55 GB
ltx-2.5-22b		21.19 GB

realrebelai LTX-2.5 GGUFs

realrebelai/LTX-2.5_GGUFs

Model	Quant	Size	Download
ltx-2.5-22b-distilled		8.83 GB
ltx-2.5-22b-distilled		11.53 GB
ltx-2.5-22b-distilled		15.09 GB
ltx-2.5-22b-distilled		13.85 GB
ltx-2.5-22b-distilled		16.82 GB
ltx-2.5-22b-distilled		18.66 GB
ltx-2.5-22b-distilled		23.63 GB

vantagewithai LTX-2.5 GGUF

vantagewithai/LTX-2.5-GGUF

Model	Quant	Size	Download
ltx-2.5-22b-dev		12.13 GB
ltx-2.5-22b-dev		12.92 GB
ltx-2.5-22b-dev		12.65 GB
ltx-2.5-22b-dev		15.24 GB
ltx-2.5-22b-dev		15.53 GB
ltx-2.5-22b-dev		15.69 GB
ltx-2.5-22b-dev		15.33 GB
ltx-2.5-22b-dev		15.98 GB
ltx-2.5-22b-dev		16.26 GB
ltx-2.5-22b-dev		18.12 GB
ltx-2.5-22b-dev		15.89 GB
ltx-2.5-22b-dev		18.62 GB
ltx-2.5-22b-dev		23.60 GB
ltx-2.5-22b-distilled		12.13 GB
ltx-2.5-22b-distilled		12.92 GB
ltx-2.5-22b-distilled		12.65 GB
ltx-2.5-22b-distilled		15.24 GB
ltx-2.5-22b-distilled		15.53 GB
ltx-2.5-22b-distilled		15.69 GB
ltx-2.5-22b-distilled		15.33 GB
ltx-2.5-22b-distilled		15.98 GB
ltx-2.5-22b-distilled		16.26 GB
ltx-2.5-22b-distilled		18.12 GB
ltx-2.5-22b-distilled		15.89 GB
ltx-2.5-22b-distilled		18.62 GB
ltx-2.5-22b-distilled		23.60 GB

Other single-file GGUFs (ApolloRaines, EllaPriest45)

Single-quant GGUFs that don't fit the multi-quant sets above: a merged fine-tune, a paired text encoder, and community re-pack mirrors of quants already listed in Checkpoints.

ApolloRaines/LTX-2.5-22b-OmniGen-v12

Gated auto merge of OmniGen v12 on LTX-2.5 (7 likes, t2v pipeline tag, base_model: Lightricks/LTX-2.5 + quantized). Ships one quant only.

Model	Quant	Size	Download
ltx-2.5-22b-omnigen-v12		18.62 GB

◦ Gated auto — sign in to download; the file tree above is public, the weights are not. Use the tree API (/api/models/ApolloRaines/LTX-2.5-22b-OmniGen-v12/tree/main) to verify filenames on gated repos — a 401 from a resolve link proves nothing about the path (HF checks auth before the filename).

EllaPriest45 single-quant mirrors

Re-pack mirrors of quants that are already listed from their primary sources, plus one paired text encoder and one Joy-LTX derivative.

Model	Quant	Size	Download
`LTX-2.5-Distilled`		15.09 GB
`Gemma-4-12b-with-proj-ltx-2.5` (text encoder)		9.51 GB
`LTX 2.5` (Joy-LTX Distilled derivative)		12.93 GB
`LTX2.5`		10.6 GB

◦ EllaPriest45's LTX repos are split into sub-repos, not one monorepo — see the EllaPriest45 (model mirror) subsection in Finetunes for the full map.

Special Quantization: PolarQuant Q5

LTX-2.3 (22B) — PolarQuant Q5 is a bit-packed quantization method using Hadamard-Rotated Lloyd-Max Quantization. It achieves optimal Gaussian weight quantization via Hadamard rotation, delivering near-lossless quality with significant size reduction.

Specification

Specification	Value
Parameters	22B
Transformer Blocks	48
Hidden Dimension	4096
Layers Quantized	1,347 (of 5,947 total tensors)

Compression Statistics:

Component	Original Size	PQ5 Packed	Reduction
Transformer (1,347 layers)	37 GB	4.6 GB	-88%
VAE + Skip (4,600 layers)	9.1 GB	9.1 GB	BF16 kept
Upscalers	1.3 GB	1.3 GB	BF16 kept
Total	46.2 GB	15 GB	-68%

Quality Metrics:

Cosine Similarity: 0.9986 (near-lossless)
Download Size: 15 GB
Beats torchao INT4 on perplexity (PPL)

Hardware Requirements:

GPU	VRAM	Status
A100 (80 GB)	80 GB	Full speed
A100 (40 GB)	40 GB	Recommended
RTX 4090 (24 GB)	24 GB	With offloading

Key Features:

Mixed precision approach: transformer heavily quantized (-88%) while VAE remains BF16
5-bit bit-packed representation (Q5)
50-65% smaller than original with zero quality loss
One-command setup with easy generation wrapper

Model	Size	Download
`LTX-2.3-22B-PolarQuant-Q5`	15 GB

Installation: pip install safetensors huggingface_hub scipy ArXiv Reference: 2603.29078

Harsha001 ROCm FP4 (AMD)

ROCm-FPX / ROCm-FP4 is an AMD-specific bit-packed FP4 format for stable-diffusion.cpp; the repo carries the sdcpp-rocmfpx.patch that teaches it the format. Both files are built from the uncensored LTX-2.5 dev stack (base_model: Lightricks/LTX-2.5) and the repo is gated - sign in to Hugging Face to download, and verify the paths against the tree API, not a range GET.

File	Precision	Size	Download
`LTX-2.5-22B-dev-Q4_0_ROCMFP4.gguf`		14.74 GB
`gemma4-12b-heretic-with-proj-ltx-2.5-Q8_0_ROCMFPX.gguf`		14.67 GB

◦ The transformer is Q4_0 in ROCm-FP4; the text encoder is Q8_0 in ROCm-FPX, the higher-precision sibling format (the encoder copy is listed in block 40). The repo also keeps the build/ scripts that produced both files and te_delta_stats.json, the encoder-delta report.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇

▓ MLX (Apple Silicon)

MLX conversions are safetensors for Apple Silicon (Metal) - a different runtime family from the CUDA builds above, not a GGUF variant. They load natively through mlx-lm, the ltx-2-mlx Python runtime, ltx-2-mlx-swift and mlx-serve. Every pack is split per component (transformer-*, connector, vae_*, vocoder, audio_vae, spatial_upscaler_*, temporal_upscaler_*, gemma4-12b-ltx-v1/), so you only download what your pipeline actually needs.

Quantized packs keep the non-quantized layers in BF16: the 4-bit / 8-bit payload is stored as packed U32 tensors next to BF16 tensors and F32 scales, so the on-disk size stays well above a pure 4-bit / 8-bit model. The bit width and group size are recorded in each repo's quantize_config.json (bits, group_size, only_transformer_blocks).

mlx-community - LTX-2.5 (bf16)

mlx-community/ltx-2.5-mlx - The reference MLX conversion of Lightricks/LTX-2.5 - full bf16, no quantization. Ships both the dev and the distilled transformer plus the Gemma-4-12b text encoder, the connector, the whole VAE family and both upscalers.

Component	Precision	Size	Download
`transformer-dev.safetensors`		37.99 GB
`transformer-distilled.safetensors`		37.99 GB
`connector.safetensors`		6.34 GB
`gemma4-12b-ltx-v1/model.safetensors`		23.81 GB
`vae_decoder.safetensors`		814.35 MB
`vae_decoder_pruna.safetensors`		690.03 MB
`vae_diffusion_decoder.safetensors`		834.32 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`duration_head.safetensors`		3.81 MB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

mlx-community - LTX-2.5 (q8)

mlx-community/ltx-2.5-mlx-q8 - The q8 text-encoder half of the pack above - it deliberately ships no transformer: pair it with the bf16 transformer-dev / transformer-distilled from mlx-community/ltx-2.5-mlx. The Gemma-4-12b encoder is stored as packed U32 at 8-bit.

Component	Precision	Size	Download
`connector.safetensors`		6.34 GB
`gemma4-12b-ltx-v1/model-00001-of-00003.safetensors`		5.32 GB
`gemma4-12b-ltx-v1/model-00002-of-00003.safetensors`		5.32 GB
`gemma4-12b-ltx-v1/model-00003-of-00003.safetensors`		2.96 GB
`vae_decoder.safetensors`		814.35 MB
`vae_decoder_pruna.safetensors`		690.03 MB
`vae_diffusion_decoder.safetensors`		834.32 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`duration_head.safetensors`		3.81 MB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

dgrauet - LTX-2.3 (bf16)

dgrauet/ltx-2.3-mlx - Reference bf16 MLX conversion of LTX-2.3. Carries the dev, distilled and distilled-1.1 transformers plus both rank-384 distilled LoRAs (384 and 384-1.1).

Component	Precision	Size	Download
`transformer-dev.safetensors`		37.99 GB
`transformer-distilled-1.1.safetensors`		37.99 GB
`transformer-distilled.safetensors`		37.99 GB
`connector.safetensors`		6.34 GB
`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`		7.61 GB
`ltx-2.3-22b-distilled-lora-384.safetensors`		7.61 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

dgrauet - LTX-2.3 (q4)

dgrauet/ltx-2.3-mlx-q4 - 4-bit MLX pack (bits: 4, group_size: 64, transformer blocks only) - the most downloaded MLX conversion of LTX-2.3. Still carries the bf16 rank-384 distilled LoRAs.

Component	Precision	Size	Download
`transformer-dev.safetensors`		11.32 GB
`transformer-distilled-1.1.safetensors`		11.32 GB
`transformer-distilled.safetensors`		11.32 GB
`connector.safetensors`		6.34 GB
`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`		7.61 GB
`ltx-2.3-22b-distilled-lora-384.safetensors`		7.61 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

dgrauet - LTX-2.3 (q8)

dgrauet/ltx-2.3-mlx-q8 - 8-bit MLX pack (bits: 8, group_size: 64, transformer blocks only) - the middle ground between the bf16 and 4-bit packs.

Component	Precision	Size	Download
`transformer-dev.safetensors`		20.60 GB
`transformer-distilled-1.1.safetensors`		20.60 GB
`transformer-distilled.safetensors`		20.60 GB
`connector.safetensors`		6.34 GB
`ltx-2.3-22b-distilled-lora-384-1.1.safetensors`		7.61 GB
`ltx-2.3-22b-distilled-lora-384.safetensors`		7.61 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

ddalcu - LTX-2.5 (q4)

ddalcu/LTX-2.5-MLX-Serve-4bit - Self-contained 4-bit pack for mlx-serve, the native Zig MLX server for Apple Silicon. The text encoder is inside the pack, so nothing else downloads. Ships a split_model.json manifest.

Component	Precision	Size	Download
`transformer-dev.safetensors`		11.32 GB
`transformer-distilled.safetensors`		11.32 GB
`connector.safetensors`		6.34 GB
`gemma4-12b-ltx-v1/model.safetensors`		6.70 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

ddalcu - LTX-2.5 (q8)

ddalcu/LTX-2.5-MLX-Serve-8bit - The 8-bit sibling of the mlx-serve pack - same layout, wider payload, plus the vae_diffusion_decoder.

Component	Precision	Size	Download
`transformer-dev.safetensors`		20.60 GB
`transformer-distilled.safetensors`		20.60 GB
`connector.safetensors`		6.34 GB
`gemma4-12b-ltx-v1/model.safetensors`		12.65 GB
`vae_decoder.safetensors`		814.35 MB
`vae_diffusion_decoder.safetensors`		834.32 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

MrMofer - LTX-2.5 (q8)

MrMofer/ltx-2.5-mlx-q8 - 8-bit MLX conversion of LTX-2.5 with the rank-450 distilled LoRA included. The text encoder lives in a text_encoder/ folder rather than gemma4-12b-ltx-v1/.

Component	Precision	Size	Download
`transformer-dev.safetensors`		20.60 GB
`transformer-distilled.safetensors`		20.60 GB
`connector.safetensors`		6.34 GB
`text_encoder/model.safetensors`		8.15 GB
`ltx-2.5-22b-distilled-lora-450.safetensors`		8.90 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`duration_head.safetensors`		3.81 MB
`spatial_upscaler_x2.safetensors`		995.74 MB
`temporal_upscaler_x2.safetensors`		261.94 MB

notapalindrome - LTX-2.3 (q4)

notapalindrome/ltx23-mlx-av-q4 - 4-bit MLX conversion of LTX-2.3 with joint audio + video (mlx-av). Quantizes the whole model, not just the transformer blocks.

Component	Precision	Size	Download
`transformer.safetensors`		11.35 GB
`connector.safetensors`		6.34 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.88 MB
`vocoder.safetensors`		258.30 MB
`audio_vae.safetensors`		63.84 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.74 MB
`upsampler.safetensors`		995.74 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.94 MB

xocialize - LTX-2.3 (bf16)

xocialize/ltx-2.3-mlx - bf16 MLX conversion of LTX-2.3 carrying the distilled transformer only - the compact option for the ltx-2-mlx Python runtime.

Component	Precision	Size	Download
`transformer-distilled.safetensors`		37.99 GB
`connector.safetensors`		6.34 GB
`vae_decoder.safetensors`		814.35 MB
`vae_decoder_pruna.safetensors`		690.03 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

MLXBits - LTX-2.3 (q4)

MLXBits/ltx-2.3-10eros-v1.3-dmd-mlx-q4 - 4-bit MLX conversion of the 10Eros v1.3 DMD distilled LTX-2.3 merge (bits: 4, group_size: 32).

Component	Precision	Size	Download
`transformer-distilled.safetensors`		12.47 GB
`connector.safetensors`		6.34 GB
`vae_decoder.safetensors`		814.35 MB
`vae_encoder.safetensors`		637.89 MB
`vocoder.safetensors`		258.31 MB
`audio_vae.safetensors`		106.51 MB
`spatial_upscaler_x1_5_v1_0.safetensors`		1.09 GB
`spatial_upscaler_x2_v1_1.safetensors`		995.75 MB
`temporal_upscaler_x2_v1_0.safetensors`		261.95 MB

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ Text Encoders

LTX-2/2.3 require Gemma-3-12b variants with text projection layers. LTX-2.5 ships with a Gemma-4-12B text encoder — see the ▣ Gemma-4-12b section below.

▣ Comfy-Org Optimized Encoders

Official and optimized versions for ComfyUI.

Model Name	Size	Download
`gemma_3_12B_it`	24.4 GB
`gemma_3_12B_it_fpmixed`	13.7 GB
`gemma_3_12B_it_fp8_scaled`	13.2 GB
`gemma_3_12B_it_fp4_mixed`	9.5 GB
`gemma_3_12B_it-int8tensormixed`	13.2 GB
`gemma_3_12B_it-int8mixedblockwise`	13.6 GB
`gemma_3_12B_it-int8mixedtensorwise`	14.1 GB
`gemma_3_12B_it-int8tensormixed`	13.2 GB
`text_projection_fp8`	1.16 GB

gemma_3_12B_it_fpmixed: Experimental quant. Should be better than the fp8 scaled
gemma_3_12B_it_fp4_mixed: 90% fp4 layers Note: mxfp8mixed quantization requires a custom fork of ComfyUI-Kitchen with mxfp8 support. Standard ComfyUI setups don't.

· · · · · · · · · · · · · ·

▣ Gemma-3-12b-it GGUF (stock encoder)

GGUF conversion of the stock google/gemma-3-12b-it — the same weights Comfy-Org ships as gemma_3_12B_it.safetensors — from unsloth/gemma-3-12b-it-GGUF, for ComfyUI-GGUF low-VRAM setups. The text projection stays a separate file: pair any of these with text_projection_fp8 (or ltx-2.3_text_projection_bf16) from the table above. The mmproj-BF16/F16/F32.gguf files in the same repo are the Gemma-3 vision projector and are not needed for LTX.

Quant	Size	Download
	23.54 GB
	6.89 GB
	6.55 GB
	4.77 GB
	6.01 GB
	5.46 GB
	6.91 GB
	7.56 GB
	7.3 GB
	6.94 GB
	8.45 GB
	8.23 GB
	9.66 GB
	12.51 GB
	3.25 GB
	3.06 GB
	4.37 GB
	3.6 GB
	4.83 GB
	4.86 GB
	6.14 GB
	7.43 GB
	8.46 GB
	10.56 GB
	14.39 GB

◦ EllaPriest45 mirror — the same Q4_K_M file is re-packed byte-for-byte (7,300,778,336 B) as Text Encoder - Gemma-3-12b-it-Q4_K_M.gguf inside the EllaPriest45/LTX2.3_base drop folder — no separate download needed.

▣ Gemma-3-12b-it QAT (Lightricks)

The official Gemma-3-12B-it encoder for LTX-2 / LTX-2.3 is the Google quantization-aware-trained release gemma-3-12b-it-qat-q4_0. Lightricks publishes it here as the unquantized checkpoint, so the Q4_0 quantization is applied locally. LTX-2.5 replaced this encoder with Gemma-4-12b (see above).

Model	Precision	Size	Download
`gemma-3-12b-it-qat-q4_0-unquantized`		24.37 GB

◦ Unquantized on purpose - the model card states the checkpoint is unquantized and must be quantized to Q4_0 with your own tool before use. The 5 shards total 24374793024 B (24.37 GB); model.safetensors.index.json maps the parameter names across them. QAT lets the quantized build keep bfloat16-grade quality at a fraction of the memory.

◦ The WanGP LTX-2 Model Pack ships the already-quantized variant (gemma-3-12b-it-qat-q4_0, 24.4 GB) plus a quanto build (13.2 GB).

· · · · · · · · · · · · · ·

▣ Gemma-3-12b Abliterated

Why Choose Abliterated Encoders?

Standard Gemma models often incorporate safety alignment that "sanitizes" or weakens specific concepts within prompt embeddings. Even when the model doesn't explicitly refuse a request, this internal filtering can dilute creative intent. For LTX-2 video generation, using a standard encoder often results in:

Reduced Prompt Adherence: Key stylistic or descriptive terms may be ignored or weakened.
Visual Softening: Visual intensity and fine details are often "muted" to fit generic safety profiles.
Concept Dilution: Complex or niche creative requests are subtly altered, leading to less faithful representations of your vision.

Abliteration bypasses these restrictive alignment layers, allowing the encoder to translate your prompts into embeddings with maximum fidelity. This ensures LTX-2 receives the most accurate and un-filtered instructions possible.

Gemma-3-12b-Abliterated (FusionCow)

Fixed versions of the abliterated Gemma-3-12b-it model by FusionCow, modified specifically for compatibility with LTX-2. The original model

Model	Precision	Size	Download
`Gemma ablit fixed`		23.5 GB
`Gemma ablit fixed`		13.8 GB

Sikaworld1990 Gemma-3-12b Abliterated

NVFP4 quantization variants by Sikaworld1990 optimized for Blackwell GPUs.

Model	Precision	Size	Download
`Gemma-3-12b QAT Abliterated FP4`		12.1 GB
`Gemma-3-12b QAT Abliterated FP4`		8.91 GB
`Gemma-3-12b HereticX Abliterated`		15 GB
`Gemma-3-12b High-Fidelity Abliterated`		14.1 GB

FP4-HF: High-fidelity mixed precision calibration
FP4-Pure: Pure FP4 quantization for maximum compression
HereticX: Uncensored variant with maximum prompt fidelity
High-Fidelity: Optimized for quality with better detail preservation

TenStrip — abliterated Gemma-3-12b with bi-projections

The encoder shipped inside TenStrip/LTX2.3-10Eros — a single fp8mixed file that lives under text_encoders/ in the checkpoint repo itself, so it is versioned with the 10Eros merge rather than released separately.

Model	Precision	Size	Download
`gemma-3-12b-it-ablit-norms-biproj-fp8mixed`		12.78 GB

ablit — abliterated (alignment layers removed) · norms — RMS-norm handling adjusted · biproj — carries the text bi-projection layers that LTX-2/2.3 need, so it is a drop-in replacement for the stock Gemma-3-12b encoder.
Verified by header: 1,299 tensors (model.layers × 1296 + model.embed_tokens + model.norm + spiece_model), dtypes F32 × 336, BF16 × 290, F8_E4M3 × 336, U8 × 337 → fp8mixed.

· · · · · · · · · · · · · ·

▣ Gemma-3-12b IT Heretic

Models by DreamFast. "Heretic" lineage bypasses alignment/restriction layers in the text encoder so LTX-2/2.3 receives the most faithful prompt embeddings. Two upstream versions (v1, v2) plus an AX1Y2JP ultra-uncensored fork and a 3rd-party mradermacher imatrix GGUF re-quant set.

Heretic v1 — DreamFast (bf16 + fp8 + 8 GGUFs)

Safetensors

Model	Precision	Size	Download
`Gemma_3_12B_it Heretic`		23.5 GB
`Gemma_3_12B_it Heretic`		12.8 GB

GGUF

Quant	Size	Download
	23.54 GB
	12.51 GB
	9.7 GB
	8.4 GB
	8.2 GB
	7.3 GB
	6.9 GB
	6.01 GB

Heretic v2 — DreamFast (5 safetensors + 8 GGUFs)

Safetensors

Model	Precision	Size	Download
`gemma-3-12b-it-heretic-v2`		24.38 GB
`gemma-3-12b-it-heretic-v2`		13.21 GB
`gemma-3-12b-it-heretic-v2`		13.22 GB
`gemma-3-12b-it-heretic-v2`		13.56 GB
`gemma-3-12b-it-heretic-v2`		8.32 GB

GGUF

Quant	Size	Download
	23.54 GB
	12.51 GB
	9.66 GB
	8.45 GB
	8.23 GB
	7.3 GB
	6.94 GB
	6.01 GB

AX1Y2JP Ultra-Uncensored Heretic ComfyUI fp8_scaled

Ultra-uncensored fork of the Heretic encoder, fp8-scaled and ComfyUI-ready. Single safetensors by AX1Y2JP.

Model	Precision	Size	Download
`gemma-3-12b-it-heretic` (ultra-uncensored)		13.62 GB

mradermacher imatrix GGUF re-quant of Heretic v2 (24 quants)

Third-party imatrix (importance-matrix) re-quantization of DreamFast's Heretic v2 by mradermacher. Covers the full i1-IQ* and i1-Q*_K* family for low-VRAM use. Sizes verified from the repo file list.

Quant	Size	Download
	3.16 GB
	2.95 GB
	4.31 GB
	4.02 GB
	3.84 GB
	3.53 GB
	5.66 GB
	5.46 GB
	5.21 GB
	4.78 GB
	6.89 GB
	6.55 GB
	4.77 GB
	4.45 GB
	6.48 GB
	6.01 GB
	5.46 GB
	6.91 GB
	7.56 GB
	7.3 GB
	6.94 GB
	8.45 GB
	8.23 GB
	9.66 GB

▣ Gemma-4-12b (LTX-2.5 Text Encoders)

LTX-2.5 replaces the Gemma-3-12B text encoder with a Gemma-4-12B encoder. Community "heretic" / uncensored forks bypass alignment layers for maximum prompt fidelity in downstream video generation.

Model	Precision	Size	Download
`gemma4-12b-heretic-ltx-2.5`		26.26 GB
`Gemma-4-12B-it-uncensored-heretic`		26.26 GB
`Gemma-4-12B-it-uncensored-heretic`		13.17 GB
`gemma4-12b-ltx-2.5-heretic`		15.37 GB
`gemma4-12b-ltx2.5-int4int8-mix`		7.52 GB
`gemma4-12b-with-proj-ltx-2.5`		15.50 GB
`gemma4-12b-with-proj-ltx-2.5`		15.37 GB
`gemma4-12b-with-proj-ltx-2.5`		11.20 GB
`gemma4_12b_ltx25_uncensored`		13.16 GB

◦ Gemma-4-12b GGUF — elix3r low-VRAM quants

ChrisColeTech/gemma4_12b_ltx25_uncensored (GGUF)

ChrisColeTech/LTX-2.5-uncensored-v1.1-FP8 — GGUF quants of the uncensored Gemma-4-12b encoder that ships with the ChrisColeTech Uncensored Turbo v1.1 (LTX-2.5) merge. The LTX text projection is already baked into the weights (the safetensors header carries text_embedding_projection, audio_projector and multi_modal_projector tensors), so no separate projection file is needed for this encoder.

Quant	Size	Download
	7.51 GB
	10.81 GB
	13.97 GB

◦ The matching mmproj-gemma4_12b_ltx25_uncensored-f16.gguf (0.12 GB) is the Gemma-4 vision projector and is not required for LTX video generation.

elix3r/gemma4-12b-with-proj-ltx-2.5-GGUF

elix3r/gemma4-12b-with-proj-ltx-2.5-GGUF — GGUF quants of the LTX-2.5 gemma4-12b text encoder (Q2_K / Q4_K_M / Q5_K_M) for ComfyUI-GGUF loading.

Quant	Size	Download
	5.96 GB
	8.41 GB
	9.51 GB

Harsha001/gemma4-12b-heretic-with-proj-ltx-2.5 (ROCm-FPX GGUF)

Harsha001/LTX-2.5-22B-dev-uncensored-ROCmFPX - the uncensored (heretic / abliterated) Gemma-4-12b encoder with the LTX-2.5 projection baked in, re-quantized to Q8_0 in ROCm-FPX for AMD GPUs. Gated repo (sign in to Hugging Face to download).

Quant	Size	Download
	14.67 GB

◦ GPU-only - the ROCm-FPX format targets AMD devices (and stable-diffusion.cpp); it will not load on the usual CUDA path. The transformer itself is the separate LTX-2.5-22B-dev-Q4_0_ROCMFP4.gguf (14.74 GB), catalogued in block 30.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ Separated Components

Separated LTX2 checkpoint by Kijai and Kijai for LTX-2.3. For alternative way to load the models in Comfy.

▣ Diffusion Models (Transformer Only)

Ver	Name	Size
2.3	`ltx-2.3-22b dev`	42 GB
2.3	`ltx-2.3-22b dev`	23.5 GB
2.3	`ltx-2.3-22b dev`	24.1 GB
2.3	`ltx-2.3-22b dev`	25 GB
2.3	`ltx-2.3-22b distilled`	42 GB
2.3	`ltx-2.3-22b distilled`	23.5 GB
2.3	`ltx-2.3-22b distilled v2`	23.2 GB
2.3	`ltx-2.3-22b distilled`	23.5 GB
2.3	`ltx-2.3-22b distilled` (experimental)	24.1 GB
2.3	`ltx-2.3-22b distilled 1.1`	42 GB
2.3	`ltx-2.3-22b distilled 1.1`	25.2 GB
2.3	`ltx-2.3-22b distilled 1.1` (experimental)	24.1 GB
2.3	`ltx-2.3-22b dev`	21.51 GB
2.3	`ltx-2.3-22b distilled 1.1`	21.51 GB
2.3	`ltx-2.3-22b distilled v3`	25.02 GB

2	`ltx-2-19b dev`	37.8 GB
2	`ltx-2-19b dev`	21.6 GB
2	`ltx-2-19b dev`	14.5 GB
2	`ltx-2-19b distilled`	37.8 GB
2	`ltx-2-19b distilled`	21.6 GB

[!NOTE]
input_scaled additionally have activation scaling, and are set to run with fp8 matmuls on supported hardware (roughly 40xx and later Nvidia GPUs).

▣ VAE (Video & Audio)

Ver	Component	Size	Download
2.5	`Video VAE`	1.47 GB	┊
2.5	`Video VAE (conv)`	1.45 GB
2.5	`Audio VAE`	0.36 GB	┊

⚠️ The LTX-2.5 VAEs used to be mirrored only under ChrisColeTech/LTX-2.5-turbo-GGUF; that repo was renamed to LTX-2.5-uncensored-v1.1-FP8 and the old ltx-2.5-*-vae-bf16.safetensors filenames no longer exist there. The rows above now point at the official Lightricks/LTX-2.5 copies, with the ChrisColeTech files kept as byte-identical mirrors (video VAE 1,472,223,346 B, audio VAE 364,866,540 B). The conv video VAE has no ChrisColeTech mirror. | | | | | | | 2.3 | Video VAE | | 1.45 GB | ┊ | | 2.3 | Cinematic Video VAE | | 1.45 GB | | | 2.3 | Pruna Video VAE | | 1.33 GB | | | 2.3 | TAE (tiny autoencoder) | | 23.53 MB | | | 2.3 | Audio VAE | | 365 MB | ┊ | | | | | | | | 2 | Video VAE | | 2.45 GB | | | 2 | Audio VAE | | 218 MB | |

▣ Embedding Connectors & Text Projection

Ver	Name	Size	Download
2.3	`Embeddings Connectors dev`	2.31 GB	┊
2.3	`Embeddings Connectors distilled`	2.31 GB

2	`Connector dev`	2.86 GB
2	`Connector distilled`	2.86 GB

▣ LTX-2.5 Diffusers Layout (Lightricks)

The official Lightricks/LTX-2.5-Diffusers repo re-serializes the whole LTX-2.5 stack in Diffusers folder layout (<component>/diffusion_pytorch_model*.safetensors), so it loads with from_pretrained() instead of the manual single-file layout used above. Every component is its own folder - download only what the pipeline needs.

Component	Precision	Size	Download
`transformer/` (4-shard dev)		37.98 GB
`transformer/` (8-shard dev)		37.99 GB
`transformer_full/` (4-shard dev)		37.98 GB
`text_encoder/` (5 shards)		23.92 GB
`connectors/`		12.69 GB
`prompt_enhancer/` (3 shards)		10.21 GB
`vae/`		1.45 GB
`audio_vae/`		106.51 MB
`vocoder/`		258.31 MB
`diffusion_decoder/`		834.31 MB
`latent_upsampler/`		995.74 MB
`temporal_latent_upsampler/`		261.94 MB
`duration_head/`		3.81 MB
`ltx-2.5-22b-distilled-lora-450-bf16.safetensors`		9.70 GB

◦ The transformer/ folder ships both shard layouts (4-shard and 8-shard) of the same weights - pick one. transformer_full/ is the full variant. connectors/ likewise ships a single file and a 2-shard split of the same weights. Everything here is bf16, no quantization in this repo.

◦ The root ltx-2.5-22b-distilled-lora-450-bf16.safetensors (9.70 GB) is the Diffusers-layout copy of the rank-450 distilled LoRA. It is a different byte layout from the single-file release in Lightricks/LTX-2.5 (8.90 GB) - the two are not byte-identical - so pick the one your loader expects. See Distilled LoRA.

◦ Budget for ~174.43 GB to mirror the whole repo, or a little over 38 GB for a single transformer.

▣ LTX-2.5 Diffusers Layout (Rin247 MECHA Forge quants)

Community quantizations of the Diffusers-layout LTX-2.5 stack by Rin247, forged with MECHA Forge (Model Engineering & Contextual Hearth Assembly). Four weight-only precisions - FP8 / INT8 / FP4 / INT4 - across the denoiser, the text encoder and every supporting component.

Component	Precision	Size	Download
`transformer/`		11.33 GB
`transformer/`		19.46 GB
`transformer/`		12.49 GB
`transformer/`		19.46 GB
`transformer_full/`		11.32 GB
`transformer_full/`		19.43 GB
`transformer_full/`		12.48 GB
`transformer_full/`		19.43 GB
`text_encoder/`		6.74 GB
`text_encoder/`		11.98 GB
`text_encoder/`		7.49 GB
`text_encoder/`		11.98 GB
`vae/`		726.18 MB
`vae/`		726.18 MB
`latent_upsampler/`		995.74 MB
`latent_upsampler/`		724.43 MB
`latent_upsampler/`		995.74 MB
`latent_upsampler/`		724.43 MB
`temporal_latent_upsampler/`		261.94 MB
`temporal_latent_upsampler/`		187.63 MB
`temporal_latent_upsampler/`		261.94 MB
`temporal_latent_upsampler/`		187.63 MB
`diffusion_decoder/`		236.32 MB
`diffusion_decoder/`		419.38 MB
`diffusion_decoder/`		262.32 MB
`diffusion_decoder/`		419.38 MB

◦ The VAE is shared between formats: FP4 reuses the FP8 VAE file and INT4 reuses the INT8 one, so only two VAE files exist for the four precisions. The card also lists connectors/, audio_vae/, duration_head/, prompt_enhancer/ and vocoder/ as part of the package - those components are absent from the repo tree.

◦ ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors (327.32 MB) is the unquantized bf16 Pixel-Spatial-Upscaler IC-LoRA, byte-identical to the file already listed under Enhance / Upscale - no need to download it twice.

◦ Recommended settings from the source card: 30 steps, CFG 3.0, resolutions under 720x1280 and up to 257 frames, with elaborate English prompts.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ LoRA

▣ Styles

Look-and-feel LoRAs — style, medium, period and concept adapters that give the model a distinctive aesthetic. (Ver = target LTX version: 2 / 2.3 / 2.5.)

LoRA	Ver	Size	Description
3D Animation Style (EllaPriest45)	2.5	0.67 GB	3D-rendered animation look (t2v 0.8 str / i2v 0.4 str per filename) — now the only weight in `LTX2.5_styles`
90s Animation Style	2.3	0.35 GB	90s Saturday-morning cartoon look
Abercrom-ME IC-LoRA	2.3	0.40 GB	Vintage A&F Quarterly americana style transfer — monochrome / sun-drenched beach / locker-room makeover (image-conditioned, 2000 steps; bundled workflow)
Animatediff V1	2	0.86 GB	AnimateDiff-style motion (step 2000)
Anime2Half-Real	2.3	0.65 GB	Anime → half-realistic conversion (rank 64, 4500 steps)
Arnold Style	2.3	2.70 GB	Schwarzenegger action-hero look — "Get to the choppa!"
Atomic Explosion	2	0.43 GB	Nuclear-blast VFX (step 2000)
Big Anime Breasts	2	1.11 GB	Anime anatomy exaggeration
Black Venom	2	0.65 GB	Dark comic-book antihero look
Body Positivity	2.3	0.33 GB	Body-inclusive realism (rank 32, step 2750)
BTS Movie Set	2.5	0.34 GB	Behind-the-scenes movie-set atmosphere — rigs, cranes, crews, green screens (T2V + I2V)
Cakeify	2	0.83 GB	Turns subjects into cake (v2)
Clay Mation Style	2.3	0.35 GB	Clay-animation look
Clay Stop Motion	2	3.42 GB	Clay stop-motion look (trigger `NGTVSPC`, step 7500)
ClayMationStyle	2.3	—	Clay animation style (gated repo)
Cozy Felt Style	2.3	0.35 GB	Cozy felt-puppet look
CRT Animation Terminal	2.3	0.43 GB	CRT-monitor look — scanlines, phosphor glow, chromatic aberration, dithering (trigger `crtanim,`; 4000 / 10000-step checkpoints)
Crisp Enhance Style	2.3	0.71 GB	Crisp detail-enhance look
Eat	2	1.11 GB	Eating animation (v2)
Fable 5 — Vintage Style	2.3	0.43 GB	Vintage minimal-illustration / stop-motion / typography (Fable 5 intro; 12 fp8 + comfy checkpoints, 5.14 GB total)
Fantasy Anime Style	2.3	0.35 GB	Fantasy anime look
Fantasy Painterly Style	2.3	0.35 GB	Fantasy painterly look
Fantasy Puppet Style	2.3	0.35 GB	Fantasy puppet look
Fantasy Realism Style	2.3	0.35 GB	Fantasy realism look
Foot Model	2.3	0.20 GB	Foot / lower-limb detail
Gantz O	2	0.65 GB	Movie-style look (10 000 steps)
Googly Eyes	2.3	0.33 GB	Cartoon googly-eyes look (rank 32, step 3000)
Hydraulic press	2	0.64 GB	Crushes objects flat (step 864)
Kawajiri Retro Anime	2	0.31 GB	Yoshiaki Kawajiri retro-anime style
Latin Dance	2.3	0.21 GB	Latin-dance motion / style (rank ~128)
Levmar lora	2.3	0.43 GB	I2V realism fine-tuned from `ltx-2.3-22b-dev` with the LTX LoRA Trainer (6000 steps, LR 1e-4; 2000 / 6000-step checkpoints; license: other)
LTX-2.3 Styles (EllaPriest45 collection)	2.3	25 GB	Style collection — anime, claymation, cyberpunk, post-apocalyptic, Dispatch, GalaxyAce, furry (42 weights)
LTXreal	2.3	0.33 GB	Realism stylization (trigger `realism`)
Luxe Sensual Style	2.3	0.35 GB	Luxe sensual look
Marbling / Ebru	2.3	0.34 GB	Ink-flow marbling / ebru fluid art (trigger `srx_ebrumotion`, 0.85 str)
MtF Transformation	2.3	1.35 GB	Male-to-female transformation style (v2.1)
Outfit Switch	2	0.43 GB	Outfit change (V2, step 2000)
Paper Cut Out Style	2.3	0.35 GB	Paper cut-out look
Pixar Toon Style (Scrappy-Doo)	2.3	0.35 GB	Pixar-style CGI toon look (trigger `P1x4r`; repo renamed from Andro0s)
Pixar Toon Style (vrgamedevgirl84)	2.3	0.35 GB	Pixar-style CGI toon look
PixelArt UME	2	—	Pixel-art style (gated repo)
POP! Inflatable Animation	2.3	0.65 GB	Comical inflate-and-pop into confetti and fabric scraps (I2V focused)
Post Apocalyptic Style	2.3	0.35 GB	Post-apocalyptic look
Product Ad Style	2.3	0.34 GB	Cinematic product-commercial look — luxury / cosmetics / tech / fashion ads (trigger `srx_commercial`, 0.85 str)
PUSH TO GLASS	2	0.86 GB	Push-through-glass transition VFX (steps 1500 / 2000)
Real2Anime	2.3	0.65 GB	Real → anime conversion (rank 64, 5000 steps)
ReStyle IC-LoRA	2.3	0.81 GB	Image-guided style transfer — re-renders video in a target style while preserving content and motion (step 8000)
Sci-Fi Cinema	2.3	0.34 GB	Cinematic science-fiction look (trigger `srx_scifilm`)
Soft Enhance Style	2.3	0.35 GB	Soft detail-enhance look
SPROUT	2	3.42 GB	Plants / foliage growing from objects (trigger `SPROUT`, step 2000)
Star Trek TNG Style	2.3	2.70 GB	Star Trek: The Next Generation look
Surfaces Realism	2.3	0.35 GB	Materials realism — kills the "plastic" look: wet asphalt, concrete pores, scuffs, wood grain (unconditional, no trigger; rank 32)
Squish – One Hand Only	2	1.11 GB	One-handed squish animation
Turkish Cinema (Yeşilçam)	2.3	0.33 GB	Nostalgic Yeşilçam Turkish-cinema look — warm tones, classic film atmosphere (0.7 str)
UE5 Style	2.3	—	Unreal Engine 5 cinematic look (trigger `ue5_style`; gated repo)
WHATUSEE	2	0.86 GB	"What do you see" transition — zooms into the eye, then pulls deep (trigger `WHATUSEE`)
Wild West Style	2.3	0.71 GB	Wild West look

◦ Gated repos (sign in to Hugging Face to download; sizes unverified): OmerHagawa PixelArt, SOLRICKS UE5 Style, vrgamedevgirl84 ClayMationStyle.

◦ EllaPriest45's LTX-2.3 / LTX-2.5 weights are split across sub-repos — the collections are indexed in block 20 under ❖ EllaPriest45 (model mirror), and the individual adapters are filed in the tables matching their type (❖ Character / Face-Swap, ❖ IC-LoRA · Edit, ❖ Motion). The 10Eros / Sulphur-2 character and action LoRAs live in LTX2.3_Characters / LTX2.3_Actions, not here.

◦ The Alissonerdx LTX-LoRAs repo also ships EditAnything Global and Inpaint Masked builds — filed under ❖ IC-LoRA · Edit / ❖ IC-LoRA · Outpaint.

◦ The SyFeee Chinese-drama set (character, canny / depth / pose IC-LoRAs, ID, plain-AV) is filed under ❖ Character / Face-Swap, ❖ IC-LoRA · Control and ❖ Audio / Foley respectively.

▣ Special & Enhancers

Combined table of enhancer, special, control, audio, camera, restoration and pipeline LoRAs, grouped into sub-tables by Type. (Ver = target LTX version: 2 / 2.3 / 2.5.)

❖ IC-LoRA · Control

LoRA	Ver	Size	Description	Download
3D Render to Photoreal IC-LoRA	2.3	0.33 GB	3D viewport to photoreal video (trigger `3DREAL`; `3DREAL-strong-v2` 0.33 GB, `3DREAL-light` 0.16 GB)
Any-Trajectory-Instruction IC-LoRA	2.3	0.33 GB	ATI spline-trajectory control
BBox-Control IC-LoRA	2.5	0.33 GB	Bounding-box / regional prompting control (rank 32, step 3000; workflow + node pack in repo)
Canny Control IC-LoRA	2	—	Canny edge conditioning
Chinese Drama IC-LoRA — Canny	2.3	1.71 GB	Canny edge / line-art conditioning — composition and architectural layout from reference outlines (step 6000; stack with the char LoRA)
Chinese Drama IC-LoRA — Depth	2.3	1.71 GB	Monocular depth conditioning (Depth-Anything-3) — transfers 3D structure, strong on interior scenes
Chinese Drama IC-LoRA — Pose	2.3	1.71 GB	DWPose 133-keypoint skeleton conditioning — transfers pose while preserving the drama style (step 6000)
Depth Control IC-LoRA	2	—	Depth-map conditioning
Depth Control IC-LoRA (RunningHub mirror)	2	0.65 GB	Depth-map conditioning (624 MiB single-file mirror)
Greenscreen Avatar IC-LoRA	2.3	—	Greenscreen avatar compositing
IC luminance map	2	—	Luminance / particle-map control
IC-LoRA-Cameraman v1	2.3	—	Cameraman framing control
IC-LoRA-Cameraman v2	2.3	—	Cameraman framing control
IC-LoRA-CrossView Prompt v0.9	2.3	—	Cross-view prompt warp
IC-LoRA-CrossView Warp v0.9	2.3	0.19 GB	Cross-view spatial warp
IC-LoRA-CrossView Warp v2	2.3	0.33 GB	Cross-view spatial warp
Ingredients IC-LoRA (Comfy-Org / Lightricks)	2.3	1.31 GB	Ingredient / multi-subject scene	┊
Ingredients IC-LoRA (Lightricks)	2.5	1.31 GB	Multi-subject / reference-sheet scene
Ingredients-Multishot IC-LoRA	2.3	—	Multi-shot ingredient scenes
LipDub IC-LoRA	2.3	2.47 GB	Lip-sync dubbing control
ltx2-compile-keytest IC-LoRA	2.3	—	Compilation / keyframe test
MergeGreen IC-LoRA	2.3	—	Green-screen merge control
Motion Track Control IC-LoRA	2.3	—	Motion-track conditioning
Pose Control IC-LoRA	2	—	Pose-skeleton conditioning
Realisdance IC-LoRA	2.3	0.33 GB	Dance pose / motion control
SAM3D Body v42 IC-LoRA	2.3	0.65 GB	3D body segmentation control
Staging IC-LoRA 512	2.3	—	Scene staging / layout control
SYSTMS FLW IC-LoRA	2.3	—	Flow / curve conditioning
Union Control IC-LoRA	2.3	—	Unified multi-condition control
Union Control IC-LoRA (LTX-2 19b)	2	0.65 GB	Unified multi-condition control for LTX-2 19b (ref 0.5 variant)

◦ LipDub is published under two repo ids - Lightricks/LTX-2.3-22b-IC-LoRA-LipDub and Lightricks/LTX-2.3-22b-IC-LoRA-DubIt are the same repository (identical last-modified timestamp, download count and the same single weight). The file on disk is ltx-2.3-22b-ic-lora-dubit-0.9.safetensors even though the card names a lipdub file - trust the tree, not the card. The Comfy-Org/ltx-2.3 split carries a byte-identical copy under split_files/loras/.

❖ IC-LoRA · Restore

LoRA	Ver	Size	Description
Black Magic IC-LoRA	2.3	0.33 GB	General restoration cleanup
Clean-Plate IC-LoRA	2.3	0.31 GB	Clean-plate object removal
Clean-Plate IC-LoRA	2.5	0.33 GB	Clean-plate object removal
Colorization IC-LoRA	2.3	0.91 GB	B&W to color
Colorization IC-LoRA	2.5	0.91 GB	B&W to color
DeArchive LTX-2.3	2.3	—	Archive-quality restoration
Deblur IC-LoRA	2.3	0.91 GB	Deblur / sharpen
Deblur IC-LoRA	2.5	0.91 GB	Deblur / sharpen
Decompression IC-LoRA	2.3	0.91 GB	Remove compression artifacts
Decompression IC-LoRA	2.5	0.91 GB	Remove compression artifacts
IC-LoRA-Colorizer	2.3	0.33 GB	Add color to footage
IC-LoRA-Deinterlace	2.3	—	Deinterlace video
IC-LoRA-MotionDeblur	2.3	—	Remove motion blur
IC-LoRA-ReFocus	2.3	—	Refocus / depth-of-field
IC-LoRA-Uncompress	2.3	—	Decompress artifacts
Instant-Shave IC-LoRA	2.3	0.65 GB	Remove facial hair
LensRemover IC-LoRA	2.3	0.33 GB	Remove lens / glass artifacts
LTX-2 IC-LoRA-Ungrade	2	—	Upscale / ungrade (LTX-2)
LTX-2.3 IC-LoRA-Ungrade	2.3	—	Upscale / ungrade (LTX-2.3)
Restore IC-LoRA	2.5	1.71 GB	Archival footage restoration (v2v) — decayed/damaged historical video
RoadReady IC-LoRA	2.3	0.33 GB	Road-surface cleanup (removes cracks/stains; LTX LoRA Jam)

❖ IC-LoRA · Outpaint

LoRA	Ver	Size	Description
IC-LoRA-Outpaint	2.3	—	Extend frame borders
In-Outpainting IC-LoRA	2.3	1.31 GB	Inpaint + outpaint
Inpaint Masked R2V	2.3	0.33 GB	Masked region-based inpainting for reference-to-video (rank 32, 3000 steps)
Inpaint Masked T2V — 2500 steps	2.3	1.31 GB	Masked region-based inpainting for text-to-video (rank 128)
Inpaint Masked T2V — 10000 steps	2.3	1.31 GB	Masked region-based inpainting for text-to-video (rank 128)
VR-360-Outpaint IC-LoRA	2.3	—	360 deg VR outpainting

❖ IC-LoRA · Relight

LoRA	Ver	Size	Description
Day-To-Night IC-LoRA	2.3	0.33 GB	Day to night relight
Day-To-Night IC-LoRA	2.5	0.33 GB	Day to night relight
Golden Hour IC-LoRA	2.3	0.33 GB	Golden-hour lighting
Golden Hour IC-LoRA	2.5	0.33 GB	Golden-hour relighting (v2v) — trigger `ADD GOLDEN HOUR`; rank 32, default release is step 500
Relight IC-LoRA	2.3	0.33 GB	Relight scene lighting
Relight LoRA (step 3000)	2.5	0.33 GB	Reference-conditioned relight (rank 32; steps 500–3000 in /checkpoints)

◦ The 2.5 Golden Hour release ships four extra checkpoints for a warmer look — checkpoints/lora_weights_step_00250 (milder) to …_step_01000 (warmest); the root file is the recommended step-500 default. The 2.3 adapter stays on its own repo.

❖ IC-LoRA · Edit

LoRA	Ver	Size	Description
EditAnything Global — 6000 steps (prodigy)	2.3	1.31 GB	Global object editing (rank 128, 6000 steps)
EditAnything Global — 9000 steps (adamw)	2.3	1.31 GB	Global object editing (rank 128, 9000 steps)
EditAnything v2.0 (EllaPriest45 mirror)	2.5	2.47 GB	Remove / replace / add / restyle objects or people (ships its own workflow JSON)
IC-LoRA-EditRefVid v1	2.3	—	Reference-video editing
LTX2.3-ICEdit-Insight	2.3	—	Insight-guided editing
Singularity OmniCine Preview 0.1	2.3	—	Cinematic edit pipeline (preview)
Singularity OmniCine V1	2.3	2.57 GB	Cinematic edit pipeline

❖ IC-LoRA · Effect

LoRA	Ver	Size	Description
Alpha-Gen IC-LoRA	2.5	1.31 GB	Alpha-matte generation / background removal (v2v) — hair, smoke, fur, fire, sheer fabric
Alpha Gen (EllaPriest45 mirror)	2.5	1.31 GB	Alpha-matte generation — filename caps it at 145 frames / 1920×1080 max, 1 str; ships its own workflow JSON
Cel-Character IC-LoRA	2.5	0.33 GB	Live-action → 2D cel character (rank 32; bg stays photographic)
Cinemagraph LoRA	2.3	0.20 GB	Selective-motion cinemagraph (i2v)
Cinemagraph LoRA	2.5	0.20 GB	Selective-motion cinemagraph (i2v)
Cross-Eyed (stereo) IC-LoRA	2.3	0.33 GB	Stereoscopic / cross-eyed
Exploded-View XPLDV LoRA	2.5	0.40 GB	Product exploded-view disassembly (i2v, v5)
FXIC LTX2 IC-LoRA	2	—	VFX / FX effect control
GameTwirl turntable LoRA	2.5	0.43 GB	i2v turntable — spins a game asset 360° in place, camera fixed (trigger `gametwirl`; step 3000)
IC-LoRA-UI (linoyts)	2.3	—	UI / screen overlay
Sandstorm IC-LoRA	2.5	0.33 GB	Sandstorm / dust-haze weather effect (v2v) — trigger `ADD SANDSTORM` plus a density line (light dust haze → near whiteout); rank 32, step 2000
Seamless-Equirectangular LoRA	2.3	—	Seamless 360 deg equirect
TTM IC-LoRA	2.3	—	Texture / transition morph
Water-Simulation IC-LoRA	2.3	0.91 GB	Water simulation effect
Water-Simulation IC-LoRA	2.5	0.91 GB	Water simulation effect

◦ The Sandstorm IC-LoRA is a v2v adapter: feed a clear clip in as the reference, keep the LoRA on the stage that still sees it and pass --skip-stage-2 so it runs at its native 960×544. Density is set by the prompt, not by the LoRA strength — keep strength at 1.0.

❖ Camera

LoRA	Ver	Size	Description
Camera Control: dolly-in (LTX-2)	2	—	Dolly-in move
Camera Control: dolly-left (LTX-2)	2	—	Dolly-left move
Camera Control: dolly-out (LTX-2)	2	—	Dolly-out move
Camera Control: dolly-right (LTX-2)	2	—	Dolly-right move
Camera Control: jib-down (LTX-2)	2	—	Jib-down move
Camera Control: jib-up (LTX-2)	2	—	Jib-up move
Camera Control: static (LTX-2)	2	—	Locked / static camera
EarthZoomOut	2	0.86 GB	Zoom-out from the subject all the way to outer space (trigger `EARTHZO`; steps 1000–2000)
Formula1 Cockpit T-Cam LoRA	2.3	—	F1 onboard T-cam view
FPV motion LoRA	2.3	0.31 GB	Smooth FPV drone motion
HERO CAM	2	0.86 GB	Hero-cam framing (steps 1500 / 2000)
Layout-to-Render IC-LoRA	2.5	1.31 GB	Render a shot from a layout / blocking sketch (v2v, `camera-control`); ships its own two-stage distilled workflow
Object POV	2	0.86 GB	Object point-of-view shot (step 2000)
ReelBids crane-down camera LoRA	2.5	0.20 GB	Crane-down camera move (steps 100–2000)
ReelBids crane-down camera LoRA v2	2.5	0.20 GB	Crane-down camera move (v2, with dataset samples; steps 100–2000)
ReelBids crane-up camera LoRA	2.5	0.20 GB	Crane-up camera move (steps 100–2000)
ReelBids dolly-in camera LoRA	2.5	0.20 GB	Dolly-in (speed token)
ReelBids orbit camera LoRA	2.5	0.20 GB	Orbit camera move (steps 100–2000)
ReelBids orbit camera LoRA v4mixed	2.5	0.20 GB	Orbit camera move (v4 mixed blend; steps 100–2000)

❖ Audio / Foley

LoRA	Ver	Size	Description
akuspace-ltx25 v0.5	2.5	0.16 GB	Audio-reactive space FX
American Accent (audio) LoRA	2.3	0.18 GB	American-accent TTS
AVC-Reward LoRA	2	1.23 GB	Audio-video consistency RL (reward-tuned)
Audio-Only-Context IC-LoRA (audio-only)	2.3	0.16 GB	Audio-only context gen
Audio-Only-Context IC-LoRA (cross-modal)	2.3	0.29 GB	Audio + video context gen
DJ LoRA	2.3	1.71 GB	DJ / mixing visuals
DramaBox TTS IC-LoRA (r128)	2.3	0.67 GB	Expressive TTS extracted from Resemble AI DramaBox (audio branch; r64 also in repo)
DramaBox TTS IC-LoRA (r64)	2.3	0.34 GB	Expressive TTS extracted from Resemble AI DramaBox (audio branch; smaller rank)
Foley LoRA	2.3	—	Foley sound synthesis
Foley-V2A LoRA	2.3	0.22 GB	Video to audio (foley)
Guitar LoRA	2.3	1.71 GB	Guitar-playing visuals
JUST-DUB-IT (LipDubbing)	2	4.93 GB	Lip-sync dubbing
OmniNFT RL-LoRA (Kijai mirror)	2.3	0.62 GB	Audio-driven RL (mirror)
OmniNFT RL-LoRA (LTX-2)	2	1.23 GB	Audio-driven RL (LTX-2)
OmniNFT RL-LoRA (LTX-2.3)	2.3	1.23 GB	Audio-driven RL (LTX-2.3)
Sax (audio) LoRA	2	0.40 GB	Saxophone-playing visuals
SyFe Plain-AV LoRA	2.3	0.86 GB	Plain audio/video LoRA — six checkpoints: production v1 / v2 (0.86 GB), bilingual and cuval builds (0.43 GB), a tv run (0.86 GB) and a unified v5 (0.11 GB)
Violin (8s) LoRA	2.3	0.86 GB	Violin-playing visuals

◦ AVC-Reward LoRA (Tsuki0512) is the LTX-2 dev adapter optimized against AVC-Reward, an event-centric audio-video consistency reward model (arXiv:2610.22321 · project page · code) that scores event-level temporal synchronization and attribute alignment between the audio and video tracks. Research usage: merge into ltx-2-19b-dev.safetensors with the repo's merge_lora.py, then run the author's inference script — not a drop-in ComfyUI LoRA.

❖ I2V Adapter

LoRA	Ver	Size	Description
Image-to-Video Adapter LoRA	2	—	Image to video adapter
Image-to-Video Adapter LoRA (siraxe variant)	2	2.62 GB	I2V adapter (mirror)
VBVR-lora-I2V	2.3	—	VBVR image to video
VBVR-lora-I2V Special	2.3	—	VBVR I2V (special)

❖ Enhance / Upscale

LoRA	Ver	Size	Description
CQ Enhancer — images (r128)	2.5	1.32 GB	Image enhancer (r128)
CQ Enhancer — videos (r128)	2.5	1.32 GB	Video enhancer (r128)
CQ Enhancer V2 (video)	2.5	1.34 GB	Generative video enhance/restore — more natural, ghost-free (use the V2 workflow)
CQ Enhancer V2 (video) — RunningHub mirror	2.5	1.34 GB	Same V2 video weights as above (1,278 MiB), mirrored by RunningHub for one-click online runs
CrispLora	2.3	—	Sharpen / detail boost
Detailer IC-LoRA	2	—	Add fine detail
Enhancer1	2.3	—	General enhancer
Enhancer2	2.3	—	General enhancer
HDR IC-LoRA	2.3	—	HDR tone mapping
LTX Image Upscaler (upscalify IC5)	2.3	0.81 GB	Conceptual upscale LoRA for blurry / pixelated stills → up to 1920×1920 — trigger `4psc4l1fy, a blurry pixelated image transitions into a static sharp high quality image` (rank 128; the bundled workflow is strongly recommended)
Pixel-Spatial-Upscaler IC-LoRA	2.3	0.65 GB	Spatial upscale x2
Pixel-Spatial-Upscaler IC-LoRA	2.3	0.65 GB	Spatial upscale x4
Pixel-Spatial-Upscaler IC-LoRA	2.5	327.32 MB	Spatial upscale x2
Refine-Details IC-LoRA	2.5	1.31 GB	Detail refinement / micro-contrast boost (v2v, upscaling)
SDR-To-HDR IC-LoRA	2.5	1.31 GB	SDR → HDR conversion (v2v) — paper arXiv:2604.11788
Skin-Hair LoRA	2.3	—	Skin / hair refinement
Upscale IC-LoRA	2.3	—	Upscale resolution

◦ The SDR-To-HDR IC-LoRA ships a small companion scene embedding that must be loaded alongside the main weights: ltx-2.5-22b-ic-lora-sdr-to-hdr-scene-emb.safetensors (12.61 MB).

◦ The LTX Image Upscaler LoRA ships with its own workflow — LTX Image Upscaler - LTX2.3 - 4psc4l1fy,….json (258 KB), mirrored from the author's EllaPriest45/LTX2.3_base drop folder (no primary Hugging Face repo exists for this LoRA).

❖ Character / Face-Swap

LoRA	Ver	Size	Description
Arcane-Jinx v1	2.3	1.71 GB	Jinx character design inspired by Arcane (step 21000)
Bender (Futurama)	2.3	0.67 GB	Bender character LoRA — triggers `2d animation`, `Bender the bending robot`
Best Face Swap — bernini r64 high-noise (WAN22)	2	0.61 GB	Face / head swap (WAN22)
Best Face Swap — bernini r64 low-noise (WAN22)	2	0.61 GB	Face / head swap (WAN22)
Best Face Swap — head_swap_v1 13500 ff (LTX-2)	2	1.31 GB	Head swap (first frame)
Best Face Swap — head_swap_v1 8750 flf (LTX-2)	2	1.31 GB	Head swap (first+last)
Best Face Swap — head_swap_v2 multimodes (LTX-2)	2	1.32 GB	Head swap (multi-mode)
Best Face Swap — head_swap_v3 r-adaptive (LTX-2.3)	2.3	1.36 GB	Head swap (adaptive rank)
Best Face Swap — head_swap_v3 r64 (LTX-2.3)	2.3	0.65 GB	Head swap (rank 64)
Best Face Swap R128 (EllaPriest45 mirror)	2.5	0.65 GB	Face swap (rank 128; R64 also in repo)
Best Face Swap R64 (EllaPriest45 mirror)	2.5	0.65 GB	Face swap (rank 64; R128 also in repo)
Donald Trump	2	0.62 GB	Donald Trump character likeness
Euni (portrait, cometa21)	2.5	0.43 GB	Character identity — Euni portrait LoRA (rank 32, 1500 steps, trigger `euni`)
GroupPhoto	2	0.86 GB	Multi-person group-photo composition — other people gather around an input subject (trigger `GR0UP`, I2V)
Homelander	2.3	1.35 GB	The Boys' Homelander character IC-LoRA — trigger `HeroHomelander` (+ optional `wearing red leather gloves` / `dark blue cape and ornate red-and-gold collar`); rank-64 v1 / rank-32 v2 (0.67 GB), 1500 steps; occasional letterbox bars
Iara v2 (identity)	2.5	0.62 GB	Character identity — Iara v2 (step 1000; 750 also in repo)
Jesse (identity, trial)	2.3	0.32 GB	Experimental character identity — Jesse (rank 32, 100-step trial on 15 stills, trigger `jesseperson`; unvalidated, no voice)
Jenna Coleman	2.3	1.35 GB	Jenna Coleman character LoRA (rank 64)
Kermit the Frog	2.3	0.67 GB	Kermit the Frog character LoRA (v2.12)
Licon MSR V1 (LTX-2.5, Multiple Subject Ref)	2.5	1.31 GB	Multi-subject reference
Licon MSR V2 (LTX-2.5, Multiple Subject Ref)	2.5	2.21 GB	Multi-subject reference (v2; V1 + per-scene validation sets also in repo)
Licon MSR V2 (Multiple Subject Ref)	2.3	—	Multi-subject reference
LTX-2.3 Dual-Character LoRA	2.3	—	Two-character scene
LTX2.3_Characters (EllaPriest45 collection)	2.3	139 GB	Per-celebrity / per-character likeness collection (394 weights; mirrors indexed in block 20)
Marge Simpson	2.3	0.67 GB	Marge Simpson character LoRA trained on LTX-2.3 Dev — triggers `2d animation`, `Marge Simpson` (use `2d animation` to avoid 3D output)
Rick and Morty v2	2.3	1.35 GB	Rick and Morty animated style (v2, trained on the Smith and Rick Sanchez characters; 6000-step checkpoint also in repo — the older BETA repo now redirects here)
SyFeee Chinese Drama char LoRA	2.3	0.86 GB	Historical-drama character trained on 78 episodes — live-action photoreal Han-dynasty / wuxia generation; triggers `char_0_person`–`char_9_person`, 0.8–0.9 str (step 3000)
SyFeee ID-LoRA	2.3	1.16 GB	Identity preservation across frames — two checkpoints (`id_lora_ours_768` step 3000 / `id_lora_ours_704` step 6000)
Tessy (identity)	2.5	0.43 GB	Character identity — Tessy (step 1500; 250–1250 also in repo)
Tessy (identity, cometa21)	2.5	0.86 GB	Character identity — Tessy in TV / podcast settings (rank 64, 500 steps, trigger `tessy`)
Wednesday Addams	2.3	0.67 GB	Wednesday Addams character LoRA

❖ Motion

LoRA	Ver	Size	Description	Download
Better NSFW motion v2.0 vSulphur	2.3	0.65 GB	NSFW motion/physics boost (Sulphur-2 tuned)
Audio-Reactive LoRA	2.3	—	Audio-reactive motion
Audio-Reactive LoRA v2 (RunningHub mirror)	2.3	1.35 GB	Audio-reactive motion (v2)
CQ Freeform Motion Transfer LoRA	2.5	0.81 GB	Transfer any motion to a different character or object — prompt strongly affects the result; consistency fades past ~30 s	┊
EllaPriest45 LTX2.3_Actions (collection)	2.3	154 GB	Huge NSFW action / motion / pose collection (188 weights), plus EditAnything v1.1 (2-stage), Best-FaceID and OmniCine — pairs with 10Eros / Sulphur-2; mirrors indexed in block 20
Handheld run	2	0.43 GB	Handheld camera-run motion (step 2000)
lynaNSFW DR34ML4Y AIO	2.3	1.94 GB	AIO (all-in-one) NSFW LoRA — V1 / V2 / V3 builds (NSFW content only)
Motion Enhancer (N54W)	2.3	2.27 GB	Boost motion dynamics
Motion Stabilizer	2.3	—	Stabilize shaky footage
Se0ulSeeker i2v NSFW collection	2.3	27.45 GB	NSFW image-to-video LoRA collection (27 weights; NSFW content only)
Slow-Motion Control LoRA (i2v)	2.5	0.65 GB	Slow-motion / speed control for i2v (auto-gated — sign in to download; repo ships a speed-control workflow and a distilled demo script)
Squish	2	0.43 GB	Squash-and-stretch squish motion
Sulphur Gyrating Hips	2.3	0.67 GB	NSFW hip-thrust motion — trigger phrase `She thrusts her hips in a sexual manner`

◦ Ydudjsb re-uploads the CQdesign motion-transfer weights byte-for-byte (812,454,088 B) together with the workflow JSON - listed as a mirror on the CQdesign row above; see also ❖ Ydudjsb.

❖ Adapter / Pipeline

LoRA	Ver	Size	Description	Download
LTX-2.5-cmf (full pipeline, 4-bit CMF)	2.5	22.07 GB	Full 4-bit CMF pipeline
Wan2.1 VAE Adapter	2	0.45 GB	Wan2.1 VAE bridge

❖ Distill

LoRA	Ver	Size	Description
Distilled LoRA 384-1.1 n4w (full rank)	2.3	7.61 GB	Distillation (full rank)
Distilled LoRA 384-1.1 n4w (rank 128)	2.3	2.58 GB	Distillation (rank 128)
DMD LoRA (r256)	2.3	5.10 GB	DMD few-step distillation — the DMD deltas extracted from JoyAI Echo, reshaped from the LTX 384 1.1 shapes with audio taken from the official distilled LoRA
DMD LoRA Hybrid v1	2.3	5.10 GB	DMD hybrid v1 — flat 0.5 blend across blocks 0–12, then a cosine ramp 0.5→0 over blocks 13–25
DMD LoRA Hybrid v2	2.3	5.10 GB	DMD hybrid v2 — blocks 0–29 replaced with the 384 distilled LoRA at 0.8 video / 1.0 audio
JoyAI-Echo_r256	2.3	4.99 GB	JoyAI-Echo distill extracted for LTX-2.3 (from jdopensource/JoyAI-Echo) — boosts prompt adherence and motion a little; use at 0.4–0.7
JoyAI-Echo-content_r256	2.3	5.10 GB	Content-only variant of the extracted JoyAI-Echo distill
DMD+Distill hybrid `b25-47`	2.3	3.69 GB	DMD for the first blocks + Distill v1.0 @0.6 on the end blocks — smoother movement, cleanest colours (ComfyUI / Kohya layout)
DMD+Distill hybrid `b25-47` (WanGP)	2.3	3.69 GB	Same merge in WanGP layout (prefix-stripped keys)
DMD+Distill hybrid `b35-47`	2.3	2.61 GB	Balanced mix of DMD movement and Distill v1.0 colours (ComfyUI / Kohya layout)
DMD+Distill hybrid `b35-47` (WanGP)	2.3	2.61 GB	Same merge in WanGP layout (prefix-stripped keys)
DMD+Distill hybrid `b35-47` (WanGP v2)	2.3	2.61 GB	Same merge re-exported in diffusers PEFT layout (`diffusion_model.*`, 4,980 tensors)
DMD+Distill hybrid `b35-47` (WanGP v3)	2.3	2.56 GB	Diffusers PEFT re-export with a reduced tensor set (4,896 tensors / 1.28 B params vs 4,980 / 1.30 B)
DMD+Distill hybrid `b45-46`	2.3	1.59 GB	Maximum DMD movement — blocks 45–46 replaced (ComfyUI / Kohya layout)
DMD+Distill hybrid `b45-46` (WanGP)	2.3	1.59 GB	Same merge in WanGP layout (prefix-stripped keys)
DMD+Distill hybrid `b45-47`	2.3	1.65 GB	Maximum DMD movement with a colour-burn fix on the final output blocks (ComfyUI / Kohya layout)
DMD+Distill hybrid `b45-47` (WanGP)	2.3	1.65 GB	Same merge in WanGP layout (prefix-stripped keys)
10Eros v1.4 fro99-AVGRank104 (averaged merge)	2.3	3.48 GB	Averaged multi-source LoRA merge for 10Eros v1.4 — verified by header: 5,172 tensors / 3,448 `lora_A`+`lora_B` pairs, per-layer dynamic rank (~104 avg), bf16
10Eros v1.4 Wan2GP fro99-AVGRank104	2.3	3.48 GB	Wan2GP-oriented build of the same averaged merge (4,302 tensors / 2,868 LoRA pairs — a reduced tensor set)
DMD Distill 72energy	2.3	1.73 GB	DMD few-step distillation (72energy build)

◦ DMD+Distill hybrids — by maximsobolev275. The DMD LoRA gives fast movement but burns colour and contrast; the official Distill v1.0 gives clean colour but sluggish motion. These hybrids take DMD for the movement (early blocks) and replace the end blocks with Distill v1.0 at 0.6 strength. Use 4–8 steps, CFG 1.0, LoRA strength 1.0 (or 0.8–0.9 for the heavier blends); the audio layers are kept from DMD. The block-range variants trade off: b45-46 = maximum movement, b45-47 = maximum movement plus a colour-burn fix on the final blocks, b35-47 = balanced, b25-47 = cleanest colours. The three key layouts are lora_unet_* (ComfyUI / Kohya), prefix-stripped (WanGP), and diffusion_model.* (diffusers PEFT).

▣ ID-LoRA (Identity-Driven In-Context LoRA)

ID-LoRA is a method that enables identity-preserving audio-video generation in a single model. It jointly generates a subject's appearance and voice, letting a text prompt, a reference image, and a short audio clip govern both modalities together. Built on top of LTX-2.3 (22B), it is the first method to personalize visual appearance and voice within a single generative pass.

Unlike cascaded pipelines that treat audio and video separately, ID-LoRA operates in a unified latent space where a single text prompt can simultaneously dictate the scene's visual content, environmental acoustics, and speaking style—while preserving the subject's vocal identity and visual likeness.

Key Features:

Text prompt controls the scene and content
Reference image preserves the subject's visual likeness
Short audio clip preserves the subject's vocal identity
Single unified generation pass for both appearance and voice

Available LoRAs for LTX-2.3:

LoRA	LoRA Rank	Size	Description	Download
ID-LoRA-TalkVid-3K	128	1.16 GB	Identity + voice (TalkVid)	┊
ID-LoRA-CelebVHQ-3K	128	1.16 GB	Identity + voice (CelebVHQ)	┊

Resources:

Project Page | GitHub | Paper (arXiv: 2603.10256)

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ ComfyUI Nodes

▣ Custom Node Collections

10S-Comfy-nodes by TenStrip - Custom ComfyUI nodes for improving motion quality when working with LTX 2.3's combined audio/video latent pipeline. Includes Latent Cross Fade Auto Concat, Audio Latent Stretch, Latent Motion Sharpener, Latent Temporal Upsampler, Latent Motion Retime, and Latent Temporal Inpainter for clean 30fps output from 24fps sampled models.
Deno Custom Nodes by Deno2026 - Practical ComfyUI custom nodes focused on fast real-world workflow improvements including (Deno) Resize Box, Multi Image Loader, LTX Sequencer, LTX Model Loader, Easy Model Download Helper, LTX Multi LoRA Loader, and LTX Prompt Guide.
PromptRelay by kijai - Enables consistent multilingual lip-sync while maintaining voice consistency across languages. Distributes video latent frames across segments with smart prompt node supporting inline and block syntax styles.
WhatDreamsCost ComfyUI by WhatDreamsCost - LTX Director 2.0 plus a variety of custom ComfyUI nodes and workflows for creating AI-generated video content including Multi Image Loader, LTX Sequencer, LTX Keyframer, Speech Length Calculator, Load Video UI, and Load Audio UI. 1.7k+ stars.
ComfyUI-Sapiens2 by kijai - ComfyUI nodes for Sapiens2 computer vision models from Facebook Research. Supports pose estimation, body-part segmentation, surface normal estimation, and pointmap estimation with model variants from 400M to 5B parameters.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ LoRA Training

For training LTX LoRAs, the community uses a variety of official scripts, community-developed forks, and cloud-based platforms.

Primary Local Training Tools

Official LTX-2 Trainer: This is the standard Python-based package for training LoRAs, full fine-tuning, and In-Context (IC) LoRAs. It is designed for Linux and requires CUDA and Triton.
- Link: Official LTX-2 Trainer Repository
Musubi-Tuner (AkaneTendo25 Fork): Widely considered the fastest and most efficient local trainer for LTX-2 and 2.3. It features significantly smaller cache sizes (up to 12x smaller than AI Toolkit) and better iteration speeds, reaching up to 2 iterations per second on an RTX 5090.
- Link: AkaneTendo25 Musubi-Tuner Fork
AI Toolkit (by Ostris): A popular third-party tool that supports LTX-2 character and image-to-video LoRAs. While beginner-friendly, some users reported issues with audio training on the main branch.
- Link: Official AI Toolkit Repository
AI Toolkit: BIG-DADDY-VERSION (ArtDesignAwesome Fork): This specific fork was created to fix broken audio and voice training in the original AI Toolkit. It is optimized for hardware like the RTX 5090.
- Link: ArtDesignAwesome AI Toolkit Fork
rs-nodes (richservo): A collection of nodes that includes a full LTX Lora trainer directly within ComfyUI. It is designed to be memory-efficient, allowing training on cards with as little as 11GB-12GB of VRAM by using ComfyUI's native weight loaders.
- Link: rs-nodes ComfyUI Trainer
SimpleTuner: A highly optimized trainer for Linux that supports LTX-2 and is noted for its ability to handle larger datasets on limited VRAM via block swapping.
- Link: SimpleTuner Repository

Cloud Training Platforms

Fal.ai: Provides a dedicated cloud trainer for custom styles and effects, though it is primarily limited to image-based training datasets.
- Link: Fal.ai LTX2 Video Trainer
RunComfy: A cloud service that offers a pre-configured AI Toolkit setup specifically for LTX-2 training.
- Link: RunComfy LTX-2 Training

Essential Dataset & Captioning Tools

Taz's Ultimate Captioning Tool: A Hugging Face space frequently used by the community to generate the long, detailed, cinematographic prompts (around 200 words) that LTX-2 requires for high-quality training.
- Link: LoRA Caption Assistant (Hugging Face)
AI Video Clipper & LoRA Captioner: A modular pipeline designed to automate local dataset creation using WhisperX and Qwen2-VL, including support for RTX 5090 Blackwell cards.
- Link: AI Video Clipper & LoRA Captioner

Training Requirements Summary

Dataset: Videos should typically be cut to 121 frames (exactly 4.84 seconds) to align with the model's architectural "8n+1" rule.
Hardware: While 16GB VRAM is possible with extreme offloading in tools like rs-nodes, 24GB is the practical minimum for quantized training. For best results and speed, 48GB to 80GB (H100 or RTX 6000) is preferred.
Precision: It is now officially recommended to train on the full BF16 model for LTX 2.3 rather than FP8 for superior quality.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

▓ Workflow & Technical Notes

❖ RuneXX

RuneXX LTX-2.3 / 2.5 Workflows:

Note: per the author, these workflows target LTX-2.3 models but are largely backwards-compatible — most work as-is with LTX-2.5 models swapped in ("Most all should work").

Movie-Maker:

Prompt-Relay-Custom-Audio:

Prompt-Relay-Dev-Model:

I2V T2V Short-Story PromptRelay-Timeline multi-sequence Dev model

Multi-ref-character-sheet:

Helper-wf:

Talking-Avatar-TTS:

Video-2-Video:

Just-Talk — add voice to silent video:

Extend-Any-Video:

Inpaint:

Shot-to-Shot-Transition:

Other:

Music-Video-Creator:

Others:

I2V T2V Basic custom audio with Gemma-API example

Custom-Audio:

First-Last-Frame:

Long-Video-Experimental:

3-Pass-Experimental:

Control-reference:

Helper-Workflows:

Other-examples:

I2V T2V Basic Custom Audio with Gemma-API

RuneXX LTX-2 Workflows old pre_feb2026

* [First Last Frame (guide node)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20First%20Last%20Frame%20(guide%20node).json) * [First Last Frame (in-place node)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20First%20Last%20Frame%20(in-place%20node).json) * [First Middle Last Frame (guide node)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20First%20Middle%20Last%20Frame%20(guide%20node).json) * [I2V Basic (GGUF)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20Basic%20(GGUF).json) * [I2V Basic](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20Basic.json) * [I2V IC-Control (pose)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20IC-Control%20(pose).json) * [I2V Simple First Middle Last Frame (1-pass K-Sampler)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20Simple%20First%20Middle%20Last%20Frame%20(1-pass%20K-Sampler).json) * [I2V Talking Avatar (voice clone Qwen-TTS)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20Talking%20Avatar%20(voice%20clone%20Qwen-TTS).json) * [I2V and T2V (beta test sampler previews)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20and%20T2V%20(beta%20test%20sampler%20previews).json) * [I2V and T2V Basic (Custom Audio)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20and%20T2V%20Basic%20(Custom%20Audio).json) * [I2V and T2V IC-Control (All-In-One Pose Canny Depth)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20and%20T2V%20IC-Control%20(All-In-One%20Pose%20Canny%20Depth).json) * [I2V and T2V Simple (1-pass K-Sampler)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20and%20T2V%20Simple%20(1-pass%20K-Sampler).json) * [I2V and T2V Simple (1-pass)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20I2V%20and%20T2V%20Simple%20(1-pass).json) * [T2V Basic (GGUF)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20T2V%20Basic%20(GGUF).json) * [T2V Basic (low vram)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20T2V%20Basic%20(low%20vram).json) * [T2V Basic](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20T2V%20Basic.json) * [V2A Foley (add sound to any video)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20V2A%20Foley%20(add%20sound%20to%20any%20video).json) * [V2V (extend any video)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20V2V%20(extend%20any%20video).json) * [V2V Head Swap Experimental (BFS lora)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20V2V%20Head%20Swap%20Experimental%20(BFS%20lora).json) * [V2V Just Dub It (experimental)(translate speech auto dubbing)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20V2V%20Just%20Dub%20It%20(experimental)(translate%20speech%20auto%20dubbing).json) * [V2V Just Dub It (with voice clone)(auto dubbing translation)(experimental)](https://huggingface.co/RuneXX/LTX-2-Workflows/resolve/main/older_comfy_pre_feb2026/LTX-2%20-%20V2V%20Just%20Dub%20It%20(with%20voice%20clone)(auto%20dubbing%20translation)(experimental).json)

═══●═══●═══●═══●═══●═══●═══●═══●═══

❖ SOLRICKS

SOLRICKS LTX-2.5 ComfyUI Workflows:

T2V + I2V with audio support, controlled from a simplified main node (switch T2V/I2V, custom audio, duration, resolution, FPS, seed, prompt enhancement). Uses a two-stage generation-and-refinement setup for higher quality. The repo also bundles the two IC-LoRAs used (ltx-2.3-22b-ic-lora-ingredients-0.9 and union-control-ref0.5).

❖ embriotapir

embriotapir LTX-2.5 MSR Dual-Stage Workflow:

Tidied-up ComfyUI workflow for LTX-2.5 Multi-Subject Reference (MSR) video generation — up to 5 reference images blended via learned MSR slot embeddings — with an optional toggleable 2x latent-upscale + low-noise refinement second stage (one-click bypass via rgthree Fast Groups Bypasser). Shared Set/Get nodes keep prompts/FPS/duration in sync across both stages.

Requires: ComfyUI-LTX2.5-MSR, ComfyUI-PromptRelay, ComfyUI-KJNodes, rgthree-comfy, ComfyUI-VideoHelperSuite.

LTX2.5 MSR — CLEAN

❖ Lightricks

LTX-2.3:

LTX-2:

LTX-2.5 Layout-to-Render:

Two Stage Distilled - companion workflow for the Layout-to-Render IC-LoRA

❖ ChrisColeTech

Example workflows bundled with the ChrisColeTech Uncensored Turbo merges — LTX-2.5 v1.1 and LTX-2.3 v1.4. The 2.5 v1.1 set uses the comfyui-gguf-loader and RTX Super Resolution custom nodes, and the repo asks for a nightly comfyui-gguf-loader build.

LTX-2.5 Uncensored Turbo v1.1:

LTX-2.3 Uncensored Turbo v1.4:

❖ vrgamedevgirl84

vrgamedevgirl84 LTX 2.3 Music Video Creator:

Music Video Creator Workflow
- Prompt Creator Workflow - Audio upload, beat detection, scene timing, lyrics analysis, style selection, prompt generation
- Text-to-Video Workflow - LoRA integration, advanced prompt controls, Remake Mode, video stitching
- Image-to-Video Workflow - Uses Z-Image Turbo and LTX 2.3
- Requirements: ComfyUI, LTX 2.3 models, Z-Image Turbo model, FFmpeg, vrgamedevgirl custom nodes

❖ EllaPriest45

EllaPriest45 LTX-2.3 / LTX-2.5 ComfyUI Workflows:

Note: the author's LTX assets are split across sub-repos (see EllaPriest45 (model mirror)). Workflows live in their own LTX2.3_workflows repo, and several LTX2.5_actions / LTX2.5_base weights ship with a matching workflow JSON alongside them.

EllaPriest45/LTX2.3_workflows — 155 ComfyUI workflow JSONs for LTX-2.3 (~0.7 GB total). A representative selection:

Workflows bundled with LTX2.5_actions weights (each sits next to its LoRA in the repo):

Workflows bundled with LTX2.5_base:

❖ TenStrip

TenStrip LTX-2.3 workflow packs — from the author of the 10Eros merge family (see ❖ 10Eros). These workflows expect the 10S-Comfy-nodes custom nodes, which are published on the ComfyUI registry / Manager — update them whenever a new workflow is uploaded. The author describes the pack as quality-first and time-costly: expect 1–5 minute run times.

TenStrip/LTX2.3-10Eros_Workflows:

I2V Basic (DMD V5) — the workflow referenced by the 10Eros v1.3 model card and by the v1.5 fp8 card
I2V FaceID v2 — Face-ID I2V implementation, from Alissonerdx/LTX-Best-Face-ID

Legacy set (Old/):

TenStrip/Workflows — earlier, lighter setups:

10Eros i2v / v2v — SDE-heavy sampling with red nodes marking the adjustment points; often better with the split sampler removed and euler_cfg only
LTX2 simple i2v — LoRA stacking — res2s with lower distilled strength (rank-175 distilled LoRA at 0.75–0.9) and many stacked LoRAs
XLP Coyottes HyperLoRA

Companion workflow shipped with a model — the v1.5 fp8 build carries its own copy of the DMD V5 graph adapted to the fp8mixed weights:

I2V Basic DMD V5 — 1.5 fp8 variant

◦ The DMD V5 graph expects the TenStrip DMD LoRA and the Distilled LoRA v1.1 experiments — for the fp8 transformer build the author points at Kijai's split files and notes the fp8mixed variant is preferable to plain fp8. The older FaceID workflow uses cg-sigmas.

❖ CQdesign

CQdesign LTX-2.5 LoRA workflows:

LTX2.5 — CQ Freeform Motion Transfer LoRA — companion workflow for the freeform motion-transfer LoRA (see ❖ Motion).

❖ Ydudjsb

Ydudjsb LTX-2.5 LoRA workflows:

LTX2.5 - CQ Freeform Motion Transfer LoRA - mirror of the CQdesign workflow, bundled with the mirror weights