HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models

Official implementation of HRDiT. Our project page is at zylwithxy.github.io/HRDiT-page.

SPA: the bundle and slide operations (T = 7, N = 3).

🔔 News

• [2026-08-06] 🎉 Initial release of the HRDiT code.
• [2026-08-16] ⚙️ SPA and HAP are available as ComfyUI nodes in wildminder/ComfyUI-DyPE: SPA, HAP and HAP Calibrate, thanks for their work!
• [2026-08-27] 🤗 The Remyx AI team has reimplemented HRDiT as a diffusers pipeline at remyxai/hrdit-flux-modular, thanks for their work!
• [2026-08-28] 📺 The five-minute ECCV presentation is on YouTube.
• [2026-08-31] 🌐 The project page is live, with 4K samples and FLUX.1-dev comparisons.

4096×4096, generated by HRDiT from off-the-shelf FLUX.1-dev. More on the project page.

Overview

Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models despite their strong text-to-image generation capabilities at limited resolutions. In this work, we find two key challenges particularly hindering the application of off-the-shelf DiT models for high-resolution image synthesis in a training-free manner, namely spatial disorder and long generation time. To address these challenges, we propose a novel method tailored to adapt off-the-shelf DiT models for high-resolution image synthesis, consisting of two components:

• SPA (Spatial Position Alignment) addresses spatial disorder by replacing each token index with a bundle index before it enters the positional encoding, and averaging over a set of bundle mappings whose boundaries slide, so that positional distinctions are restored without any training.
• HAP (Head-adaptive Attention Pruning) addresses long generation time by giving every attention head its own attention scope, derived once from the off-the-shelf model, and pruning the attention computations outside it.

HRDiT is training-free: it runs on off-the-shelf FLUX.1-dev weights and needs no fine-tuning.

🔧 Installation

Tested with CUDA 12.6 on NVIDIA GPUs. HAP is built on PyTorch FlexAttention and needs PyTorch 2.7 or newer.

conda create -n hrdit python=3.10 -y
conda activate hrdit

pip install torch==2.7.0 torchvision --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

Download the off-the-shelf base model:

huggingface-cli download black-forest-labs/FLUX.1-dev --local-dir ./pretrained_models/FLUX.1-dev

🚀 Inference

Generate a 4K (4096x4096) image from a text prompt:

python inference.py \
    --prompt "A textbook shows a large group of people." \
    --model_path ./pretrained_models/FLUX.1-dev \
    --output_dir ./outputs

The image is produced progressively at 1024 -> 2048 -> 4096. The final 4K image is written to --output_dir, and the 2048 intermediate to <output_dir>_res2k/.

To generate 2K only:

python inference.py --prompt "..." --resolutions 2048

To run several prompts from a file (one prompt per line, or a JSON list of strings):

python inference.py --prompt_file prompts.txt --num_prompts 10

Higher image quality

HAP only shortens generation; SPA is what restores spatial structure. Setting scope_plan=None in the pipe(...) call of inference.py turns HAP off at every step while SPA still runs: attention stays dense, generation takes longer, and image quality is better.

Hardware

Generation runs on a single GPU; the pipeline does not shard the model across devices. A 4096x4096 image peaks at about 46 GiB, of which 33 GiB is the FLUX.1-dev weights in fp16. On a 48 GB A6000 card, run with

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True python inference.py ...

so the allocator does not fragment. --resolutions 2048 peaks at about 36 GiB.

Key arguments

argument	default	meaning
--resolutions	2048 4096	target resolutions, generated in order
--scope_plan	configs/scope_plan_flux.json	per-head attention scopes used by HAP
--spa_steps	3 0	number of leading denoising steps using SPA, per resolution stage
--group_num	80	SPA bundle granularity
--global_anchor	32	HAP keeps every N-th token block globally visible
--seed	3407	random seed
--dtype	fp16	fp16 or bf16

Deriving a scope plan

configs/scope_plan_flux.json holds the per-head attention scopes HAP prunes with, and is ready to use as shipped.

To derive a plan for a different attention budget, see HRDiT-HAP, which contains the profiling and plan-search pipeline. Its output drops straight into --scope_plan.

Citation

@inproceedings{xue2026hrdit,
  title={{HRDiT}: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models},
  author={Xue, Yu and Qu, Haoxuan and Li, Zhuoling and Xu, Hongbin and Yin, Jianxiong and See, Simon and Rahmani, Hossein and Liu, Jun},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

Acknowledgements

The code is built upon HiFlow, FLUX and I-Max, thanks for their work!