ItsSNOWMOBILE/aftertake

Then comes the Aftertake

▶ Watch the 93-second film · Download the MP4 · Open its editable projects

A film about Aftertake, made in Aftertake. It cuts through generated portals, castles, recast characters and finishing looks to its own song, which was made with open-weight ACE-Step 1.5, with each section under its sung line. All 72 shots, 84 titles and 26 transitions are native, editable Film layers, cut to beats measured by aftertake_detect_beats. The film also shows how a take is built: the real ComfyUI graph, the @comfyorg/sdk calls, a recorded run on an RTX 5080 and a live agent session over MCP. Its generated plates were then upscaled ×2 on Comfy Cloud with Aftertake's own Upscale studio, and the scenes re-rendered unchanged.

What you can make

	One-tap effects One photo in, a vertical clip to share out. Pick one of ten effects: Animate a frame moves the camera on LTX-2.5, an automatic check re-renders a blank, frozen or short take, and Film adds titles, a transition, an ACE-Step song and a small watermark you can turn off. Every effect is one JSON file.
Storyboard One idea becomes a film. A local language model plans 3 to 5 shots you can edit, each renders with LTX-2.5, ACE-Step scores them, and the cut lands on the song's beats with titles and transitions, ready to keep editing in Film.
Create video Text to video with sound, native LTX-2.5.	Create image Stills and first frames with Z-Image Turbo.	Transform a region Paint a timed cue; the effect composites onto your footage.	Replace a character Track a person with SAM3, recast them with MiniMax H3.
Move the camera Animate a frame with 27 camera moves, from 360° orbit to FPV drone.	Transfer motion Your character, their performance, with Wan Animate 2.	Color & finish Looks, curves, LUTs, grain and Hide details.	Cut the film Titles, transitions, narration and music on the beat.

Every card is a real result made in Aftertake. Start from 18 editable recipe starters (product shots, TV spots, UGC, unboxings, try-ons and more, adapted from Higgsfield's public skills), or from a blank prompt.

Highlights

Recast everyone in the shot. Swap one person, then the next: each pass changes only the pixels inside its own tracked mask. Two-person example	Keep what's in front. Depth runs on your NPU or Neural Engine and puts back anything nearer the camera than the person, like this mug. Café example
Same frame, same crop: the source enlarged, and SeedVR2 ×2.	Enhance any clip. Upscale ×2 with SeedVR2, smooth motion to 48 fps or slow motion ×4 with RIFE, or compose a song with ACE-Step 1.5 from style tags and lyrics. All five operations were verified on an RTX 5080. Results and timings Repair the mask. When tracking drops a person for a moment, Repair mask fills the gap on the CPU in seconds and shows exactly what it added. Know the time before you render. Every render button shows how long the job usually takes on your GPU, from your own past runs.

Also in the studio: 2–4 takes of one image, video or shot, compared side by side with synced playback · a moves gallery of 27 camera moves, each with a real preview loop · punch titles and transitions (slam, pop, glitch, whip, zoom, wipe) · beat detection that snaps cuts to music · tracked subject copies · layers and masks with 27 blend modes · depth and light for stills · local narration · selective HSL, curves and 3D LUTs · Hide details, which blurs every place a username appears in a screen recording.

Quick start

You need Node.js 22+ and Git. The launcher checks your computer, installs what's missing and opens the studio in your browser.

git clone https://github.com/ItsSNOWMOBILE/aftertake.git
cd aftertake
node scripts/aftertake.mjs --examples

Or double-click start.cmd (Windows), start.command (macOS) or start.sh (Linux). --examples opens a workspace full of recorded results; leave it off to start with your own. node scripts/aftertake.mjs doctor prints the checklist.

No GPU is needed for the examples, editing, Film, Color & finish or agents. To generate new footage, add a model profile on Windows or Linux with an NVIDIA GPU (tested on an RTX 5080 with 16 GB):

node scripts/aftertake.mjs setup-gpu --video     # or --image, --character-swap, --upscale, --song, …
node scripts/aftertake.mjs setup-npu             # depth on the Neural Engine or an NPU, optional

No NVIDIA GPU, or missing a model? Comfy Cloud is optional, studio by studio: add a key from a paid Comfy Cloud plan in Render setup, then choose Comfy Cloud for any of images, text to video, upscaling, smooth and slow motion, or songs. Everything else keeps your local models.

Show it to someone? node scripts/aftertake.mjs share runs a separate copy behind an access link, with a daily render allowance for each visitor. Share a live link puts it online with a free Tailscale Funnel.

All model profiles

Profile	Flag	Models
Text to video with sound	--video	LTX-2.5
Images	--image	Z-Image Turbo
Reference-frame editing	--frame-edit	Qwen-Image-Edit
Painted region effects	--region-edit	LTX-2.5 with the in/outpainting adapter
Camera guidance	--camera	LTX motion-track IC-LoRA
Motion transfer	--motion-transfer	Wan Animate 2
Character replacement	--character-swap	SAM3 + MiniMax H3 (needs FFmpeg)
Still-image depth	--depth	Depth Anything 3 Small
Upscale ×2	--upscale	SeedVR2 3B
Smooth and slow motion	--interpolate	RIFE 4.26
Songs	--song	ACE-Step 1.5

Add --models-dir=PATH to keep models on another disk. Every download is pinned by revision, size and SHA-256. Render setup in the studio shows which tools are ready and the command that fixes each one. See hardware and GPU selection and the studio guide.

How it's built

flowchart LR
  UI["Studio UI<br/>React · Vite"] --> Server["Local studio server<br/>Node · Express"]
  Agent["Any MCP client<br/>(your agent)"] --> MCP["Aftertake MCP<br/>55 tools"] --> Server
  Server --> SDK["@comfyorg/sdk<br/>isolated worker"] --> Comfy["ComfyUI 0.41<br/>+ Aftertake nodes"] --> GPU[("NVIDIA GPU<br/>open-weight models")]
  Server --> Media["Media runtime<br/>PyAV · Pillow · FFmpeg"]
  Server --> Depth["ONNX Runtime depth<br/>Neural Engine · NPU · GPU · CPU"]

• Native ComfyUI workflows. Every generation is an API-format ComfyUI graph submitted through the official Comfy SDK to your local engine. Download the graph from any result to continue in ComfyUI.
• Open weights only. No partner nodes or closed APIs. Inference runs on ComfyUI's own model nodes; a small pack of Aftertake nodes loads, masks, composites and saves media around them.
• Every result is a recipe. Jobs keep their request, seed, content hashes and the graph that actually ran. Film projects keep every cut, title and sound. Crashes and reconnects rejoin the same job instead of starting a new one.
• Agents are first-class. The MCP server exposes the same studio: generate, track and replace, enhance, plan an edit, voice a script, export, and review decoded frames. Your own agent supplies the language model.
• The editing runs on the CPU. Film export, finishing, layers, mask repair and narration need no GPU; depth for foreground protection uses the best accelerator you have.

Models and licenses

Task	Model	License
Video, text to video, frame animation, region effects	LTX-2.5 and LTX-2 adapters	LTX model license
Images	Z-Image Turbo	Apache-2.0
Reference-frame edits	Qwen-Image-Edit	Apache-2.0
Person tracking · character replacement	SAM3 · MiniMax H3 + Character Swap LoRA	see model cards
Motion transfer	Wan Animate 2	Apache-2.0
Upscale ×2	SeedVR2 3B	Apache-2.0
Frame interpolation	RIFE 4.26	MIT
Songs	ACE-Step 1.5	Apache-2.0
Still depth · foreground depth	Depth Anything 3 Small · V2 Small	Apache-2.0
Narration	Kokoro	Apache-2.0

Model weights keep their own terms, separate from Aftertake's MIT code. Each profile's exact files are pinned in docs/models-*.json; see credits and licenses.

Made with Aftertake

Take the lead · 94 s One shot becomes a campaign: an H3 replacement, a tracked subject in a design, animated headlines and a portrait cut. How it was made	Hotel Lobby, recast · 26 s A supplied portrait becomes the performer on the left, from tracked selection to matched before and after. Workflows and project
Make it yours · 40 s A vertical ad whose hook changes in the real editor, then tracked copies and character replacement. Editable projects	One performance, extra copies A dancer tracked once and repeated without regenerating face or motion. Replays on the CPU. Example

The example library holds dozens of recorded projects with their inputs, outputs, recipes, executed graphs and attribution. Explore them without installing anything in the browser editor.

Documentation

Guide	What's inside
Studio guide	Every studio, its controls, setup profiles and troubleshooting
Text to video	Native video and audio sampling, formats and render evidence
Motion graphics · scenes	Titles, shapes, transitions, groups and depth
Color & finish	Looks, curves, LUTs, optical effects and Hide details
Narrated cuts	Scripts, measured voices, captions and pacing
MCP reference	Agent setup, all 55 tools and recovery behavior
Hardware	GPU selection, memory policy and precision
Share a live link	Let someone use your studio through an access link and a tunnel
Architecture	How the build works and what is reproducible
Media attribution	Footage, fonts, music and generated references

Development

pnpm install
pnpm dev        # studio server with reload
pnpm test       # 400+ TypeScript tests
pnpm lint
pnpm build

Python media tests run with python -m unittest discover -s tests -p "test_*.py" in the media runtime (.runtime/film-venv) or the engine runtime (.runtime/venv). Contributions are welcome; see CONTRIBUTING.md.

Status

Aftertake is an independent, experimental project built for Comfy's Open Source Higgsfield challenge. Generative results still need human review: likeness, hands, occlusions and camera adherence vary by shot, and a passing technical check is not a judgment of visual quality. Optional Comfy Cloud rendering was tested end to end on a Standard plan on October 22, 2026: one image, song, smooth-motion clip, upscale and text-to-video clip, each finished by the studio's usual checks (record). Aftertake does not claim full Higgsfield feature parity.

License and acknowledgments

Aftertake's application code and custom nodes are MIT licensed. Model weights, third-party runtimes, fonts and footage keep their own terms.

Built with ComfyUI, the Comfy SDK, H3 Relay, LTX, MiniMax H3, SAM3, Wan Animate, Qwen, Z-Image, SeedVR2, RIFE, ACE-Step, Depth Anything and Kokoro. Some finishing ideas were inspired by Scumble (attribution). See the full credits and media attribution.