Comfy Openjutsu

A standalone app for swapping the person in a video for the character in one image, with a Comfy workflow. Upload a clip and a character image, choose a 5 to 15 second part, and get the part back with the character in it. The scene, the performance and the clip's own sound stay.

It is the app behind Openjutsu on comfy.org, built on Akatz Labs' H3 Character Swap workflow and MiniMax H3.

You need a Comfy Developer Platform account, an API key, and a deployment with the nodes and models listed in build/model-files.md. The workflow was run on a 96 GB GPU (RTX PRO 6000); smaller configurations have not been tested. Generation uses deployment credits, and the optional first frame is a separate Comfy Router charge.

1. Install Node.js 22.6 or newer. Prepare a Comfy Developer Platform deployment with the patched h3_character_swap node pack, MiniMax H3, SAM 3.1 and the two LoRAs by following build/model-files.md. Cloning this repo does not create that deployment.

2. Copy this repository and make local settings:

   git clone https://github.com/Comfy-Org/comfy-openjutsu.git
   cd comfy-openjutsu
   cp .env.example .env.local

3. Add the API key for the workspace that owns your deployment to COMFY_API_KEY. Set COMFY_BASE_URL to the deployment URL from the Developer Platform; the SDK reads both variables from .env.local on the server.

4. Install and start the app:

   npm ci
   npm run dev

5. Open http://localhost:3000. Pick one of the examples, or choose a clip of one continuous shot with a person in its first frame, and a character image. Pick where the part starts and how long it is, then select Swap character.

The first job may take several minutes while the deployment starts and loads its models; a 5 second part at 480p took about two minutes on a warm worker. The server reads the key from .env.local; do not expose it in a NEXT_PUBLIC_ variable or commit the file.

Swap character submits workflows/swap.api.json. lib/swap.ts fills it in: the clip goes to node 30, the character image to node 9, the part's start and length to the trim node and node 31, the quality to node 31 (480 or 768), and the prompt and seed to node 801. The server uploads the files with the SDK and submits the graph; node 809 saves the result, which the page plays from /api/jobs/<id>/outputs/<output>.

Every prompt starts with the workflow's own instruction, using H3's reference tags, and then adds the optional extra direction:

Replace the selected person in <Video 1> with the character in <Picture 1>. Preserve the source performance, scene and audio timing.

Before the swap, the page takes the frame where the part starts and posts it with the character image to /api/anchor. The server asks GPT Image 2.5 (openai/gpt-image-2.5-sunburst, quality low) through Comfy Router to put the character, outfit included, into that frame, using the same COMFY_API_KEY. The swap then starts from that image (the anchor_image and anchor_fit nodes), which carries the character's look from the first frame. If it fails, or you untick it, the swap runs without it and those two nodes are removed. See lib/anchor.ts.

The graph is workflows/H3 Character Swap.api.json from akatz-labs/h3-character-swap-tool at 0670e4f, with these changes:

Every other setting is Akatz's: 8 steps, CFG 1, er_sde, 124-frame windows with 18 frames of history, a 64 px mask expansion, and SAM finding the largest person.

The form and its defaults are in components/app-runner.tsx, styled in app/globals.css. The four examples are the ones on comfy.org, read from media.comfy.org through /api/examples; change the list in lib/examples.ts. lib/swap.ts holds the part limits, the qualities and the prompt, and lib/anchor.ts the first-frame prompt.

To edit the workflow, open it in ComfyUI and export Workflow (API). Replace workflows/swap.api.json and update the node ids in lib/swap.ts if they moved. If the change adds a model or custom node, add it to the Comfy build and publish a new release before running the app.

The deployment Build must include every node and model used by the workflow. The versions, files, and download commands are in build/model-files.md.

Install the Comfy CLI and sign in to the Comfy account for your deployment. Run these commands from the prepared ComfyUI folder containing models/ and custom_nodes/:

pip install comfy-cli
comfy setup --where cloud
cd /path/to/ComfyUI
comfy build init --name comfy-openjutsu --models-dir ./models --custom-nodes-dir ./custom_nodes
comfy build status
comfy build push --release --target linux/nvidia
comfy deploy refs compute

Create a deployment from the released Build, choose an available GPU and region, and put its URL in .env.local. Stop or delete the deployment when you finish to end GPU charges.

These commands package the local ComfyUI install; they do not download model files. After deployment, run one 5 second swap at 480p and confirm a video comes back.

• Clips up to 100 MB, MP4 or MOV; the part is 5 to 15 seconds.
• The clip should be one continuous shot. A hard cut breaks the swap at the cut.
• SAM looks for a person in the part's first frame. A part that opens on a frame with no person (a fade from black, a title card) fails before it generates.
• With several people, the largest one is swapped.
• A silent clip gives a silent result.
• Reloading the page clears the current job and result from the UI.

MIT. See LICENSE. The workflow is derived from Akatz Labs' MIT-licensed workflow, and the node-pack patch in build/ is GPL-3.0 like the pack it changes; both notices are in THIRD_PARTY_NOTICES. Model licenses are separate; see the end of build/model-files.md.