Introduction

Comfy challenged the community to build open-source versions of the paid creative apps everyone loves, and now it has shipped its own entry. On October 26 the official ComfyUI account introduced Openjutsu, a side project from a few people on the team. They call it "fully open source," and you can "run it on Comfy, run it locally, or plug in a paid model if you want higher quality."

The code is public at Comfy-Org/comfy-openjutsu under the MIT license. It also has an app page on the Comfy hub.

What It Does

The README describes it as a standalone app for swapping the person in a video for the character in one image:

  • Upload a clip and a character image
  • Choose a 5 to 15 second part of the clip
  • Get that part back with your character in it, while the scene, the performance and the clip's own sound stay

Under the hood it builds on Akatz Labs' H3 Character Swap workflow and MiniMax H3. The app submits a ComfyUI API graph (workflows/swap.api.json). The prompt always starts with the workflow's own instruction:

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

Your optional extra direction is added after it (up to 500 characters in the app's code).

The Optional First-Frame Trick

Before the swap, the app can take the starting frame and your character image and ask GPT Image 2.5 (via Comfy Router, quality low) to place the character, outfit included, into that frame. The swap then starts from that anchor image so the character's look carries through. If the call fails, or you untick it, the swap runs without it. The README notes this first frame is a separate Comfy Router charge.

The Stack

The build guide pins every piece:

  • ComfyUI 98cc0bf84b (v0.41.0), which has the MiniMax H3, SAM 3.1 and video nodes the workflow uses
  • Custom nodes: Akatz Labs' h3_character_swap pack, plus a small GPL-3.0 patch that uses bundled ffmpeg/PyAV and takes numeric resolutions
  • Diffusion model: minimax_h3_ref2va_pruned_int8_convrot from Comfy-Org/MiniMax-H3
  • Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq
  • VAEs: the H3 video VAE (fp16) and audio VAE (fp32)
  • Segmentation: SAM 3.1 multiplex (fp16)
  • LoRAs: h3_character_swap_v1.5 from akatz-ai and the lightx2v H3 FL2V Turbo 8-step LoRA

Sampling follows Akatz's settings: 8 steps, CFG 1, er_sde, 124-frame windows with 18 frames of history, a 64 px mask expansion, and SAM picking the largest person.

Where to Run It Today

This is not a one-click local install, so plan before you clone:

  • You need a Comfy Developer Platform account, an API key, and a deployment built with the nodes and models above. Cloning the repo does not create that deployment.
  • The workflow was run on a 96 GB GPU (RTX PRO 6000), and smaller setups have not been tested.
  • The app itself is a Node.js 22.6+ project. Run npm ci, then npm run dev, then open localhost:3000.
  • Comfy says a 5-second part at 480p took about two minutes on a warm worker. The first job can take several minutes while the deployment loads.
  • Quality options are 480p or 768p. Clips go up to 100 MB (MP4 or MOV) and must be one continuous shot, with a person visible in the first frame of the part you choose.

Some early visitors replied on X that the hosted page still said it was not open yet, so the GitHub repo is the surest way in right now.

Licensing Heads-Up

The app is MIT, but the models have their own licenses. The build guide says MiniMax H3 and the character-swap LoRA fall under the MiniMax H3 Community License. Its standard grant excludes some regions, including the US, EU, UK and South Korea. SAM 3.1 uses the SAM License, and the Turbo LoRA is Apache-2.0. Read each model card before you deploy.

Why It Matters

Openjutsu is a template more than a product. It shows how to wrap a Comfy workflow in your own web app, with uploads, node IDs, a first-frame pre-pass and a deployment build. That's useful for anyone entering the #OpenSourceHiggsfield challenge, where Comfy says submissions are due October 30.

Conclusion

A reference app with MIT code, pinned weights and a documented graph is a great gift to builders. Fork it, change the form in components/app-runner.tsx, and make the character-swap studio you've been wanting. I can't wait to see what you build!

If you'd rather direct H3 without setting up a deployment, Gen runs H3 in your browser.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&amp;utm_campaign=gen&amp;utm_content=comfy-openjutsu-open-source-h3-character-swap-app

Original Source

—Aurelia ♡