Introduction

Another #OpenSourceHiggsfield entry just landed, and it's a big one. On October 27 LoRA trainer Alisson Anjos (alissonerdx on X and Hugging Face) introduced Henge, an open-source video studio built on ComfyUI + MiniMax H3. It's named after Henge no Jutsu, the transformation technique from Naruto, and the pitch is simple: keep the performance, transform the character.

The app code is public at alisson-anjos/Henge under the MIT license. You can run it on your own ComfyUI, in one RunPod pod, or with the UI on your laptop and the GPU on RunPod Serverless.

What It Does

You give Henge one guide video and a few pictures, then pick a mode:

  • Character Swap replaces the person in the video with a character picture
  • Motion Transfer drives a brand-new character with the guide's performance
  • Head Swap changes only the head, using a face picture
  • Multi-character swap marks several people and swaps them at once, even across camera cuts
  • Sharpness cleans up an upload or one of your results, with optional 2K/4K upscaling

The README shows a 30-second fight clip with both fighters swapped, rendered as three shots at 480p. The four headline demos ran on RunPod Serverless at 480p, with Sharpness taking a swap result up to 1080p.

Not Just a Prompt Wrapper

Each mode uses its own trained LoRA, not plain H3 with a clever prompt. They come from Anjos's BFS LoRA collection:

  • Character Swap v1.1
  • Motion Transfer v1.0 and TARA v1.0
  • Head Swap v1.0 (r32)
  • Little More Sharpness v1.0, plus an H3 latent upscaler

Two technical ideas stand out:

  • Guide-latent training. Borrowed from LTX's guide-latent IC-LoRAs and brought to H3, the guide video's VAE latents share the output's frame grid, so frame N is built on frame N of the guide. Pose, camera and timing stay locked.
  • TARA-H3 (Temporally Aligned Reference Attention). This is Anjos's new method for Motion Transfer. It keeps the guide as a native video reference and nudges each output moment to look at the matching guide moment, without adding trained layers. It's marked experimental, and a technical report sits in his ai-toolkit fork.

The Studio Around It

This is where Henge feels like a real product:

  • Shot-by-shot long clips. Long videos split at camera cuts or at a fixed length, with context frames carried between shots. You can preview, redo or download each shot on its own.
  • Subject guidance. Mark people with SAM 3.1 tracking and bind each one to its pictures, as 2D color marks, a 3D body, or a latent mask that leaves the rest of the frame pixel-identical.
  • A built-in assistant (Claude, OpenAI or a local model) that can inspect footage, mark characters, write prompts and launch jobs. You can also drive Henge from your own AI client through its MCP server.
  • Turbo options include DMD 8-step, DMAD 4-step, pDMD 4/2-step and lightx2v 8/4-step, or none.
  • Audio can keep the original track or the soundtrack H3 generates per shot.
  • Every job downloads as the exact ComfyUI API graph it ran, which is lovely for anyone who wants to learn the workflow.

Where to Run It Today

Pick one of three setups:

  • A. Local with your own ComfyUI. The README asks for an NVIDIA GPU with 32 GB VRAM and 64 GB RAM (think RTX 5090), plus CUDA 13 PyTorch, Node.js 22.6+, Python 3.10+ and ffmpeg.
  • B. All in one RunPod pod. You pay while the pod is on. The first boot downloads the models.
  • C. Local UI + RunPod Serverless. No local GPU needed. You pay per second while a job runs, with a 1–3 minute cold start after idle.

Budget disk space: every profile together is about 125 GB of models (H3 base, text encoder, VAE, planner VLM and the task LoRAs). The custom nodes are published in the Comfy Registry, and Comfy Cloud support is still being validated.

The local install starts like this:

git clone https://github.com/alisson-anjos/Henge.git && cd Henge
npm ci && npm run build
ln -s "$PWD/backend/comfy_nodes/henge" /path/to/ComfyUI/custom_nodes/henge
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git /path/to/ComfyUI/custom_nodes/bfsnodes

There's also an AI-guided install doc you can hand to a coding agent, which checks your machine and walks through the option you choose.

Licensing Heads-Up

The app code is MIT, but the weights are not relicensed. The Motion Transfer LoRAs on Hugging Face use the BFS Motion Transfer license, MiniMax H3 has its own community license, and the README points to third-party notices for everything else. Read each model card before you ship client work. Henge also says it's independent of Higgsfield and Comfy Org.

Why It Matters

On October 26 Comfy shipped Openjutsu as a reference app for single character swaps. Henge goes further, with dedicated LoRAs per task, a motion-transfer method, multi-person tracking, long-clip shot planning and an MCP hook. It's a strong sign that open H3 tooling is growing into full studios, not just one-off workflows.

Conclusion

If you've got a 32 GB card or a RunPod account, Henge is one of the most complete open performance-transfer stacks you can try today. Start with Character Swap on a short clip, then try TARA Motion Transfer and see how well your character keeps the beat.

Want to play with H3 without setting up 125 GB of models first? Gen runs it in your browser.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=henge-open-source-h3-character-swap-motion-transfer-studio

Original Source

—Aurelia ♡