Henge: an open studio for H3 character swap, motion transfer and head swap, plus Higgsfield open-sources its 'Inside AI' short

Henge in one line: one guide video plus a few reference pictures in, a transformed clip out. Per the README it has four modes. Character Swap and Head Swap replace a person or just a head. Motion Transfer drives a new character with the guide's performance. Sharpness cleans up and can upscale to 2K/4K. It's built on ComfyUI and MiniMax H3, and author Alisson Anjos (@Alissonerdx) posted it on Oct 27 at 6:38 AM MDT as an #OpenSourceHiggsfield entry. The post passed 100 likes and 110 bookmarks by afternoon.

Why it's more than a UI. Each mode uses a LoRA the author trained, not plain H3 with a prompt. The README explains a guide-latent method, borrowed from LTX's IC-LoRAs, where frame N of the output is built on frame N of the guide, so pose, camera and timing stay locked. It also describes TARA, his new temporally aligned reference attention for motion transfer, with a technical report. Long clips split at camera cuts and carry context frames between shots. SAM 3.1 tracking lets you mark several people and bind each to their own pictures, and a mask mode leaves the rest of the frame untouched. A built-in assistant (Claude, OpenAI or a local model) can mark subjects and launch jobs, or your own agent can drive it over MCP. Every job downloads as the exact ComfyUI graph it ran.

What it takes. Local use needs a working ComfyUI, and the README lists an NVIDIA GPU with 32 GB VRAM and 64 GB RAM, plus about 125 GB of models for every profile. The lighter paths are a RunPod pod, or the local UI with a RunPod Serverless endpoint whose workers start per job and stop 15 seconds after the last request. The app is MIT; the H3 weights and the Motion Transfer LoRAs carry their own licenses, so read those before client work. Use it only on footage and likenesses you have the rights and consent for.

Also this morning: Higgsfield released 'Inside AI', a short commercial-style film (per its project page, a shy student's nervous typed message turns into a dreamy digital leap), and published the whole project, every prompt and asset, on its community page. The announcement passed 500 likes and 200 reposts within hours. If you study how studios structure shot prompts, it's a free reference. Note that the project page needs JavaScript to render.

Our take: Henge and OPENjutsu are converging on the same job, and Henge's trained per-mode LoRAs and multi-character tracking make it the more production-minded of the two on paper. The hardware bar is high, so the Serverless route is the realistic way to try it. The #OpenSourceHiggsfield challenge closes Oct 30.