Openjutsu swaps characters and objects in video locally with ComfyUI and MiniMax H3
OpenjutsuFree, open source (MIT); runs on your own GPU with ComfyUI
Openjutsu is a free, MIT-licensed, open-source alternative to Higgsfield Genjutsu: drop in a video and a character photo and the main person is replaced while the camera, cuts, motion and soundtrack stay. It runs on open weights on your own GPU (MiniMax H3 for video, SAM 3.1 for tracking, Qwen3-VL 8B for descriptions) through a small Next.js app over a ComfyUI workflow.
What it does
Simple mode replaces the main person; Advanced mode handles 1 to 4 characters, each matched to a tracked person per shot. Object swap changes, removes or re-backgrounds what SAM tracks (hair, a jacket, a car), and From picture uses FLUX to pose the character from a frame, with an optional relight pass.
Worth knowing
Video generation uses MiniMax H3, whose licence excludes the EU, UK, South Korea and the US. You can choose a smaller describer model (Qwen3-VL 4B FP8 or Gemma 4 E2B INT8) if VRAM is tight.
Open-H3-Regenerate-2K-IC, released on 9 November 2026 by the NVlabs SoL-Engine team, is an in-context LoRA for MiniMax H3 that regenerates a low-resolution draft at 1080p, 2K or 4K in 3, 4 or 8 steps. It is an open alternative to MiniMax's H3-Regenerate-2K, which is only a hosted API, and it keeps the draft in context so faces, lip sync, text and layout stay as they were.
Tencent WeChat Vision's WeVisOne-VL, an Apache-2.0 vision model in 2B and 4B sizes, runs natively in ComfyUI since 9 November 2026, when kijai's pull request was merged. A single Detect node takes an image and a phrase and returns boxes, per-instance alpha mattes, masks, points, counts or OCR boxes, without SAM or custom nodes.
ZONOS3-Mini, released by Zyphra on 9 November 2026, is a compact open text-to-speech model with voice cloning and voice design, under the Apache 2.0 licence. It has 800 million parameters with about 100 million active per token, and runs in real time in a browser through WebGPU, on laptops and on phones.
Fizgig is a free, Apache-2.0 LoRA and fine-tune studio for image, video and music models such as Krea 2, MOSS-Image, Qwen Image 2.1, FLUX Klein, MiniMax H3 and ACE-Step. Version 9.2, released on 8 November 2026, moves to PyTorch 2.15 and a faster Muon optimizer step, so a MOSS-Image LoRA on an RTX 5090 trains in about 20 minutes instead of 49 with likeness unchanged.
SnapRef is a reference adapter for the Anima anime image model: you give it one or more pictures of a character and the prompt only describes the scene, while hair, eyes and outfit come from the reference. The stable v2a release of 8 November 2026 adds a weight for each reference, two modes and six ComfyUI workflows.