Original Title: Viggle/Viggle-Remake
Two sets of LoRAs (rank 256) trained by Viggle on the MiniMax-H3 base transformer: swap mode replaces people and characters, supporting multiple replacements at once; edit mode handles all other modifications, including replacing, deleting, or inserting objects, as well as changing backgrounds, weather, or lighting. The inputs are the original video, 1–4 hand-drawn keyframes (which can be pinned to specific frames in the original clip), and prompts written in MiniMax-H3's six-segment format. The output preserves the original video's motion, camera work, pacing, and audio. Officially described as v2, previously teased on the Viggle-Animate card: Animate only swaps characters, while Remake can also modify objects and scenes. The repository includes an MCP service; agents like Claude Code and Codex, given a video and a single sentence, will automatically segment shots, draw keyframes, render window by window, self-check results, and then combine them with the original audio into a complete film.
Why It Matters Not just character swapping: it also supports adding or removing objects, changing scenes and lighting, and comes with an MCP service that allows coding agents to automatically segment shots, draw keyframes, and assemble the entire video.
Rising Score 208 · Likes 374 · Downloads 183188
Reasons
- Official examples include creating 22 versions of a single shot from La La Land, as well as swapping characters in a full music video lasting 116 seconds with 69 edit points; longer videos are split into windows at edit points for separate rendering.
- A recommended path is provided based on GPU availability: use the Viggle app if you have no GPU; for 12–24GB GPUs, manually run workflows using ComfyUI's int8 / GGUF files; for 32GB (RTX 5090) GPUs, you can run the agent plus the ComfyUI rendering endpoint, with official estimates of approximately 28 minutes per window at 1280×736 resolution and 345 frames; for GPUs with more than 80GB, use the reference script sample.py.
- There are also smaller LoRAs with rank 128 / 64 (1.34 / 0.67GB), which officials claim produce nearly identical outputs; examples/demo.sh can be used to verify whether the LoRA is actually loaded.
Usage
- First, download the MiniMaxAI/MiniMax-H3 base model (approximately 144GB in files) and directories such as lora/ and inference/ from this repository (approximately 5.5GB); do not git clone the entire repository (which totals approximately 210GB including all weights).
- Manual mode: python inference/sample.py --mode swap (or edit) --video original_video --prompt prompt_text --ref keyframe --pin frame_number --frames 124 --short-edge 480; the number of frames must be 17n+5, and officials recommend not exceeding 294.
- Agent mode: download only the mcp/ directory, install the server side with pip install -e ., then install the worker on a GPU machine; refer to PROMPTING.md for prompt formatting guidelines.
Risks
- It is possible to insert a real person's likeness into a video without their consent, as the model itself does not perform identity or authorization verification; officials require that it be used only on material you have the right to modify and that AI generation be clearly labeled.
- The weights are derived from MiniMax H3; commercial use requires prominent labeling of “MiniMax H3” in the interface, and annual revenue exceeding 2000 ten thousand USD requires prior written authorization from MiniMax; copyright for film clips used in examples remains with the original owners.
- Single rendering takes approximately 12 seconds; six-segment format prompts are mandatory; known issues listed by officials include drift in multi-shot sequences without pinned frames, long shots of non-humanoid characters, and minor detail artifacts; this entry has not been run.