Skip to main content
Back

Viggle Remake Lets Developers Recast Any Video With Open Weights

Viggle open-sourced a two-LoRA system on MiniMax H3 that recasts actors, swaps props and changes weather in existing footage while preserving motion, camera and audio.

Model, Full Access
  • Viggle released Viggle Remake, two open LoRAs on MiniMax H3 for editing existing video.
  • Swap mode recasts people and characters; edit mode changes objects, weather, light and style.
  • Keeps original motion, camera, timing and soundtrack; works on whole films, not just clips.
  • Ships an MCP server so Claude Code or codex can remake a film from one sentence.
  • Base plus weights are ~210 GB; GGUF builds run on 12 GB cards.
  • MiniMax H3 Community License applies; identity comes from painted frames, no built-in consent check.

Viggle has released Viggle Remake, an open-weight video editing system built on MiniMax H3. It re-renders existing footage with new characters, clothing, props, weather, lighting, or visual styles while preserving the source motion, camera work, timing, and soundtrack.

The model card demonstrates a shot from La La Land remade 22 ways and a 116-second Uptown Funk video recast across 69 cuts. Both examples use one-sentence instructions sent to a coding agent.

Two adapters divide the work

Remake ships as two rank-256 LoRA adapters for MiniMax H3’s base transformer. A LoRA is a compact set of trained weight updates that modifies a base model without requiring another full checkpoint. One adapter handles people and characters; the second covers props, backgrounds, weather, lighting, and broader scene changes.

Mode Command flag Intended edits
Swap --mode swap Humans, animals, and stylized characters
Edit --mode edit Objects, backgrounds, weather, lighting, and style

The release also includes a ComfyUI node pack, 8-bit and GGUF builds for lower-memory systems, and a Model Context Protocol server that lets compatible coding agents process a film shot by shot.

Painted keyframes steer the shot

Users guide an edit by exporting one to four frames from the source video and painting the desired changes into them. Each reference can optionally be pinned to its original frame index, which helps the model maintain timing and identity across longer or multi-shot sequences.

The accompanying text prompt follows MiniMax H3’s six-section template. Remake then propagates the painted changes through the shot, removing the need to prepare pose skeletons, segmentation masks, or separate identity embeddings.

Local inference starts at 144 GB

MiniMax H3’s base files occupy about 144 GB, while the complete Remake repository reaches roughly 210 GB when every variant is included. The reference Python sampler uses bfloat16 precision and requires at least one GPU with about 80 GB of memory. Each render window spans approximately 12 seconds of footage, with longer videos divided into windows at detected cuts.