Changing a video without disturbing its performance is harder than generating a new clip: edits must stay aligned with the original motion, camera, cuts, and timing. This project treats a painted frame as the visual anchor for each shot, then renders the surrounding footage to match—making it suited to substantial recasts and scene changes rather than just short, isolated effects.
Key Capabilities
- Separate character-replacement and general scene-editing modes let users recast people or change objects, settings, weather, and lighting without treating every edit as the same task.
- An MCP server can coordinate shot detection, keyframe painting, rendering, quality checks, and soundtrack assembly across a longer film. That makes the workflow agent-driven, but generated details can differ between runs.
- The source soundtrack is carried through rather than generated, so the original audio remains synchronized with the edited footage.
- LoRA weights, ComfyUI workflows, and quantized formats offer different ways to run the model, though the full base model is large and high-resolution local inference requires substantial GPU memory.
Who It’s For
Great fit if you have footage you’re authorized to modify and want to preserve its filmed performance while changing its cast, props, or setting. Look elsewhere if you need lightweight local inference, reliable fine text or glass rendering, or generated audio. The model cannot verify consent or identity, so likeness rights and clear AI-generation labeling remain the user’s responsibility.
