tritant/ComfyUI_MiniMax_H3_Extender

What it solves

This project provides a custom node for ComfyUI that enables the generation of long, continuous video sequences using the MiniMax H3 model. It solves the problem of maintaining visual, motion, and audio continuity across multiple shorter clips, allowing users to build extended videos without jarring transitions or loss of identity.

How it works

The extender uses a clip-by-clip generation workflow with a disk-based latent cache and a "Motion Context" system to ensure seamless progression. It supports Ref2VA conditioning, allowing users to provide image, video, and audio references to maintain character and scene consistency. The system includes a validation mechanism where users can lock in accepted clips, ensuring that subsequent clips are built upon a stable visual foundation without needing to regenerate previous work.

Who it’s for

AI video creators and ComfyUI users who want to produce longer, more structured narratives with consistent characters and environments across multiple scenes.

Highlights

• Continuous Motion Context: Maintains seamless motion and visual flow between generated clips.
• Multi-Reference Support: Supports up to 9 image references, 3 video references, and 3 audio references for precise control over identity and timing.
• Latent Refine/Upscale: An optional second pass that upscales latents and refines the results for higher quality.
• Existing Video Continuation: Ability to start a generation from an existing video file, using its final frames as context.
• Per-Clip Customization: Individual control over prompts, seeds, LoRAs, and color corrections for each clip in the sequence.
• Project Persistence: Save and load .ext project files that embed references and settings to resume work across sessions.
• Dynamic Image Guides: Support for FL2VA clips with multiple image guides placed at specific frames to control transformations.