NikoDemon80/ComfyUI-H3-Motion-Context

What it solves

This project provides a way to chain multiple MiniMax H3 video and audio clips together so that motion and sound continue seamlessly across the cuts. It prevents the common issues of color drift, softening of images, and audio "ticks" or restarts that typically occur when trying to extend a video generation using standard anchoring methods.

How it works

The system uses a specialized node to slice the tail of a previous clip's latent representation (both picture and sound) and feed it into the next generation. By working directly with latents rather than decoding them to pixels and re-encoding them, it maintains bit-for-bit accuracy and avoids quality loss. It also rewrites the time coordinates of the pinned audio so the model treats it as a continuation of the existing track rather than a separate, similar-sounding clip.

Who it’s for

Users of ComfyUI and the MiniMax H3 model who want to create longer, consistent videos with synchronized audio without visible or audible seams at the join points.

Highlights

• Latent-based chaining: Avoids the lossy round-trip of decoding to pixels, preventing color drift and softening.
• Audio continuation: Ensures audio continues as a single track rather than restarting or sounding like a "cover band" imitation.
• H3 Motion Context Save/Load: Dedicated nodes to carry latents across different generation runs to avoid circular connection errors in ComfyUI.
• Seam Probe: A built-in tool to measure if a join is a true continuation or a convincing imitation.
• Safety checks: Automatically verifies ComfyUI's H3 layout arithmetic before rendering to prevent bad outputs after software updates.