H3-LongTakeNoCuts: ComfyUI nodes for a three-minute MiniMax H3 shot without cuts
H3-LongTakeNoCuts is a free set of ComfyUI nodes and a workflow that fight the quality loss when you extend a MiniMax H3 shot with motion context. Its 18 October update adds character masking to the refine pass, and the author shows a three-minute one-take made with it.
What it does
When motion context or latent save/load extends a video without cuts, degradation shows up around segments 5 to 6: each segment treats the previous latent as ground truth, so drift and detail loss add up, like a photocopy of a photocopy. The fix adds a short refine resample at 0.5 to 0.6 denoise that restores character detail and the waxy or burnt look, with a character mask made by SAM3, then blends the result back into the latent.
Two nodes do the work: Keyframe Mask (default 15:1, 22:0, 30:1) keeps the character matched to the previous shot at the start, and Blend Latent By Frames (default 0:0, 22:0, 44:1) fades from the current latent to the refined one between frames 22 and 44, because motion context trims the first 22 frames.
Results and limits
The author says it brings the degradation down to a minimum and works very well with moving shots, but in half and static shots the subject still drifts toward higher contrast and saturation. It's a workaround rather than a fresh-latent regeneration, and how motion context handles audio in this setup is still an open question.
Worth knowing
It's built on ComfyUI's native motion-context example workflow with minimal changes and needs NikoDemon80's ComfyUI-H3-Motion-Context nodes; forks such as ComfyUI-H3-Motion-Context-MultiRef load the wrong node and throw an error. KJNodes is also required. The code is MIT; MiniMax H3 itself is under its Community License: Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use.
Try it yourself
In ComfyUI/custom_nodes run git clone https://github.com/xyzDist/H3-LongTakeNoCuts.git, load its workflow, chain motion-context segments of [a slow tracking shot following one character through a market], and keep Blend Latent By Frames at 0:0, 22:0, 44:1.