ComfyUI adds a MiniMax H3 Context Windows node for longer clips with matching sound
A change merged into ComfyUI on 30 October adds context window support for MiniMax H3 and a new MiniMax H3 Context Windows node. The sampler then works on a long clip in overlapping slices of picture and sound instead of all frames at once, which keeps memory tied to the window size.
What it does
The pull request by kijai adds basic context window support for MiniMax H3 and an easier node for it. According to ArtRealmAI's reading of the merged code, windows snap to H3's 17-frame chunks, the audio latent is split at the same frame boundaries as each video window, first-frame, last-frame and keyframe guides are cropped and re-anchored to each window, reference conditioning stays attached, and FreeNoise is on by default to help windows blend. You put the node between the H3 model loader and your usual sampler and set the empty latent to the full length you want.
Worth knowing
It landed after the v0.43.0 release (27 October), so you need ComfyUI master or a nightly build until the next release. ArtRealmAI lists the defaults as context_length 124 and context_overlap 39 frames (both must be 17n + 5, so 22, 39, 56, 73 and so on) with the pyramid fuse method. The PR calls it basic support and ships no example workflow; each window only sees its own frames plus the overlap, so very long clips can still drift, and larger overlaps blend better but take longer.