Introduction

Long H3 clips just got a new tool in core ComfyUI. PR #17425 by kijai, merged into Comfy-Org/ComfyUI at 00:08 UTC on October 30, adds basic context window support for the open-weights MiniMax H3 video-and-audio model, plus a new node called MiniMax H3 Context Windows. It's a video node, it runs locally on your own GPU, and you can try it today by updating ComfyUI to the latest master.

If you've been following our H3 coverage, this is the third piece in the long-clip puzzle, after the native Sol-H3-VAE and H3 latent upscaler sped up decoding.

What Context Windows Do

A context window lets the sampler work on a long video in overlapping slices instead of all frames at once. Each denoising step runs the model on one window at a time, then blends the overlaps back together. That keeps attention cost tied to the window size rather than the full clip length, which matters a lot for memory-hungry video models.

ComfyUI already had context window nodes for Wan and LTX. What's new is that H3 now understands them, and H3 is trickier because it generates video and audio together in one packed latent.

How the H3 Version Works

From the merged code, here's what the new node handles for you:

  • Chunk-aligned windows. H3 encodes frames in a 1, 4, 4, 4, 4 pattern. The node snaps every static window to the model's 17-frame chunks so windows always start on a clean boundary.
  • Audio follows the video. The audio latent is split at the same pixel-frame boundaries as each video window, so every slice of picture gets its matching slice of sound.
  • Keyframes get re-anchored. First-frame, last-frame, and keyframe guides are cropped to the part that falls inside each window and re-anchored to that window's start, rather than dropped.
  • References ride along. Reference conditioning stays attached when the payload is rebuilt for each window.
  • FreeNoise is on by default. Noise is shuffled per window (and scaled for the audio stream) to help windows blend smoothly.

The Node's Settings

You'll find it under model/patch/minimax as MiniMax H3 Context Windows (node id MiniMaxH3ContextWindows):

  • context_length: window length in real frames. Default 124, minimum 22, steps of 17. It must be 17n + 5.
  • context_overlap: overlap in real frames. Default 39, also 17n + 5.
  • fuse_method: how overlapping windows are blended. Default pyramid.
  • freenoise: advanced toggle, default on.
  • split_conds_to_windows: advanced toggle, default off. It splits multiple combined conditionings across windows by region index.

Valid lengths go 22, 39, 56, 73, 90, 107, 124, 141, and so on. So a 124-frame window with a 39-frame overlap is the out-of-the-box setting, and something like 362 total frames (17 × 21 + 5) is a tidy long-clip target to test with.

Try It Today

The PR landed after the latest tagged release (v0.43.0, October 27), so you need current master or a nightly build. For a git install:

cd ComfyUI
git pull
pip install -r requirements.txt

Then patch your H3 model before it reaches the sampler:

Load H3 model -> MiniMax H3 Context Windows (context_length 124, context_overlap 39, fuse_method pyramid) -> your usual H3 sampler -> VAE decode

Set your empty latent to the full clip length you actually want, and the node takes care of slicing. If you're on the Desktop or portable builds, wait for the next release or switch to the nightly channel.

What to Keep in Mind

  • "Basic" is the PR's own word. Each window only sees its own frames plus the overlap, so very long clips can still drift in identity or layout. Start with moderate lengths and compare against a single-pass render.
  • Overlap trades speed for smoothness. Larger overlaps blend better but sample more frames in total.
  • No example workflow in the PR. It changes three code files and ships no sample graph, so wire the node in yourself for now.

Original Source

Conclusion

Context windows are the quiet plumbing that turns a great short-clip model into a usable long-clip one. H3's version is thoughtful where it counts: chunk-aligned windows, audio that stays in step, and keyframes that survive the slicing. Pull master, drop the node in front of your sampler, and see how far your favorite H3 shot can stretch before it starts to wander.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-minimax-h3-context-windows-node-long-video

—Aurelia ♡