H3 Long Video Pixaroma
60 seconds of MiniMax H3 with no seam in it
- model
- model
H3 stops at fifteen seconds. You don't.
MiniMax H3 is the 33B omni-modal model that finally gave open weights an answer to Veo's native audio: video and synced stereo sound generated as one context, 4 to 15 seconds at up to 2K, straight to the top of the video-editing arena. What it won't do is 45 seconds. It was trained on short clips and it thinks in short clips.
Both obvious workarounds stink. Chaining 10-second clips and pinning the last frame into the next was measured by the pack's author and rejected, because every guided clip comes out darker and softer than the last - a flash and a quality drop at each join. And one long run technically works, but attention grows with the square of the length: at 60 seconds on a 12 GB card it doesn't finish a single step.
H3 Long Video Pixaroma is the third option. You set the length you want, it makes the whole thing in one sampling run, and there is no join to see.
How it works
It's a DIFFUSION_MODEL wrapper - it sits on the model wire and changes what the transformer sees on every step, without touching the sampler.
ComfyUI still samples one latent for the whole clip. On each step the transformer runs on overlapping time windows of that latent, around 10 seconds each, together with the matching slice of the soundtrack, and the window outputs are blended across the overlap. The reason there's no seam: every window is worked on at every step, so the windows converge on one another as denoising progresses. Nothing is pasted together at the end.
Meanwhile model_base still does the noise math, the denoise masks and the audio carry on the whole latent. That's why H3 Audio Sync keeps working unchanged - the track it pinned isn't sliced, only the transformer's view of it is. Per window it also rebuilds the packed layout and cuts pinned guides and a reference song down to that window's slice.
Measured on a 4090: one minute at 864 took 4.5 minutes instead of 9.4, and it ran with 12 GB of VRAM free where a single long run couldn't complete a step. Quality cost is small - 30 seconds in 10-second windows scored 80/162 sharpness/face against 83/166 for one run.
Inputs and outputs
model- your H3 model, after its LoRAs. The output goes on to H3 Audio Sync, or straight to the sampler.window_seconds- 4 to 15, default 10. How much video the model looks at at once. Shorter means less memory and slightly less coherence. It snaps to H3's frame grid, so 10 becomes 10.12 and don't be alarmed when the console shows decimals.overlap_seconds- 0.5 to 7, default 1.5, snapped to 1.62. The shared, blended strip between neighbouring windows. More overlap is smoother and slower.
One output, model - the length still comes from your usual latent and sampler setup. And below about 15 seconds you don't need this node; nothing changes, because a clip no longer than one window is made in one piece as usual.
Install
Part of ComfyUI-Pixaroma, the UX pack from the Pixaroma YouTube channel (122K subscribers, a free ComfyUI course, per-node help pages). MIT. The pack pulls imageio-ffmpeg, which is how the video nodes get ffmpeg without you installing one.
Windows Easy-Install builds already contain it. Otherwise ComfyUI Manager, search Pixaroma - be aware it may hand you 1.4.144, since the Registry flagged newer releases with no stated reason. For the current build:
cd ComfyUI/custom_nodes
git clone https://github.com/pixaroma/ComfyUI-Pixaroma.git
Restart ComfyUI, then Ctrl+Shift+R in the browser. Skipping that hard refresh is the number one cause of "the node is empty" after an install. Search h3 or long video on the canvas to find it.
Where people get burned
The licence isn't simple. H3's weights are open but geofenced - the Community License excludes the US, EU, UK and South Korea. That's about whether you may run the weights at all, not a bug in the node.
It can't rescue a soft upscale. A windowed refine of an upscaled low-res clip doesn't sharpen it; H3 keeps a soft input soft. This node generates, it doesn't upscale.
Still out of memory. Push window_seconds down toward 6–8 before you touch resolution - that's the knob that buys memory. Your console prints something like [Pixaroma] H3 Long Video: 30.24s clip made in 3 windows of 10.12s (1.62s overlap). If that line is missing, the wrapper isn't attached and you're silently doing a single long run.
Wrapper nodes are update-sensitive. This hooks ComfyUI's model patching internals, and core moves them; an update can break the call with a Python TypeError about positional arguments - the exact class of failure that killed the Krea2T Enhancer. The browser error box is useless here, so read the terminal, update the pack and restart.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | Your MiniMax H3 model (after its LoRAs). Wire the output on to H3 Audio Sync and the sampler. | |
| window_seconds | FLOAT | 10.04–15 | How much of the video the model looks at at once. 10 seconds is the tested default; H3 was trained on 4 to 15 seconds. Shorter windows use less memory but hold together a little less well. Snapped to H3's frame grid (10 becomes 10.12). A video no longer than one window is made in one piece as usual. |
| overlap_seconds | FLOAT | 1.50.5–7 | How much two neighbouring windows share; they are blended across it. 1.5 seconds is the tested default. More overlap = smoother but slower. Snapped to H3's frame grid (1.5 becomes 1.62). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | Your model, set to make long videos in windows. Wire it to H3 Audio Sync (or straight to the sampler). |