<h2 id="minimax-h3:-15-s-image-to-video-at-about-1-mp-with-audio-(low-memory)">MiniMax H3: 15 s image-to-video at about 1 MP, with audio (low-memory)</h2><p>An image-to-video workflow for MiniMax H3 that renders about 15 seconds (357 frames) at roughly 1 megapixel, with audio. It uses the LayerStream: H3 Chunking node to keep activation memory in check on long clips.</p><h3 id="why-the-chunking-node">Why the chunking node</h3><p>On long or large H3 clips the model's activations, not the weights, can fill the VRAM. They grow with the number of tokens (frames × resolution). The node runs the same transformer math in slices instead of one big pass. In my tests, PyTorch's own allocations stayed under a hard 8 GB cap for this 15 s, 1 MP clip.</p><h3 id="does-it-change-the-picture">Does it change the picture?</h3><p>No. Output was bit-identical to the stock run in my tests. I compared the chunked and stock results frame by frame at 6 steps, on a 124-frame clip and on a full 15 s clip, and they matched exactly. The node doesn't approximate anything.</p><p>Two honest caveats:</p><ul><li><p>Audio was compared byte for byte only on a short clip (56 frames). I haven't run that comparison on the full 15 s yet.</p></li><li><p>Chunking is slower. About 10 % in my test.</p></li></ul><h3 id="what's-in-the-workflow">What's in the workflow</h3><ul><li><p>Start image → H3 image-to-video, 6 steps with the 8-step turbo LoRA, <code>res_multistep</code> sampler</p></li><li><p>Megapixel Size node: sets width and height from your start image (about 1.05 MP, multiple of 32, so a square image gives 1024×1024)</p></li><li><p>Audio VAE decode, so the video has sound</p></li><li><p>Saved as a video with audio</p></li></ul><p>A tip I learned the hard way: use 6 steps, not 2. At 2 steps this turbo LoRA gives artifacts (for example a distorted nose in motion). That happens with or without my node.</p><h3 id="requirements">Requirements</h3><ul><li><p>ComfyUI 0.39.0 or newer</p></li><li><p>The osiworx-pack custom node pack. Install it from ComfyUI-Manager or from <a target="_blank" rel="ugc" href="https://github.com/osi1880vr/osiworx-pack">https://github.com/osi1880vr/osiworx-pack</a></p></li><li><p>MiniMax H3 files: diffusion model (hybrid int8), text encoder, video VAE, audio VAE (fp32), and the 8-step turbo LoRA. The file names in the loader nodes are from my machine, so pick your own.</p></li></ul><h3 id="tested-on">Tested on</h3><p>Windows, RTX 4090 (24 GB), ComfyUI 0.39.0. I haven't tested cards below 24 GB or other operating systems. On my 24 GB card, a launch flag (<code>--vram-headroom</code>) already solved the memory problem on its own. The node mainly lowers the memory floor and keeps usage predictable. </p><h3 id="limits">Limits</h3><ul><li><p>The node patches parts of ComfyUI's H3 code, so a future ComfyUI update could break it. Tell me if it does.</p></li><li><p>The lowest activation memory I measured for H3 is about 4.2 GiB. Below that it needs more work.</p></li><li><p>Model files are not included.</p></li></ul><p>Details, measurements and what I haven't tested are in the README: <a target="_blank" rel="ugc" href="https://github.com/osi1880vr/osiworx-pack">https://github.com/osi1880vr/osiworx-pack</a></p>

Files with SHA256: 21a6396bcff368b9efa4d2ba801d9984f6083c5f6696f1baf74876b75e4ffcf1