MiniMax H3 in ComfyUI: Veda setup and speed comparison

This tutorial explains how to add Veda sparse attention to a MiniMax H3 workflow in ComfyUI, then compares generation times and motion quality. The speaker describes a small predictor that selects attention tiles to compute without changing the model's backbone weights. The claimed gains of up to 6.8 times faster attention and roughly three times faster overall generation are separate from the timings in the demonstration.

Setup covers installing the custom node through ComfyUI Manager or the custom nodes folder, placing predictor files in ComfyUI/models/veda/, and connecting the node before the sampler. Text-to-video and image-to-video share a predictor; reference-to-video uses a dedicated file. The demonstrated pipeline combines Veda with Sage Attention and turbo LoRAs. The speaker also explains how to substitute Comfy Kitchen Attention.

For a 15-second example with 362 frames at 0.7 megapixels, the speaker reports a base sampling pass of 1 minute 57 seconds with Veda, compared with 2 minutes 51 seconds using Sage Attention alone. A separate sampler upscales the output. Coin-flipping and hand movements still show problems, and the speaker prefers simpler motion or dialogue scenes for this setup. RTX 30-series support is described as having a code path but remaining unverified by the author.