How To Speed Up MiniMax H3 in ComfyUI With Veda Sparse Attention (Patreon)

Video : https://youtu.be/r39Ej4L2dO4

Hey everyone.

New video is up. This one is Veda, a sparse attention method for MiniMax H3 that made my render times drop by around a minute on a 15-second clip, without touching the model itself.

The idea is simple once you see it: about 90 percent of the attention H3 computes every step barely contributes to the final video. Veda distills a tiny predictor from the full model that scores the attention tiles, keeps the 10 percent that matters, and skips the rest. In my test, a 15-second R2V clip went from 2 minutes 51 seconds down to 1 minute 57 seconds, and the node report showed 28 percent of attention computed, 72 percent skipped.

The part I like most: it is not a LoRA. Your weights and style stay exactly the same, and it stacks with what you already run — Sage Attention, Turbo LoRAs, enhancement LoRAs, all fine together. Install is a couple of minutes: ComfyUI Manager, search Veda, install, then drop the predictor safetensors into a new veda folder under models. There is one predictor for text-to-video and image-to-video, and a separate one for reference-to-video.

Honest verdict from my testing: for drama scenes, people talking, camera motion and general footage, the quality holds up and the speed gain is real. For complex fast motion — my test was fingers flipping a coin — you get morphing, and maybe 5 or 6 out of 10 clips are usable. So pick it per scene type. If you want the absolute detail on high action, raise your sampling steps and ease off the Turbo LoRAs instead.

The free 2-stage R2V setup from the video is attached below, grab it and run it. Supporters, the full version with the upscale pass and the Singularity 3-sampling setup is in the supporter post.

— Benji

Resources:

- Veda-on-ComfyUI custom node: https://github.com/veda-sparse/Veda-on-ComfyUI

- Veda T2VA predictor (text/image-to-video): https://huggingface.co/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview

- Veda R2VA predictor (reference-to-video): https://huggingface.co/Veda-Sparse/Minimax-H3-R2VA-Veda-Preview

- Predictor files: minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors / minimax_h3_r2va_veda_preview_fp8.safetensors — both go into ComfyUI/models/veda/

- MiniMax H3 ComfyUI weights (Comfy-Org): https://huggingface.co/Comfy-Org/MiniMax-H3

- Veda paper: https://arxiv.org/abs/2605.30325