MiniMax H3 Roundup: VEDA Skips 90% of Attention Steps for 7.1x Speedup

optimisticalish · reddit · 2026-10-06

A MiniMax H3 ecosystem roundup covering recent tools and resources:

VEDA for ComfyUI (preview 0.2): Claimed to compute H3 attention steps, decide which matter, and safely skip about 90% of them, boosting speed "7.1x on attention alone" in an H3 workflow, with gains growing with clip length. Trained at 1.0Mpx and 10 seconds; requires ComfyUI 0.38.0 and a 275MB predictor file placed in ../ComfyUI/models/veda/. Workflows included. The author tested the default workflow on a 3060 12GB: 8 seconds at 0.5Mpx with a 6-step turbo LoRA and no Spectrum took 8 minutes total, acceptable visuals but poor audio — no benefit versus their usual turbo/Kitchen/Spectrum workflow, possibly more noticeable on a 5090.
SSS Rank Minimax H3 LoRA: aims for "smooth, slow, sync" animation.
Claymation LoRA: a simpler look than H3's default claymation styles, but only handles visual style, not stop-motion timing/moves.
Seamless loop Skill: for H3 and ComfyUI, written from a real job and its mistakes, turning a still image into a looping ambient video; works in Claude Code and ships three Comfy .PY scripts.
ComfyUI Contest: build open creative apps with features usually stuck behind a paywall; $10,000 cash for the winner plus RTX 5090 cards.

More from Multimodal