<h1 id="fasth3-6-step-lora-for-minimax-h3-comfyui-native">⚡ FastH3 6-Step LoRA for MiniMax H3 — ComfyUI-native</h1>Cut your H3 render time by ~2.5× with zero custom nodes. This is FastVideo's FastH3 4-step distillation (Dense / Data-Free adapter) converted into a plain ComfyUI LoRA — drop it in <code>models/loras</code>, add it to your LoRA stack, set steps to 6. No git clone, no patched loaders, no extra kernels.Works with stock ComfyUI and any Comfy-Org pruned MiniMax H3 checkpoint. Verified on <code>int8_convrot</code> (fl2va + ref2va); bf16 and fp8_scaled use the same layout and should work too.<hr /><h2 id="what-this-is-(and-isn't)">What this is (and isn't)</h2>FastVideo released FastH3 — MiniMax H3 distilled with DMD2 down to 4 sampling steps. Their headline "14× faster" numbers come from VSA sparse attention, which needs their own CUDA kernel and launcher (multi-GPU). That part is not in this file.This is the Dense variant's LoRA adapter, converted from FastVideo's diffusers-style layout to ComfyUI's fused-QKV H3 layout. What carries over: the full attention + MLP distillation (653 tensors). What can't be expressed as a ComfyUI LoRA: the AdaLN timestep modules and the time embedder — which is exactly why you run 6 steps instead of 4. The extra steps stand in for the missing timestep recalibration.Result: the same H3 you already run, finishing in 6 steps instead of 20 (or 8 with turbo lora).<h2 id="settings">Settings</h2>LoRA strength (model)1.0 — do not lower it, that makes things worse. LoRA strength 1.0. Steps 6 (5 = serviceable drafts · 4 = flicker/bloom, unusable · 7–8 = marginal gain). CFG1.0 (distilled — no CFG) Sigma shift your normal values; video 8–12 / audio 3 tested fine Sampler / schedule reuler / beta, or whatever you use for H3<h2 id="where-it-shines">Where it shines</h2>Seed hunting. Scout 4 candidates in the time one used to take, pick the best, then refine with the base model (leave this LoRA out of your refine/upscale pass — the polish deserves the undistilled weights). Tested with my Stubelius Director packs: FastH3 on the hunt, clean H3 on the finish.<h2 id="caveats-read-before-judging-it">Caveats — read before judging it</h2><ul><li>T2VA-trained preview. FastVideo's Preview v1 targets text-to-audio-video. It loads on ref2va / fl2va checkpoints, but reference-mode quality isn't characterized yet — test your reference scenes before trusting it there. Omni Ref is FastVideo's stated next focus.</li><li>Not the VSA speedup. If you want the sparse-attention 14×, look at Kijai's <code>vsa</code> branch + comfy-kitchen <code>sol_attn</code> (advanced, breaks on ComfyUI updates) or the ComfyUI-MiniMax-H3-FastVideo node.</li><li>Fine detail / motion dynamics are where 6-step distills give ground first. Refine with the base model. Reference mode works. Tested on ref2va with multi-character reference scenes: identity held at 6 steps. (FastVideo's preview is T2VA-trained and they list Omni Ref as "next" — but the converted dense adapter already carries over to ref2va in practice.)</li></ul><h2 id="credits-and-license">Credits &amp; license</h2><ul><li>FastVideo / hao-ai-lab — FastH3 distillation (DMD2), Dense/Data-Free adapter</li><li>NikoDemon80 — ComfyUI-FastH3-Lora-Converter (the layout conversion this file was made with)</li><li>MiniMax — H3 base · Comfy-Org — pruned checkpoints</li></ul>Inherits the MiniMax H3 Community License. Redistributed converted adapter; no weights of my own — I just did the conversion so you don't have to.

Minimax FastH3 lora

This is the Dense variant's LoRA adapter, converted from FastVideo's diffusers-style layout to ComfyUI's fused-QKV H3 layout. What carries over: the full attention + MLP distillation (653 tensors). What can't be expressed as a ComfyUI LoRA: the AdaLN timestep modules and the time embedder — which is exactly why you run 6 steps instead of 4. The extra steps stand in for the missing timestep recalibration.

Files with SHA256: 0eef87b6254a4be843fea8c02cb9eb6f8954e8f83f6ade1f9016657e30dec1cf