⚡ FastH3 6-Step LoRA for MiniMax H3 — ComfyUI-native
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 models/loras, 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 int8_convrot (fl2va + ref2va); bf16 and fp8_scaled use the same layout and should work too.
What this is (and isn't)
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).
Settings
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
Where it shines
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.
Caveats — read before judging it
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.
Not the VSA speedup. If you want the sparse-attention 14×, look at Kijai's
vsabranch + comfy-kitchensol_attn(advanced, breaks on ComfyUI updates) or the ComfyUI-MiniMax-H3-FastVideo node.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.)
Credits & license
FastVideo / hao-ai-lab — FastH3 distillation (DMD2), Dense/Data-Free adapter
NikoDemon80 — ComfyUI-FastH3-Lora-Converter (the layout conversion this file was made with)
MiniMax — H3 base · Comfy-Org — pruned checkpoints
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.
Description
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.
FAQ
Comments (5)
This lora give me really bad quality video genertion cmpared to the v1.1 turbo 4 step 768 lora and the 0.1 4 step600 ema.
I tested at 4, 6 and 8 steps. Same bad qualité and audio. and no speed generation gain of course. As it need the Fastvideo node and the Fastvideo speed increase is optimized only for blackwell and new nvidia GPU.
I made 5 tests, and deleted this lora. Need to wait the real comfyui implementation, and a better version.
There's definitely speed improvements going from twenty steps to six steps. Try six steps without the lora and then with the lora. I think the quality of this one is a contender for the best so far. I used the released script by the original author to generate my own rather than downloading this one -- I made mine from int8-pruned but the hashes don't match, but there's definitely some increased quality at lower steps.
@CosmicSpores all my tests was with the two loras i cited in first comment, and this one, and without lora. for 4, 6, and 8 steps. Seen no speed gain, and lower quality video/audio.
See my last two video generated few hours ago, Preview resolution 844x480 at 4' steps, with the minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600_comfy_fro_pruned.
Even at 8 step with the fastH3 lora i can't get this quality.
https://civitai.red/models/2909163/minimax-cumshot?modelVersionId=3290361
@Pat3dx Might be worth trying to create your own copy from the source. Instructions aren't too intense: https://github.com/NikoDemon80/ComfyUI-FastH3-Lora-Converter/ . I would try it out. I'm not getting undercooked results at 6 steps. Hm
@CosmicSpores ty :p.
But as we all know, the real issue is the R2V model.
To get the same video/audio quality than the FL2V model, R2V need like 8 or 10 additional step more.
The FL2V model natively produce better quality than the R2V.
Even at 8 step i can't get the same quality than my examples videos at 480p. When FL2V need only 4 step for decent quality, R2V need like 10-15 step to produce the same quality.
I hope the comapny will retrain their R2V model, there is a issue in this one :p, and all lora change nothing.
This lora,, not helped me with speed and quality, but maybe others users will have more luck.