wanx-troopers

MiniMax H3

2026.08.31

PDD appears to be a way to speed up H3 - consists of a LoRA + some extra layers

beta packs more steps into low sigmas where detail sets, which is why it beats euler/simple on flow-matched H3, while karras’s radically different shape performs poorly [1][2]. Turbos kill motion by cutting the high-noise phase where motion gets established

Apparently there are many attempts to accelerate H3:

• Lightx2v
• smth called “Acc Loras”/this ?/; base model 20 steps, accelerated 8 steps
• GH:siraxe/H3_slider_experiments:H3_SPEED_slider

brendacaro: get enough steps, or use some huen_2s like I do for some passes to really get the sound dailed in

A technique of motion transfer is being tested based on pre-populating the working latent.

Interesting:

• HF:Kijai/MiniMax-H3-experimental:minimax_h3_ref_lora_rank_256_bf16
• GH:seitanism/ComfyUI-H3-Motion-Context-MultiRef V2V, references, extensions
• HF:LBH-123-AI/Minimax_h3_latent_Upscaler

RuneX

on my 3090 int8 comfy kitchen is slightly faster. But Sage + mem eff sage patch is about as fast.. Benefit of comfy kitchen is that you dont need to install anything

2026

“A lot of animation styles”. Support merged to Comfy official.

• Model flavors: fl2va and ref2va (multiple references, at least 3, possibly more) “both models do both tasks actually, just more finetuned for their specialties”
• Workflow flavors: i2v, r2v, t2v
• Pruning is removing unnecessary weights, seems like 13B out of 33B have been pruned out, “from what I tested pruned is completely fine, just bit different outputs”
• We have a sort of an edit mode similar to Bernini
• Kijai-style colored bodies understdood natively - via a reference video
• there are reference workflows shipped with Comfy
• video VAE is 5Gb “innately tiled, it uses very little memory … tiled VAE … does nothing on this model”

Weights

• HF:Comfy-Org/MiniMax-H3 “Int8 supposedly works great on Blackwell too”
• prompting guide
• HF:MiniMaxAI/MiniMax-H3 “there’s a docs folder in here with prompting guides btw”
• HF:Kijai/MiniMax-H3-experimental int8 VAE, w4a8 pruned versions of fl2va and ref2va
• HF:lilcheaty/MiniMax-H3-NVFP4 MiniMax_H3_ … _pruned_nvfp4.safetensors files exist.. somewhere - there are complains of noticeable quality degradation
• h3-license, including a link to a form that companies and individuals in UK/US/EU/Soth Korea can use to still obtain a license

text encoder … qwen 32b - even with 4bit its 20Gb - available in nvfp4 as qwen3vl_32b_minimax_h3_nvfp4_awq

nown issues at this point:

• audio breaks with stochastic samplers (SDE/Ancestral)
• audio probably breaks with inpainting and when using denoise under 1.0

int8 error is so low that it shouldn’t be huge difference at all … twice as fast and only ~0.9% relative quant error [compared to fp16]

it has it’s own streams for image/video/audio, you can mix and match them.. up to 9 images, 3 videos, 3 audio

it can be slower if you use a large image, the image size affects it

The old man from <Picture 1> is sitting in a studio in front of a blank canvas, he paints the <Picture 2>.

Drozbay:

Works fine with RES4LYF samplers, though you need to update RES4LYF to most recent version (from today) to get around a small error with nested tensors. Also you really do have to set SDE noise to 0.0 right now (eta in ClownSamplers), it totally breaks the audio otherwise.

Acceleration

HF:drbaph/MiniMax-H3-Turbo-Lora-ComfyUI:minimax_h3_turbo_v4_step600_ema_pruned_comfyui

GH:xmarre/ComfyUI-Spectrum-MiniMax-H3 accelerator node “Is this better than Easy Cache or can you use them together”

patientx: “int8 model + nvfp4 clip + sage-attention + spectrum”

Larryvrh Turbo Lora v 4.0

This site is open source. Improve this page.