MiniMax H3 pruned fp8 scaled
Pruned FP8 version of the video model
fl2va:
(I2V, FL2VA, T2V)
ref2va:
-Other required files:
nvfp4 Text Encoder: https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors?download=true
Description
MiniMax H3 ref2va pruned fp8 scaled
Note: Use the same VAEs and Text Encoder as the "fl2va pruned fp8 scaled" version
FAQ
Comments (17)
What the difference between int8 and fp8,the latter more fast?
In simplified terms: FP8 is more precise, it uses floating-point. INT8 uses integral.
int8 can be slightly faster depending on hardware.
fp8 in 40 series and up , for others int8 then w4a8 then GGUF if both vram and ram is low
576*928 5秒视频在5060Ti 16G / 32G内存上跑了5:27
I generated one 3s clip with this, and instantly yeeted like 200 GB worth of LTX and Wan models. Holy diff O_O
Yeah. I can't even believe it myself... Like... Uncensored, works, very well, righ tout of the gate. I'm actually finding it unreal to be honest.
@DaddyWolfgang lol, same, and also the fact that you can do so much with basic ahh prompts @.@ Not having to feed it a novel just to control the camera feels like the years of my lifespan stolen by LTX are coming back
Guess I gotta eat crow. I never expected an uncensored model from that company. EVER. Glad I didn't put money on it being censored or not because I'd have lost.
LOL
I just emptied 200gb of wan and ltx and dumped 10 + custom workflows. Were free!!!!
On a 4070 12GB + 48GB - Generated a 5 seconds 544p clip in just 105 secs with the experimental Turbo LORA, upscaled to 720p (ish) with RTX Super Res. Not as fast as LTX, but the prompt adherance is WAYYYYY better.
@I_XXIV Hi! Please share your workflow.
If have 64RAM and 3080ti 12gb VRAM? Then what type of h3 is best and balanced in terms of the generation time and output quality?
This one will work well for you; if you want, you can try the Int8 Convrot and then decide which one suits you best.
Tried it with my 5070ti + 32gb ram, its much slower than the INT8
Since the ComfyUI 0.32 update, FP8 scaled is basically useless because of the INT8 ConvRot versions.