How is the speed and quality compared to int8?
Wow, Thanks, ANybody tried the in4 quant? How is the speed and quality compared to int8
want to know
speed 50% faster, quality 300% worse, or i'm doing something wrong, just selected this int4 pruned model instead of the int8 pruned from kijai
speed 50% faster, quality 300% worse, or i'm doing something wrong, just selected this int4 pruned model instead of the int8 pruned from kijai
the int4 text encoder works fine for me
It seems like... it is not good.
Workflow: https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.jsonYes, This looks really bad. Wel Kijai is working on an int4 model hopefully we can see some speed improvements with good quality
Yes, This looks really bad. Wel Kijai is working on an int4 model hopefully we can see some speed improvements with good quality
oh really, that sounds good. where's this information from?
i just checked the mixed int4 int8, and it was a horror show, so far it seems like int4 destroys H3's quality...
i just checked the mixed int4 int8, and it was a horror show, so far it seems like int4 destroys H3's quality...
https://huggingface.co/tsolful/Minimax_H3_INT4Mixed/tree/main, Checkout this quant. This seems to work. It consumes way less vram than the pruned int8
i just checked the mixed int4 int8, and it was a horror show, so far it seems like int4 destroys H3's quality...
https://huggingface.co/tsolful/Minimax_H3_INT4Mixed/tree/main, Checkout this quant. This seems to work. It consumes way less vram than the pruned int8
which one of the two do you recommend?
I havent tried it. The author says. The int4BQ saves almost 5gb vram and int4Q saves 2gb compared to int8. You can try the int4QB if works, then you can checkout the int4Q