Text-to-Image
Diffusers
English
SVDQuant
FLUX.1-dev
FLUX.1
Diffusion
Quantization
ICLR2025

Are there any quality or speed differences between the int4 and fp4 models?

#2
by nymical - opened

Which one would you recommend?

nymical changed discussion title from Are there any quality of speed differences between the int4 and fp4 models? to Are there any quality or speed differences between the int4 and fp4 models?

Since I unfortunately didn't get any replies here. I've just found out that fp4 is for 50-series GPUs, while int4 is for others.
For reference, I tried fp4 before and it didn't work on my 3060. So, I downloaded the int4 and is working well on my PC.
Still have no idea about quality or speed though, but I guess it doesn't matter now, at least for non-50-series owners.

nymical changed discussion status to closed
MIT HAN Lab org

fp4 should have better quality, but it is only for 50-series.

MIT HAN Lab org

the speed is almost the same.

Thank you for confirming!
Keep up the amazing work you guys are doing. :)

Sign up or log in to comment