any ways to increase speed on single 24gb vram gpu?

jjihwan / FIFO-Diffusion_public

Official implementation of FIFO-Diffusion

https://jjihwan.github.io

277 stars 17 forks source link

any ways to increase speed on single 24gb vram gpu? #4

Closed Gitterman69 closed 3 weeks ago

Gitterman69 commented 1 month ago

is there any way to increase speed (16 minutes for 100 frames on 4090 while utilising 100% gpu resources) without multi gpu usage!? its amazing we finally have a “long video” creation tool with good quality… any hints/tips/tricks would be highly appreciated.

also some ways to decrease quality in order to speed up things would be helpful

thanks so much

jjihwan commented 1 month ago

Thank you for your interest in our research!

In terms of hardware, I don’t have an idea to accelerate inference with a single GPU. You might try to load multiple pipelines on a single GPU using our torch multiprocessing code, but I don’t think it will dramatically reduce the time.

BTW, reducing the “video_length” (i.e., f in the paper) in the code from 16 to 12, 8, or sth will definitely save sampling time. However, it may decrease quality, and we have not tested these cases yet 😅