continue-revolution / sd-webui-animatediff

AnimateDiff for AUTOMATIC1111 Stable Diffusion WebUI
Other
3.04k stars 253 forks source link

When text2gif and controlnet are used at the same time, the speed becomes very slow. Is it something I did wrong? #334

Open baby11233 opened 10 months ago

baby11233 commented 10 months ago

Use text2img, start animatediff at the same time, frame rate 10, total number of frames 80, number of contexts in a single batch: 16, closed loop: NO, upload video, and open a single controlnet at the same time. Generate image size: 512*910, use graphics card RTX 3080, completion time It lasted 1 hour and 20 minutes. I also tried comfyui and used controlnet to create animated diff. Basically, the same parameters only took more than ten minutes. However, since I am not familiar with comfyui... I have to continue to use webui, but at this speed The gap is really too big

baby11233 commented 10 months ago

屏幕截图 2023-11-21 223715 屏幕截图 2023-11-21 223725

baby11233 commented 10 months ago

屏幕截图 2023-11-21 224434 300多秒/步...

baby11233 commented 10 months ago

上下文单批数量设置为8的时候| 20/20 [05:55<00:00, 17.77s/it] 不到6分钟完成, 但是结果就很难入眼...

ENTPRESTIGIOUS commented 8 months ago

same

marcsyp commented 7 months ago

+1 -- What I am seeing is that generation is slow, yes, but the postprocessing is PAINSTAKINGLY slow. For instance, default parameters with no controlnets takes 63s, add one controlnet for even a tiny portion of the generation (starting at 0, ending at 0.1) takes 3m30s. It's making using controlnets nearly impossible.

Note: I am on torch 1.13.1+cu117 due to a current project that requires it.