openai / consistencydecoder

Consistency Distilled Diff VAE
MIT License
2.1k stars 76 forks source link

Significant performance problems, see profiler screenshot #3

Open felix-red-panda opened 7 months ago

felix-red-panda commented 7 months ago

image_2023-11-06_23-07-38

Running the consistency decoder takes several seconds and most of this time is spent in a stalled state and reducing the number of diffusion steps leads to no meaningful speed increase. The default SD1.5 decoder is ~100x faster running the code example in the readme.

I'm on Pytorch 2.0.1 on Linux kernel 6.1 with an RTX 3060

sokolegg commented 7 months ago

Same problem

SKDDJ commented 7 months ago

:eyes:

sokolegg commented 7 months ago

Personally upgrade from old diffusers (12) to new one was helpful