My project is hosted at:
Sometimes the results are not so good. Using the same images and prompts, the results are good with a local model.
To troubleshoot this issue, I’ve downloaded the docker and run it locally. However, the locally run docker produces good results, just like the non-docker local model.
So I’m confused.
The only differences I can think of:
- The cuda version. On my local workstation, the cuda is 12.2. The spaces seems to use 12.1. But I guess it shouldn’t make much difference.
- Different random number initialization. Although the seed is fixed, the random numbers may be generated differently, leading to different results.
- Hardware. I actually also tried to run on my own A100 and the results are good. So I wonder, does HF make any customization to the A100? For example sacrifice numerical precisions for speed?