FID score changes a lot during the model training

[viyjy]

Hi, I found that FID score becomes larger and larger during the Stage 2 (512x512) training of SD-2.0-base, but the loss keeps at the same level, roughly about 0.12~0.13. Any thoughts? Thanks.

[viyjy]

Update: The FID and CLIP score after the Stage 1 (256x256) training of SD-2.0-base is 10.79 and 25.43, respectively.
However, the FID/CLIP becomes very weird during the Stage 2 (512x512) training of SD-2.0-base as shown below:

Step | FID   | CLIP
-------------------
400k | 25.22 | 24.19
500k | 13.63 | 25.11
600k | 22.31 | 24.21
720k | 21.94 | 24.5
770k | 23.89 | 24.6
-------------------

[viyjy]

This is related to this issue #88