"Recency bias" / model "degradation"

[SAC020]

I have put the subject in quotes because I am using tentative terms to describe what I am seeing.

Trained a model, everything worked great, generated beautiful photorealistic images with great styling flexibility.

After a few hours of playing with different prompts it started to generate sketchy, unstyle-able pieces of crap, especially when building batches. The first photo in a batch can be reasonably good and styled, but the subsequent ones are very poor.

No matter how I add "photo", "photorealistic" etc in the prompt, or, the opposite, leave the prompt as basic as possible, it is generating consistently poor results.

I am stumped, and this is not the first model I am seeing this happening. I realize the CKPT is not changed (looking at the file timestamp), and indeed using a fixed seed with a fixed prompt does generate identical images.

So, is there some sort of "caching" or "training" or "bias" going on in SD? Is the random seed generation not really random? I cannot explain otherwise what I am seeing.