random seed leakage

[altsoph]

The current "seed" and "noise_prompt_seeds" arguments don't guarantee a fixed output.

Here are two results with the exact same arguments, including

pixray.add_settings(seed=42, noise_prompt_seeds=[42,])

They are pretty similar, which means the initial random image is the same.
However, there is some source of unseeded randomness, I think it will be great to locate and fix it :)

[dribnet]

Any chance you get better results if you leave off the noise_prompt_seeds? Just curious if that might be the source of it.

[altsoph]

Any chance you get better results if you leave off the noise_prompt_seeds? Just curious if that might be the source of it.

I tried it with the same results

[robinsloan]

I suspect that CLIP is the culprit!

See: openai/CLIP#13

I suspect it's due to CUDA's nondeterministic behavior, as it doesn't happen in the CPU mode. Slight numerical imprecision is sometimes inevitable, especially when dealing with half-precision models. A workaround is to "warm up" the model as you did, and after a third call or so it'll more likely emit deterministic outputs.