search

nerdyrodent / VQGAN-CLIP

Just playing with getting VQGAN+CLIP running locally, rather than having to use colab.
Other
2.6k stars 428 forks source link

Why do I get different ouputs when using the same input on a different repo? #29

Closed shaolinseed closed 2 years ago

shaolinseed commented 3 years ago

I got an amazing output that i got on the google collab notebook and am trying to replicate it to a larger scale running locally on my 3090. For some reason the outputs appear to have a different style than that of collab (im using the same model, prompt, seed and save interval..)

Is there something thats been altered ? Is it to do with the optimizer or learning rate? (these can't be specified on the collab notebook).

Thanks alot for this bit of software it has given me hours of experimenting and fun!

nerdyrodent commented 3 years ago

The defaults are the same as the notebook, so the outputs should be fairly similar, though randomness will still be a factor. There is a "-d" option with this one to make things a bit more deterministic, but there is still randomness.

shaolinseed commented 3 years ago

The defaults are the same as the notebook, so the outputs should be fairly similar, though randomness will still be a factor. There is a "-d" option with this one to make things a bit more deterministic, but there is still randomness.

hmm, theres definately somehting different going on. I've tried multiple images and it appears your version has a different style to the outputs i'm getting using sborquez docker version and the collab notebook. I will post some examples later today to show you what i mean - it's almost like a more blocky style, hard to explain.

shaolinseed commented 3 years ago

nerdy_alien standard_alien

shaolinseed commented 3 years ago

this is what i mean each time i get a certain style when using your repo , been trying to figure out why but not sure..

(the top image is your repo vs bottom which is based off another notebook)

nerdyrodent commented 3 years ago

Mine is based of the linked notebook, so if you're not using that that same linked notebook then that will account for the differences.

nerdyrodent commented 3 years ago

I've now added the "--cut_method" option which allows for older cutout methods to be used, such an in the pre-kornia and pooling days. Note that the aug option only works with the latest cut method.

rlallen-nps commented 3 years ago

Different resolutions will produce different output. I don't know of a way to produce the same image when using a different resolution.

Additionally, I've seen some notebooks use a learning rate optimizer, like StepLR, which can be overlooked since it's not in the normal parameters but definitely has an effect on the output.

nerdyrodent commented 3 years ago

Different resolutions will produce different output. I don't know of a way to produce the same image when using a different resolution.

That is expected behaviour. Even the same size won't generate the same image each time.