Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Images not reproducible using/not using StylePile #42

Open
mikebind opened this issue Jan 15, 2023 · 2 comments
Open

Images not reproducible using/not using StylePile #42

mikebind opened this issue Jan 15, 2023 · 2 comments

Comments

@mikebind
Copy link

Expected behavior: If I take the PNG info from an image created while using StylePile and enter it into txt2img tab, it should reproduce the same image (given that the model is the same, etc.)

Actual behavior: I am entirely unable to reproduce images created using StylePile using txt2img. Here are two images, the first created using StylePile, and the second created using the PNG info from the first image entered into txt2img.
00026-3304665564
00029-3304665564
Same model (SD 1.5 inpainting). Auto1111 up to date as of today. StylePile extension up to date as of today (extension update check says latest). I wondered if this might be a VAE effect, but messing with VAE settings didn't help (I got VERY minor changes to the txt2img version, much smaller than the changes between the StylePile version and the txt2img version). The images are similar enough that I assume the starting latent seed is really the same between the two. I did not change the width, height, sampling method, or sampling steps. Batch size and count were unchanged at 1 and 1. No face restoration, tiling, or hires fix. Absolutely nothing set in StylePile except for the [X] sequential prompt set to change the step at which prompt editing occurred. Set batch count to prompt count was checked, and increase seed with batch size was checked. All I can think is that maybe there is some newer setting which is not being passed along by StylePile, or there's some old default which is being put in. In any case, everything sufficient to reproduce (given the same model weights) should ideally end up in the PNG info, regardless of the input, but that is definitely not happening in this case. Also, just to make sure it's not something odd about my system, I did confirm that I have no problem reproducing images that were generated with txt2img by loading their PNG info, just ones generated using StylePile.

@mikebind
Copy link
Author

OK, I learned a bit more. I was able to almost reproduce the StylePile image using txt2img, but there is still some odd behavior. I realized that the StylePile script was still selected on the txt2img tab when I sent over the PNG info from the PNG info tab, instead of being changed back to "None". When I changed the script back to "None", txt2img generated an image which much more closely matches the one generated using StylePile. It's not quite pixel-perfect, but it's close enough that I would not have noticed at a glance that they were different (unlike the original pair). This is the new txt2img version:
00036-3304665564
If I test by selecting the StylePile script instead of "None" and press "Generate" to run txt2img (with nothing set in any of the StylePile selections and nothing to substitute in the prompt), I get this output:
00037-3304665564
(notice the clear changes to the hat and sleeves). Again, I think the expected behavior would be no difference between these two images; they have the same prompts same settings, and I haven't asked StylePile to make any changes at all, but still the outputs are subtly different. What is going on here?

@some9000
Copy link
Owner

Sorry for the super late reply. I think this may be related to the fact that Stylepile adds a bunch of "default" negative prompts that "according to my research" you would almost always want to use. Guess a checkbox or such could be added to select whether the user wants that. Otherwise if you take the same ones and go back to StylePile they would stack.

Actually I may need to look into what happens with the exact same prompts with and without it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants