Replies: 2 comments 6 replies
-
Uncropped works fine. There have been issues with xformers 0.16, so make sure you are not using that (use the older 0.14 or the newer 0.17 instead). Maybe also share your training configuration and the version info you get when starting dreambooth? |
Beta Was this translation helpful? Give feedback.
-
Yeah, there is definitely no dropping of images going on. The bit of assert code in the trimming logic is just there to ensure everything is trimmed as intended. All of your instance images should be used regardless of scale. If you train using perfectly square images - then no cropping or trimming will occur. If you use non-square images, they will be scaled as a factor of the maximum resolution, with buckets created for the various "common" aspect ratios around that. Any cropping that happens should be a minimum, as I perform a combination of upscaling and cropping to try and fit the images to the target bucket(s). |
Beta Was this translation helpful? Give feedback.
-
Recently, I've been trying to get training working. I've trained with this ext successfully in the past, but this time nothing works (outputs stay the same, can't remember how many steps was it, but half an hour on 3090 ti, I think about a thousand). The thing is, this time I haven't cropped the images (and I would prefer not to). Textual inversion and hypernetworks don't work too (outputs garbage), so I suspect that the aspect ratio is to blame. Could this be the case, and if so, what's the best way to solve this?
Beta Was this translation helpful? Give feedback.
All reactions