Conjoined Twins Even Triplets #183
-
|
I'm dumb. I don't even know how to search for this... So, I got things working and starting to understand basic environmental stuff. I don't know where to go or who or how to ask but, I'm going to give it a shot here. Trying to use 3060TI 8GB to get some consistent output between 512x512 and 1024x1024. Trying to find something that's pleasing at 512, then run 1024 for final output and the 1024 version ends up being some bizarre conjoined version of the 512 version. Very similar but, bizarre perversion of the 512. Needless to say all settings remained the same between them. Is this expected behavior or, should I go digging further? I totally understand that this is all very experimental. [2022-09-18 12:17:43,327] {transport.py:1874} INFO - Connected (version 2.0, client OpenSSH_7.6p1) [2022-09-18 12:17:43,675] {transport.py:1874} INFO - Authentication (publickey) successful! [2022-09-18 12:19:33,922] {neongradio_ultimate.py:782} INFO - prompt: engraving of golden retriever sitting, W: 512, H: 512 [2022-09-18 12:21:16,345] {neongradio_ultimate.py:867} INFO - saving images [2022-09-18 12:21:17,334] {neongradio_ultimate.py:889} INFO - memory_final = 34.033664 [2022-09-18 12:22:25,973] {neongradio_ultimate.py:782} INFO - prompt: engraving of golden retriever sitting, W: 1024, H: 1024 [2022-09-18 12:37:25,079] {neongradio_ultimate.py:867} INFO - saving images [2022-09-18 12:37:28,514] {neongradio_ultimate.py:889} INFO - memory_final = 34.033664 PLMS Sampler: 98%|██████████████████████████████████████████████████████████████████▋ | 49/50 [00:47<00:00, 1.10it/s] |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
Yes, this behavior is expected. The stable-diffusion model was created using 512x512 images. If you try to create an image with any other resolution, then the results become less stable/predictable. I suggest making a 512x512 image of your subject, then upscaling to 1024x1024 with realesrgan or something similar. Alternatively, you can use img2img to upscale to different resolutions, but again this may produce weird results. |
Beta Was this translation helpful? Give feedback.
Yes, this behavior is expected.
The stable-diffusion model was created using 512x512 images. If you try to create an image with any other resolution, then the results become less stable/predictable. I suggest making a 512x512 image of your subject, then upscaling to 1024x1024 with realesrgan or something similar.
Alternatively, you can use img2img to upscale to different resolutions, but again this may produce weird results.