Replies: 1 comment
-
|
I think I just realized that you can set the plugin to generate at the full resolution when inpainting by changing the setting "Preferred Resolution" under Styles, checking "Override," and entering the desired pixel resolution. That will prevent the plugin from generating multiple images at the resolution of the selected area, and instead upscale the selected region so it is near the optimal resolution, do the generation, and then downscale it back into the masked area, exactly what I was talking about. So this option is already present, I just didn't know how to enable it. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
In the documentation it says:
Other Diffusion UIs sometimes offer options to generate at a higher resolution and downscale results to fit the image, to “improve quality/details”. This is an anti-pattern: downscaling means keeping a lower quality version of the result that was generated. It is always better to upscale existing low-resolution content instead.I think the reason many other diffusion UIs use that approach is because many models (pre-Flux) are trained on images at a specific resolution, such as SDXL at 1024x1024, and anything much smaller than that or non-square often results in inferior janky generations. You can do larger batches when inpainting smaller resolutions, but the results are often much worse than if you generated just one image at the full resolution of the model, and then scale it back down to fit the inpainting region. You often have to make many inferior generations that takes far longer than just one or two high quality generations. As suggested in the docs, you could scale up the whole canvas to generate inpainting at higher resolutions, but there are times when users may just want to work on inpainting small details at the full resolution capabilities of the model, and don't want to scale up the entire canvas to extremely large dimensions to do so. They might also not care that those generations are downscaled to fit the image.
I think such an option would be a great feature for those who want to always use the full resolution capabilities of the selected model.
Beta Was this translation helpful? Give feedback.
All reactions