Images not getting diffused correctly at 840x1176 (windows xl sm80 1 1176 unet)

[masslevel]

On a resolution of 840x1776 840x1176 (windows xl sm80 1 1176 unet) images are not getting diffused correctly. It goes away if you lower the resolution to ex. 848x1184 (windows xl sm80 1 1184 unet works) and AIT switches to a different module.

EDIT: resolution typo corrected to 840x1176

It's not seed, prompt or sampler dependent.

@Shaistrong kindly helped confirming the issue.

Tested on Windows 11, 4090 (536.67), latest version of ComfyUI and the AIT custom node

first stages during diffusion (ksampler image)

final output image

[hlky]

It goes away if you lower the resolution to ex. 848x1184 (windows xl sm80 1 1184 unet works) and AIT switches to a different module.

840x1176 is smaller than 848x1184, and should use the same module. Can you confirm the resolution that works, and which modules are being selected?
The selected module will be displayed on the terminal.

Can you try a 64px resolution such as 832x1152? There have been some reports of 8px increment not working as expected on XL models, as far as I'm aware v1/v2 is unaffected.

[masslevel]

I made a typo in the resolution values in my initial post. Sorry about that, @hlky ! (updated the first post)

The incomplete diffusion happens when using a resolution of 840x1176 (windows xl sm80 1 1176 unet).

The image diffuses correctly when using:
832x1152 (windows xl sm80 1 1152 unet)
848x1184 (windows xl sm80 1 1184 unet)

This confirms what you've said about 8px increments not working with SDXL. So not an issue with AIT or any of the modules.

Thank you for explaining it!

[Shaistrong]

@hlky can you guys add a node for empty latent image [AIT]? so people don't accidentally use 8px increments

[hlky]

@masslevel

848x1184 (windows xl sm80 1 1184 unet)

Interesting. 848 is 16px increment, 1184 is 32px increment. Whereas 840 and 1176 are only divisible by 8. It would be useful to test more 16px and 32px increments.

@Shaistrong Yes they have been re-added here

[Shaistrong]

@masslevel

848x1184 (windows xl sm80 1 1184 unet)

Interesting. 848 is 16px increment, 1184 is 32px increment. Whereas 840 and 1176 are only divisible by 8. It would be useful to test more 16px and 32px increments.

@Shaistrong Yes they have been re-added here

16px increments seem to work, it's just 8px increments that AIT doesn't diffuse correctly.

[clockworkwhale]

As requested on another forum, sharing my results testing SDXL in ComfyUI with AIT, latest builds as of time of this post. AIT consistently introduces large changes to images in the direction of reduced coherence.

Prompt "Photograph of a young woman holding an apple", basic workflow with no use of refiner. All images generated at 1024x1024.

In all cases, far left side is non-AIT image, middle and/or right side are AIT with same seed.

[masslevel]

@clockworkwhale I've also noticed differences - but only very minor ones. Maybe it has something to do with the settings (sampler etc).

You might also want to open a new issue for this since this topic is about something else.

[hlky]

@masslevel I had asked for the tests be shared on this issue. There has always been some slight numerical inaccuracy, however the differences in the above tests are more than typical. I believe it to be a related regression.

@clockworkwhale Thank you. Can you share the workflow json so I have the seed etc?