You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have been trying out this workflow in replicate and comfy-ui and found the results to be pretty good. Only problem I see is that when pattern is enabled, it takes a lot more time (4x more time for 2x upscaling a 1024x512 image). I looked around the implementation and noticed the code was performing tiled decoding when pattern is enabled. When I tried other regular SDXL workflows, enabling circular padding on Conv2d layers for seamless patterns doesn't cost any extra time and works perfectly almost everytime. Would it be possible to have a similar implementation here as it would be better, faster, simpler and more efficient than tiling+offsetting+etc.?
If I knew where to modify the diffuser pipeline object, I would have done it myself and raised a PR but I haven't got my head around this repo setup. Thank you!
The text was updated successfully, but these errors were encountered:
Hi,
I have been trying out this workflow in replicate and comfy-ui and found the results to be pretty good. Only problem I see is that when pattern is enabled, it takes a lot more time (4x more time for 2x upscaling a 1024x512 image). I looked around the implementation and noticed the code was performing tiled decoding when pattern is enabled. When I tried other regular SDXL workflows, enabling circular padding on Conv2d layers for seamless patterns doesn't cost any extra time and works perfectly almost everytime. Would it be possible to have a similar implementation here as it would be better, faster, simpler and more efficient than tiling+offsetting+etc.?
If I knew where to modify the diffuser pipeline object, I would have done it myself and raised a PR but I haven't got my head around this repo setup. Thank you!
The text was updated successfully, but these errors were encountered: