Closed green-anger closed 11 months ago
Some info from check_inputs:
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
...
# DemoFusion specific checks
if max(height, width) % 1024 != 0:
raise ValueError(f"the larger one of `height` and `width` has to be divisible by 1024 but are {height} and {width}.")
The first check is redundant in this case.
This still leaves the question open, what are the supported resolutions? Is it the only limitations to the size? Why is it there in the first place, algorithm limitations?
Hi @green-anger. Thanks to our community, many img2img demos can be found now, e.g., https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL.
As for the supported resolution, any aspect ratios are supported, but you need to make sure the larger one of height
and width
has to be divisible by 1024. You can understand it as only integer times upscaling is supported.
Most of the examples you provided are squares, but some have different orientations, portrait and landscape. What are the supported resolutions (or in this case better say about aspect ratio), are they the same as for SDXL?
The seconds part is mostly a dup from other issues, I just want to expand it a bit. Is it possible to generate an image with a SDXL model and then use it as input in DemoFusion to either increase level of detail (keep the same resolution) or upscale? Can DemoFusion be considered upscaling + ControlNet Tile model replacement?