CCSRv2 results too different from original image compared to CCSRv1

[pierre1618]

i tried this model with tiles (default values) and i found out that compared to CCSRv1 , the CCSRv2 result change too much from the original image when scales are smaller than 4x, which parameters can i change to achieved that? I already tried to change the guidance scale and prompt but does not improve much .

[raulc0399]

did you find any solution? for upscale 2 i get really bad results

[csslc]

Thanks for your question. What is the resolution of your input? The SD-based SR methods are quite sensitive to the input resolution, which should ideally be larger than 512×512.

[bonlime]

Tried different input and upscale factors, CCSRv2 seems to be generating too much of gan artifacts and deviating too much from the input image. v1 seems much better in comparison