Proposal
Hey guys,
I’d like to request a small but useful enhancement regarding seed handling in multi-stage pipelines (particularly HiRes Fix)
Currently, Inference uses a single seed across the full generation pipeline. When using multi-pass workflows:
[example: Base generation >> HiRes Fix >> additional HiResFix],
This can sometimes lead to undesirable artifacts when doing multiple upscale/refinement passes, like: blotching or loss of fine detail, etc... even when other parameters remain unchanged.
Changing the seed per pass tends to produce a more stable and "consistent" results.
this issue does not happen with all models, if you're looking to recreate this particular problem the Anima base v1 is an ideal choice, the other model that benefited from this was either Z-Image Base or Z-Image Turbo,
Just do two path of hires fix at x1.5 (using upscaler of your choice.)
Proposal
Hey guys,
I’d like to request a small but useful enhancement regarding seed handling in multi-stage pipelines (particularly HiRes Fix)
Currently, Inference uses a single seed across the full generation pipeline. When using multi-pass workflows:
[example: Base generation >> HiRes Fix >> additional HiResFix],
This can sometimes lead to undesirable artifacts when doing multiple upscale/refinement passes, like: blotching or loss of fine detail, etc... even when other parameters remain unchanged.
Changing the seed per pass tends to produce a more stable and "consistent" results.
this issue does not happen with all models, if you're looking to recreate this particular problem the Anima base v1 is an ideal choice, the other model that benefited from this was either Z-Image Base or Z-Image Turbo,
Just do two path of hires fix at x1.5 (using upscaler of your choice.)