Minimal inference snippet / PiEaPixelDiffusionDecoder class
#1
by alekssdev - opened
Hi! Thanks for releasing PiEa !
Could you share a minimal inference snippet (or the PiEaPixelDiffusionDecoder class)? Specifically: input pixel range/normalization (RGB vs BGR, [-1,1] vs [0,1]), the timestep/noise-level encoding fed to time.net, the noise schedule, and how the degraded condition is produced (downscale factor, blur, noise).
We reconstructed the U-Net from the key shapes and it partially works, but a sampler mismatch leaves a strong color cast and residual grain. Thanks!