>>8654037
the "best" snr-based timestep weighting scheme you can do in sd-scripts for sdxl vpred is snr / (snr + 1)**2 and you can achieve that if you use 'debiased estimation' and 'scale vpred loss like noise pred' together (green line), without min snr
(well if you don't count the bug in sd-scripts which doesn't let snr reach zero for weighting even with zero terminal snr enabled, any snr-based weighting doesn't make sense for ztsnr)