>>83923608> If anything, it should just learn to produce slightly noisy image every timeI worded it wrong. I meant exactly what you are saying. It will learn to always leave some noise, because inputs always have some noise. So instead of full noise at timestep 1000 -> fully denoised image at timestep 0 as it should be with zsnr, it will produce a noisy image even at timestep 0. It's like chopping off 100 timesteps from one side and stretching the rest. Technically, that's not a znsr schedule anymore?
It's fine now, but it will become an issue if training is very long. Zoom in and you'll see how this noise looked on ef-hll
>not sure if it even works for sd-scriptsit should just work if you install it and add --optimizer_type "optimi.Lion" --optimizer_args "weight_decay=0.01" ...
But it looks like things like gradient_release need some changes
https://github.com/kohya-ss/sd-scripts/pull/1381>>83923782I see. I only quickly skimmed though that that discussion on github. I should read it properly