Finetuning Latent Upscale


I would like to finetune a Latent Upscale Model from the Pipeline (Latent Upscale Diffusers) but I couldn’t find any resource on its training. I built it similarly to the training code of the normal T2I fine-tuning (without LoRA). I get the model to train but I get the typical diffusion artifacts on my outputs (just much worse) meaning my training in fact makes the model worse. Could anybody share some light on this and maybe give some tips on how to transform the inference code from the pipeline into code for actually training such a model?

Thank you.

I’m looking into a similar thing, did you have any luck fine tuning your upscaler?