Why Change VAE?
#25
by marcdemory - opened
What is the purpose of using this VAE instead of the default one on the SDXLControlNet pipeline?
Does it generate better results, improve memory usage, etc.
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
We kept the vae in fp16 during training and iirc there are issues with the standard sdxl vae in fp16. I believe https://huggingface.co/madebyollin/sdxl-vae-fp16-fix has more details
williamberman changed discussion status to closed