VAE configuration issue

[spezialspezial]

Describe the bug

It seems vae configuration might be broken in recent commits (~7f58a76f). Even though I can verify the correct config contents are read in I still end up with the wrong vae config. Most notably the scale_factor for SDXL models is that of SD1.5 resulting in washed out generations. I suspect the error to be somewhere in create_diffusers_vae_model_from_ldm but not sure at the moment. Just me? Any ideas?

Reproduction

1. Instantiate SDXL pipe from single file.
2. Check your vae scaling factor. If it's not 0.13025 there might be an issue

Logs

No response

System Info

diffusers 0.26.0.dev0 7f58a76, Linux, Python 3.10.12 with variadic function support, torch 2.1.2+cu121 + cuda 12.1

Who can help?

@sayakpaul

[spezialspezial]

looks like the VAE always ends up with the default argument value from create_diffusers_vae_model_from_ldm and the yaml config scale_factor item is never used. See create_vae_diffusers_config

[sayakpaul]

Cc: @DN6

[DN6]

@spezialspezial Could you provide a code example I could use to test please?

[spezialspezial]

sure thing

pipe = StableDiffusionXLPipeline.from_single_file("YOUR_PATH/sd_xl_base_1.0_0.9vae.safetensors")
print(pipe.vae.config.scaling_factor)

[DN6]

Should be fixed with #6779

[spezialspezial]

Awesome, much appreciated!