This version vs SD 1.4(CompVis)

[yuhaoliu7456]

I test two images using SD 1.4 from CompVis and this consistencydecoder, and is seems that no obvious difference between of them. Anyone else found this?

[Godofnothing]

@yuhaoliu7456 could you provide more context? I tried myself and the outputs of SD v1.4 VAEDecoder and ConsistencyDecoder differ significantly.

Code:

vae = AutoencoderKL.from_pretrained(
    "CompVis/stable-diffusion-v1-4", 
    subfolder="vae",
    revision="fp16", 
    torch_dtype=torch.float16,
).to(device)

sample_vae = vae.decode(latent).sample.detach()

sample_consistency = decoder_consistency(latent)

VAE

ConsistencyDistillation

[yuhaoliu7456]

@yuhaoliu7456 could you provide more context? I tried myself and the outputs of SD v1.4 VAEDecoder and ConsistencyDecoder differ significantly.

Code:

vae = AutoencoderKL.from_pretrained(
    "CompVis/stable-diffusion-v1-4", 
    subfolder="vae",
    revision="fp16", 
    torch_dtype=torch.float16,
).to(device)

sample_vae = vae.decode(latent).sample.detach()

sample_consistency = decoder_consistency(latent)

VAE ConsistencyDistillation

Can you provide this original image and I can test it by myself? Thanks

[Godofnothing]

@yuhaoliu7456 Sure, you can upload it from this url:
https://img.championat.com/c/900x900/news/big/p/l/real-madrid_1651732892490413154.jpg

[Kallamamran]

Is it possible to load a local VAE encoder instead of pointing to Huggingface?

[Godofnothing]

@Kallamamran if you have directory with model weights and config.json you can try:
AutoencoderKL.from_pretrained(path_to_local_dir, subfolder="vae" or no subfolder)