Taesd for SD 1.5

[andreineamtu]

Do you have a version for Stable Diffusion v1.5 that can be loaded using from_single_file method as VAE?
I tried it and got this error:
"diffusers.loaders.single_file_utils.SingleFileComponentError: Failed to load AutoencoderKL. Weights for this component appear to be missing in the checkpoint."

[madebyollin]

No, TAESD is not an AutoencoderKL model so you can't load TAESD using AutoencoderKL.from_single_file.

To load TAESD in diffusers, you can instead use AutoencoderTiny.from_pretrained, like this:

import torch
from diffusers import AutoencoderTiny
taesd = AutoencoderTiny.from_pretrained("madebyollin/taesd", torch_dtype=torch.float16)

For more detailed information, see the AutoencoderTiny diffusers docs.

[andreineamtu]

Yes that's how I did, but this means to load it from HuggingFace.
I wanted to load it from my local machine. Is that possible?

[madebyollin]

Yeah, from_pretrained supports local folders, so you can download the taesd folder with git clone https://huggingface.co/madebyollin/taesd and then construct the model from that taesd folder using AutoencoderTiny.from_pretrained("/path/to/taesd").