Weight mismatch with SD 2.1

ali-vilab / AnyDoor

Official implementations for paper: Anydoor: zero-shot object-level image customization

https://ali-vilab.github.io/AnyDoor-Page/

MIT License

4.01k stars 366 forks source link

Open ifeherva opened 10 months ago

ifeherva commented 10 months ago

If I understand the approach correctly the in blocks of SD 2.1 were frozen during training.

Yet, when I compare the weights between the checkpoint shared here vs the SD 2.1 on huggingface the weights, don't match. What am I missing?

For example the keys I am comparing from the 2 state dicts:

Anydoor checkpoint: _model.diffusion_model.inputblocks.0.0.weight vs. SD 2.1 loaded as UNet2DConditionModel: _convin.weight

ifeherva commented 10 months ago

I figured out what was happening, the link in the README points to SD-v2.1 however you used SD-2.1-base. I recommend to update the README :)