How can I train from scratch?

ali-vilab / AnyDoor

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

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

MIT License

3.95k stars 359 forks source link

How can I train from scratch? #109

Open mao-code opened 1 day ago

mao-code commented 1 day ago

I noticed that the "resume_path" in the run_train_anydoor.py file should be placed with the pre-trained anydoor checkpoint. What if I want to train the anydoor model from scratch?

How can I get an un-trained anydoor weights? (I have already had the SDv2.1, ControlNet for SDv2.1 and DINO weights) Is there a model structure class so that I can train the whole modules(ControlNet, linear projector after DINOv2 and UNet)?

Thank you very much!

mao-code commented 14 hours ago

I noticed that the anydoor.yaml contains all the model structure. For the projector layer, it can be found in the last part of the config file or this file "https://github.com/ali-vilab/AnyDoor/blob/main/ldm/modules/encoders/modules.py"