Quetion about training_method

[PotatoBananaApple]

I have been trying to find information on the training methods: noxattn, innoxattn, selfattn, xattn, full, xattn-strict, noxattn-hspace, noxattn-hspace-last

But i can't just quite find any information on how each of them affects the training. If you could pinpoint to resource or briefly explain how does it affect the slider training i would be really happy!

[rohitgandikota]

All the models in our paper are trained on default setting noxattn which only attaches LoRAs on all layers except cross attention layers in the UNET.

Similarly following our previous work "Erasing Concepts in Diffusion Models" we released some other setups

• xattn means sliders are attached only on cross attention layers
• full is for all layers in the UNET
• selfattn is for self attention layers

The remaining are experimental which are not well tested

[PotatoBananaApple]

All the models in our paper are trained on default setting noxattn which only attaches LoRAs on all layers except cross attention layers in the UNET.

Similarly following our previous work "Erasing Concepts in Diffusion Models" we released some other setups

* `xattn` means sliders are attached only on cross attention layers

* `full` is for all layers in the UNET

* `selfattn` is for self attention layers

The remaining are experimental which are not well tested

Thank you for answer! Gotta test around different settings!