An unofficial extension that implements Custom Diffusion for Automatic1111's WebUI.
Custom Diffusion is, in short, finetuning-lite with TI. Instead of tuning the whole model, only the K and V matrices of the cross-attention blocks are tuned simultaneously with token embedding(s). It has similar speed and memory requirements to TI and supposedly gives better results in less steps.
You can find the UI in the Train/Train Custom Diffusion tab. Just train as you would a normal TI embedding. Under the training log directory, alongside with name-steps.pt you should also see name-steps.delta.safetensors, which contain finetuned delta weights (~50MB at half precision uncompressed).
Custom Diffusion proper includes regularization. To generate regularization images, go to Custom Diffusion Utils/Make regularization images. You can then optionally supply the generated images directory as regularization when training.
The trained deltas will be under models/deltas (--deltas-dir); you can also copy over logged .safetensors versions. The delta weights can be used in txt2img/img2img as an Extra Network. You can select them under the extra networks tab like hypernets. Use the token embedding like a normal TI embedding.
This is an unofficial implementation based on the paper and the features and implementation details may be different to the original.