[SD 1.5][LORA]AYEISHA v1.1

0xSoul

Test model AYEISHA.

Shocasing a new training method I developped that I call "multires-training". The goal is to use the same images but sized at different resolution to train the LoRA. As you can see this model is able to produce "ok" faces and details when the subject is small in the overall picture. Even small squinting eyes are better rendered.

This is possible because I trained the model at 512x512, 384x384, 256x256 and 128x128 resolution. In the picture where the model is smaller it allow SD to use the data learned from smaller resolution that more closelly match the size of the model in the picture.

At the moment the code required to train with multiple resolutions is in the dev branch: bmaltais/kohya_ss at dev (github.com)

You can use the tools/resize_images_to_resolutions.py to create the multi-resolution dataset for training.

Let me know what you think.