from pixiv HIKA

v2.0 update : no difference in datasets, but use text encoder for general use, and increase total learning steps

v3.0 : training using only his latest images, dataset does not overlap with v2.0

v2.0 vpred update : I trained it on the v2.0 dataset (500 images), using noobai vpred as the base model. It's an unrefined dataset crawled from Pixiv, so rather than converging on a single art style, it feels more like a hodgepodge. Anyway ¯\_(ツ)_/¯

v3.0 vpred update : I trained this version by adding a bit more data to the existing v3.0 dataset (150 images). As an experiment, I tested training at 1280x1280 resolution. The sample images were also all generated at resolutions of 1024x1024 or higher.

( v3.0 IL | v3.0 vpred )

It seems to maintain the original body proportions 'slightly' better at high resolutions. It's not a dramatic difference, but it gave me an interesting idea to make multiple 1280-resolution LoRAs and merge them into a model.