v0.1
Unlike the previous test version this was trained a) more aggressively b) directly on Synta-XL v2 checkpoint (unlike previous version that was trained on illustrious-xl v0.1).
This version is more stylized due to stronger training.
Concept I really like: "round window, stained glass, patterned glass".
----------------------------------------------------------------------------------------------------
Basics to LoRa:
Trained on 80 images generated with Chroma by adcom: https://tensor.art/u/754389913230900026/posts
The goal was to improve sdxl composition by using images that have good aesthetic proportions, are already in sdxl format (1344 x 768) and thus won't be cropped during training via bucketing and loose compositional power afterwards.
Thus I recommend to use this LoRa strictly at vertical 1344 x 768 format. Actually I don't recommend using this LoRa at all! It's just a test.
Next time I'll train with images without text and perhaps a bigger dataset.