A (kinda failed) attempt to train an SDXL LoRA model upon on a mixed style dataset of images from Dreamworks, Disney and Pixar. I have uploaded this<a rel="ugc" href="https://huggingface.co/datasets/Norod78/DreamDisPix-blip2-captions"> Blip2-Captioned Dataset to Huggingface</a>. I usually prefer to train in LoRA ranks 4-16 because so far it always seemed to be enough and provided a reasonably small checkpoint but in this case, I only started getting results when I went for Rank-32, even tried 24 and it was not enough. My conclusion is that it should have probably been trained as several different lower ranked LoRAs each for the specific visual sub-style and it would have probably looked way better for each.

Dream Dis Pix XL

A rank 32 SDXL LoRA trained upon <a rel="ugc" href="https://huggingface.co/datasets/Norod78/DreamDisPix-blip2-captions">this dataset</a>

Files with SHA256: 8233dbcf24a610701686c5eb096aa02aa18f7c5d14db9009e53fcc9973acff00