SDXL TLDR: 1. SDXL seems to vastly outperform v1.5 and v2.1 in terms of user preferences and is even (roughly) tied with @midjourney v5.1!! 2. SDXL is BIG (2.6B Unet params) --> slower + more VRAM than prev SD 3. Two CLIP txt-encoders instead of one (conditioning vectors are concatenated) resulting in better img-txt alignment 4. Slightly improved VAE 5. Much better ways of dealing with low-res training imgs (condition the model on img-size), random cropping (condition the model on crop location) and non-square imgs (condition model on aspect ratio) 6. SDXL has an optional refinement stage trained specifically on denoising small amounts of noise (when there's already lots of signal) from high quality images.