SourceSDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
The paper identifies a reproducible failure mode in SD 1.5 and SD 2.1: generated objects appear cropped or truncated at the edges of the image (e.g., a cat's head cut off). The authors attribute this to random cropping used during training as a data augmentation step, which leaks into the generation process. SDXL addresses this by conditioning on crop coordinates, and the paper shows that SDXL does not exhibit this artifact under the same prompts.