Custom Diffusion enables few-shot fine-tuning of Stable Diffusion and other text-to-image diffusion models given 4-20 images of a new concept. The method optimizes only key and value projection matrices in cross-attention layers, enabling fast training (~6 minutes on 2 A100 GPUs) and lightweight per-concept storage (75MB). The approach supports single and multi-concept customization including object-style combinations and new category additions.