cloneofsimo/lora

Decomposes weight updates into low-rank matrices (ΔW = AB^T) applied primarily to attention layers, reducing fine-tuned model size to 1-6MB while maintaining or exceeding full fine-tuning quality. Implements three distinct training approaches—LoRA-DreamBooth (with prior preservation), Textual Inversion, and Pivotal Tuning Inversion—enabling flexible control over style vs. identity trade-offs. Integrates with Hugging Face `diffusers` library and supports inpainting, CLIP+UNet+token joint training, and checkpoint merging for composable style combinations.

7,529 stars. No commits in the last 6 months.