lora for diffusion

LoRA for diffusion enables efficient fine-tuning to learn specific styles, subjects, or concepts with minimal resources. **Application**: Customize Stable Diffusion for particular characters, art styles, objects, or domains without training from scratch. **How it works**: Add low-rank decomposition matrices to attention layers, train only these small adapters (~4-100MB), freeze base diffusion model weights. **Training setup**: 5-50 images of target concept, captions describing each image, few hundred to few thousand training steps, single consumer GPU (8-24GB VRAM). **Hyperparameters**: Rank (typically 4-128), learning rate, training steps, batch size, regularization images. **Trigger words**: Use unique identifier in captions ("photo of sks person") to activate learned concept. **Comparison to DreamBooth**: LoRA is more efficient (smaller files, less VRAM), DreamBooth may capture subject better but requires more resources. **Community ecosystem**: Civitai, Hugging Face host thousands of LoRAs for styles, characters, concepts. **Combining LoRAs**: Can merge or use multiple LoRAs with weighted contributions. **Tools**: Kohya trainer, AUTOMATIC1111 integration, ComfyUI workflows. Standard technique for diffusion model customization.