LoRA

LoRA stands for Low-Rank Adaptation, but at the practical level it is a way to teach an image generation model something new without paying the cost of full retraining. You collect 10 to 30 images of whatever you want the model to learn, a person's face, a specific illustration style, a product's appearance, and run a training process that produces a small file (typically a few hundred megabytes). That file encodes the new knowledge and can be applied on top of any compatible base model.

In the Stable Diffusion and Flux communities, LoRAs are the primary way creators customize outputs. The Civitai community platform hosts hundreds of thousands of them: anime art styles, specific character designs, photographic aesthetics, fictional IP looks. You download one, load it alongside a base model in a tool like ComfyUI or AUTOMATIC1111 (a popular Stable Diffusion web interface), and the model starts generating images in that style.

For builders, LoRAs are significant because they offer a lightweight path to brand-consistent image generation. Instead of prompting your way to the right look every time, you train a LoRA on your brand's visual identity and use it as a persistent style anchor. Adobe Firefly's Custom Models feature works on a similar concept for creators who don't want to manage model files directly. The key limitation is that LoRAs are usually model-specific: one trained for Stable Diffusion 3.5 won't work on Flux without retraining.