Image LoRA

An image LoRA is the same low-rank adaptation idea used for text models, applied to a diffusion model's U-Net (and sometimes its text encoder). You take a base model like SDXL or Flux, train a tiny set of adapter weights on 10-200 reference images of a style ("Studio Ghibli backgrounds," a specific person's face, a brand's product photography aesthetic), and ship the result as a small file you load at inference. The usage pattern is very different from text LoRAs. Text LoRAs are typically fine-tuned for a task or domain and used one at a time, often merged into the base. Image LoRAs are stacked: a typical ComfyUI or Forge workflow might load a base SDXL checkpoint, then apply a "cinematic lighting" LoRA at strength 0.7, a character LoRA at 0.9, and a "film grain" LoRA at 0.4 — all simultaneously — and the model blends the influences. Sites like Civitai host tens of thousands of community-trained LoRAs, and a serious SD/Flux user might have hundreds locally. The reason the ecosystem ships thousands of LoRAs and not full fine-tunes is the same reason text LoRAs took over: they're cheap to train (a few hours on a single 4090), small enough to share freely, and composable. Flux LoRAs in particular have become the dominant way to inject specific subjects or aesthetics — Flux's prompt adherence is so strong that a well-trained LoRA can lock in a character's face across hundreds of generations. As with ControlNet, this whole world only exists because Stable Diffusion and Flux ship open weights. You can't add a LoRA to Midjourney.