LoRA (low-rank adaptation)
Related tutorialA LoRA (low-rank adaptation) is a small, lightweight file that adjusts a base image model so it reliably reproduces one subject, character or style. Instead of retraining the full model, which needs huge datasets and compute, a LoRA learns from a handful of reference images and stores only the changes, making it fast to train and easy to share.
In tendre.red, you train your own LoRA from 10 to 15 photos of a character, plus a trigger word, in the app's LoRA training tab. WD14 auto-captioning helps the training read identity and framing, and you choose the training engine: Local, on your own GPU, or Cloud, on tendre.red's servers for credits. Once trained, load the LoRA in LoRA generation and use the trigger word to produce new images of that same subject in any scene or pose.