Glossary
LoRA Training Basics
The step-by-step process of teaching a diffusion model a new face, object, or style using a small set of training images.
Training a LoRA requires three things: a base model (e.g., Flux Dev), a dataset of images, and a training script (Kohya-ss, SimpleTuner, or a managed training service).
Dataset preparation is the most important step. For a face LoRA: collect 15–30 images showing the subject from varied angles, lighting conditions, and expressions. Crop tightly to the face or slightly wider; remove backgrounds if possible. Mislabelled or low-quality images hurt more than adding good ones helps — quality beats quantity.
Captioning: each training image needs a text caption. Auto-captioners (BLIP-2, Florence-2) generate initial captions; you then prepend a unique trigger token (e.g., 'ohwx person') that the LoRA will learn to associate with the subject. During inference, including that token 'activates' the learned identity.
Hyperparameters: key settings are learning rate (1e-4 to 1e-5), network rank (r=16–64 — higher rank = more expressive but larger file and more overfitting risk), and number of training steps (500–2000 for faces). Overfitting produces images that look exactly like the training photos but fail to generalise to new poses and prompts.
Synexa skips training altogether: a creator's face is locked from reference photos, so there is no dataset to caption and no LoRA file.
Frequently Asked Questions
- How many images do I need to train a good LoRA?
- 15–30 images is the practical sweet spot for face LoRAs. Object or style LoRAs may need 50–200 images with greater variety.
- What is a trigger word and do I need one?
- A trigger word (or token) is the unique string you include in prompts to activate the LoRA's learned concept. It is required for identity LoRAs and optional for style LoRAs.
- How long does LoRA training take?
- On a single A100, a face LoRA typically trains in 10–30 minutes.
- Can I train on images that contain multiple people?
- Training images should ideally show only the target subject. Multiple people in frame confuse the model about which identity to learn.