What Is a LoRA?
A LoRA (Low-Rank Adaptation) is a small model that teaches an existing AI image generator a new concept — your face, a specific art style, a product, a character, or any visual concept. Instead of training a full model (which costs thousands), a LoRA adds targeted knowledge at a fraction of the cost and time.
What You Need
- 15-30 high-quality training images of your subject
- A computer with a decent GPU (8GB VRAM minimum) OR a cloud training service
- Stable Diffusion installed locally OR an account on Civitai/Replicate
- 30-60 minutes of training time
Step-by-Step Guide
Step 1: Prepare Training Images
Collect 15-30 images of your subject. For a person: different angles, lighting conditions, expressions, and outfits. For a style: varied subjects in the target style. For a product: multiple angles on plain backgrounds. Quality matters more than quantity — blurry or low-res images degrade output quality.
Step 2: Caption Your Images
Each image needs a text description (caption) that tells the AI what's in the image. Use BLIP or WD Tagger to auto-generate captions, then edit them manually. Include a unique trigger word (e.g., "ohwx" for a person, "mystyle" for an art style) that will activate your LoRA during generation.
Step 3: Train
Local (Kohya_ss): The standard LoRA training tool. Set learning rate to 1e-4, train for 1500-3000 steps, use SDXL as the base model. Training takes 30-60 minutes on a modern GPU.
Cloud (Civitai/Replicate): Upload your images and captions, select your base model, and the platform handles the rest. Costs $1-5 per training run. Easier but less control.
Step 4: Test and Iterate
Generate test images using your trigger word. If the results are too weak (doesn't look like your subject), increase training steps. If the results are "fried" (distorted, overfitted), reduce training steps or lower the learning rate. Most LoRAs need 2-3 training iterations to get right.
Popular LoRA Use Cases
- Personal avatar: Generate professional headshots, social media content, and artistic portraits of yourself
- Product photography: Train on your product, then generate it in any setting or context
- Art style: Develop a signature visual style and apply it consistently across all content
- Character design: Create a consistent character for comics, games, or branding
Legal Note
You can train LoRAs on your own images, images you have rights to, or public domain images. Training on other people's copyrighted work or likeness without permission raises legal and ethical issues. When in doubt, use only images you created or have explicit permission to use.
