A new affordable method to train AI image models on personal photos allows users to generate custom AI images for approximately $3. The technique, detailed in a blog post by software developer Cory Zue, combines the Flux image model with LoRA (Low-Rank Adaptation) training technology on the Replicate platform.
The process requires users to upload 10-15 diverse photos of themselves and takes less than an hour to complete. The system automatically generates image captions using large language models, eliminating the need for manual labeling. After training, users can create various AI-generated images of themselves in different scenarios.
The method utilizes Replicate’s GPU rental service and Hugging Face’s model storage platform. Training costs approximately $2.50 per model, with additional image generation priced at $0.03 per image. The author reports mixed results in image quality, noting that adding descriptive prompts like age and gender improves accuracy.
The system works best with single-subject images and shows limitations when attempting to combine the trained model with other people in generated images.