How to Generate Realistic People by Dozens of Attributes

The customer needed to generate a dataset of photorealistic images of people based on dozens of specified attributes such as age, gender, facial features, and hair. The main difficulty was controlling so many attributes, as putting everything in one prompt caused conflicts and lost features. The project was done by the Singularis team before the current generation of generative services. They compared DALL-E, Midjourney, and various Stable Diffusion versions, and the customer chose a local solution based on Stable Diffusion. The pipeline used multiple approaches: text-to-image for base images, image-to-image for adding or changing features, ControlNet and IP Adapter for reference-based control, and Compel for long prompts. They also created an attribute compatibility matrix to decide which features could be set together and which should be added in stages. Splitting complex combinations into stages helped maintain facial similarity when adding accessories like glasses or changing gender. They also prepared keyword recommendations for each attribute to improve predictability.