Ideogram is rolling out a new feature called Custom Branded Models, letting companies upload a library of 15 to 100 branded images and get a custom AI model trained on their specific visual style. The system automatically generates labels, handles the training, and spits out a model that can then generate new, on-brand images in bulk.
Once training is complete, you get a custom_model_uri that plugs directly into Ideogram's existing generation API—no need to manage GPU infrastructure or worry about the underlying compute. It's a significant upgrade from the previous Style Reference feature, which only allowed uploading up to three reference images for style hints. Now you're training a dedicated model on an entire image library.
The API now supports the full training workflow: creating a dataset, uploading images, kicking off training, and running batch generation.
https://twitter.com/ideogram_ai/status/2046993830952771724
The company emphasizes that the quality of your image labels matters more than the sheer number of images. It recommends adding descriptive text to each image to better align the model's output with your intended style and semantics.
The feature is available to Team and Pro subscribers, with deeper customization offered for Enterprise customers. Sneaker brand Skechers has already used it to generate over 100,000 branded assets.
Ideogram was founded by researchers who worked on Google's first text-to-image system, and the product has long been known for its strong text rendering capabilities. This move into custom model training is part of a broader trend where AI platforms are offering more specialized tools for enterprise clients, similar to how Google's AI Studio is expanding options for paying subscribers and Anthropic is securing massive compute resources to scale its models.