Adobe has opened the public beta of a key feature in Firefly: custom models. This tool is designed for users and companies to train a generative AI model with their own visual assets. The goal is to generate images while maintaining consistency in style, color palette, or character design, addressing the usual unpredictability of these tools. The models are private by default.
How custom model training works 🤖
The process is based on uploading a set of your own images that define the desired visual identity. Firefly's system analyzes this data to learn specific attributes such as color schemes, illustration styles, or typographies. Once trained, the model generates new content that replicates those characteristics. Adobe states that the data used for training and the generated results remain in the user's private environment, without being used to train public models.
Goodbye to six-fingered hands and surreal watermarks 👋
With this, we might leave behind the golden era of logos floating in cosmic soups and corporate characters with three arms. Now, AI will make brand consistency errors with the precision previously reserved for interns. You can generate a hundred images of your mascot and they will all have the same corporate blue tone, even if it still has a slightly psychotic gaze. It's an advance: at least the visual schizophrenia will be consistent.