Adobe Launches Customizable AI Models: Train Firefly on Your Own Data for a Unique Style

Adobe has announced the public beta release of Firefly Custom Models, a new generation of generative AI tools that allow users to train models on their own visual materials. The solution is designed for creative professionals and companies looking to maintain a consistent aesthetic for characters, illustrations, or photographs when generating content at scale.

The key feature of the new offering is the ability to fine-tune the model to a unique style. Instead of starting from scratch every time, creative teams can “teach” the AI to replicate signature strokes, color palettes, lighting conditions, and character traits.

Before training begins, Adobe asks users to confirm they have the rights and permissions for the uploaded materials. The documentation emphasizes that uploaded data must not infringe on copyright, image rights, or the privacy of third parties.

By default, all custom models are private. Adobe guarantees that images used to train them will not be used to improve public foundation models. In a press release, the company stated: “To grow a brand, you need a steady stream of assets that consistently express your essence. After training, your custom model becomes part of your workflow, allowing you to create content at scale without losing what makes your work unique.”

How It Works

To create a custom model, users need to upload between 10 and 30 images. The system evaluates the suitability of the dataset using a “Model Score”: the higher the quality and consistency of the source images, the more accurately the AI will replicate the style.

Currently, three training methods are available:

1. Photography Style — The model learns unique lighting parameters, color grading, and overall mood to generate new images with the same visual feel.
2. Illustration Style — Fine-tuning based on a characteristic drawing style, textures, and execution technique.
3. Characters — Creating a consistent character that can be used across different scenes and stories while maintaining recognizable traits.

Uploaded files must meet technical requirements: JPG or PNG format, a minimum resolution of 1024×1024 pixels, and a consistent aspect ratio (up to 16:9 for landscapes or 9:16 for portraits).

The introduction of customizable models is expected to significantly streamline workflows in high-volume content projects, ensuring visual consistency without the need for manual adjustments to each element.