Bria trains an 8B image model using licensed data: matching Qwen, open source but not for direct commercial use

Beating AI News Flash: AI imaging company Bria has released Fibo 2.0, with 8 billion parameters, integrating text-to-image and image editing capabilities into a single model for the first time.

Users can modify photos through natural language, and can also use structured JSON to specify character positions, lighting, and text, minimizing the risk of breaking other parts when modifying one area. It also supports mask editing and multi-image references.

Bria has also released the full Dev version, as well as Edit Turbo, which requires only 4 inference steps and is specifically designed to accelerate photo retouching.

On Artificial Analysis's AA-Image-Editing v2.1 leaderboard, Fibo 2.0 ranks first among open-source models with 1087 Elo, slightly ahead of Qwen-Image-2.1's 1083. On the text-to-image leaderboard, Qwen leads with 1042 points over Fibo's 1039. Neither gap reaches statistical significance, placing them essentially in the same tier.

Fibo 2.0 is trained entirely on licensed images. Bria has previously partnered with copyright holders such as Getty Images, paying to acquire training materials. However, the model weights use the CC BY-NC 4.0 non-commercial license — developers can download for research and fine-tuning, while commercial use requires a separate agreement with Bria.