Fibo-2.0-dev: Image generation and editing in a single 8B model

Handles text-based image generation and image editing within a single model. Bria's Fibo 2.0 Dev is an 8B parameter-scale model that, unlike conventional approaches, integrates separate pipelines for generation and editing to operate as a single deployment. It ranked #1 in image editing and #2 in text-to-image among open-weight models according to Artificial Analysis.

The core is structured captions based on the Visual Generation AI Language (VGL). By explicitly defining attributes such as lighting, composition, and camera parameters in JSON format, it enables precise editing that modifies only specific attributes while keeping the rest unchanged. Using the bria-vgl package automatically converts short prompts or natural language instructions into these structured prompts.

It offers improved quality over previous versions in text rendering, texture representation, and human anatomical accuracy. It also supports mask-based editing and multi-reference synthesis using up to 4 reference images. As an undisstilled base model designed for professional fine-tuning, it serves as a starting point for LoRA training or full fine-tuning.

The license is CC BY-NC 4.0, providing open-source access for non-commercial purposes only. Commercial use requires a separate agreement with Bria. It is trained on 100% licensed data, featuring legal clarity regarding copyright and attribution tracking.