Fibo 2.0

Fibo 2.0 is an open-weights image generation and editing model developed by Bria, featuring 8 billion parameters. It unifies text-to-image generation and complex image editing tasks within a single set of model weights. Built upon Bria's Visual GenAI Language (VGL) framework, Fibo 2.0 is trained on long, structured captions and exclusively utilizes licensed data. The release includes the full base model, Fibo 2.0 Dev, designed for high-quality control and fine-tuning, as well as Fibo 2.0 Edit Turbo, a distilled checkpoint optimized for fast, few-step production editing.

Architecture and Capabilities

Fibo 2.0 employs a flow-matching architecture centered around a single-stream transformer backbone. Unlike traditional systems that treat generation and manipulation separately, the model reads structured prompts or free-form text instructions that translate into structured document modifications, ensuring high fidelity during edits. Key capabilities include instruction-based general editing, support for masked regions, and multi-reference inputs that accept up to four reference images to preserve consistent subject identity, layout, and lighting. Furthermore, the model demonstrates significant improvements in text rendering, fine surface textures, and human anatomy compared to prior iterations.