With Fibo 2.0, Bria bets on image control

Bria introduces Fibo 2.0, an 8-billion-parameter AI model combining generation, targeted editing, and JSON-based image control.

Modifying the background of an advertising photograph without touching the product, replacing a prop, or rolling out a campaign across multiple formats: these operations require precise control over the changes made to an image. With Fibo 2.0, Bria AI combines visual generation and editing into an eight-billion-parameter model trained exclusively on licensed data, according to the company. Available as accessible weights, APIs, and integrations for creative workflows, it relies on a structured description language that allows defining scene characteristics and the elements to be modified. In leaderboards published as of October 20, 2026, Bria claims the top spot among open-weight models for image editing. An Image Described as a Set of Parameters

The distinctiveness of the Fibo family stems from its use of the Visual GenAI Language (VGL). Rather than solely passing an instruction written in natural language, the system converts the request into a structured JSON description. Objects, their position, lighting, colors, composition, photographic characteristics, and typographic elements can be described separately.

The process addresses a common difficulty in AI image creation: modifying an attribute without causing unexpected changes elsewhere in the composition. When a user requests different lighting or a new background, they generally want to preserve the product's appearance, its label, and other previously approved elements. With VGL, modifications are formulated as changes to specific scene properties, rather than as a new, free-form description of the entire image.

Bria provides a tool, bria-vgl, capable of expanding a short prompt into a detailed schema, without requiring the user to manually write JSON. The structured document can then be inspected, modified, and reused. This representation facilitates tracking campaign variations, even if it does not guarantee that a new generation will reproduce every pixel identically. Generate and Edit with the Same Weights

With Fibo 2.0 Dev, text-based generation and the editing of an existing image rely on the same set of weights for the first time in the Fibo family. The model accepts a scene description just as easily as a photograph accompanied by an editing request. Users can replace objects, modify materials, adjust lighting mood, or work on text within the image.

Editing can be guided by up to four images: a primary image and up to three additional references. For a perfume campaign, for example, the first image can provide the bottle to be preserved, the second a lighting mood, and the third a desired composition. References can be explicitly designated in the prompts to avoid confusing their roles.

Another mode uses a mask to restrict modifications to a defined area. Pixels located outside the mask are preserved, allowing an object to be replaced without regenerating the entire visual. This function applies to single-image edits. Multi-reference operations use different controls, particularly for selecting the final output from among several aspect ratios.

Bria also distinguishes between Generate, Refine, and Inspire modes. The first two serve to build or modify an image using structured descriptions. The third analyzes a reference to extract certain stylistic or compositional features, then generates a new visual based on that description. In this latter case, the generator does not receive the reference image directly: it operates based on the text information generated upstream. Two Versions for Quality and Speed

The family includes two main checkpoints, intended for different constraints. Fibo 2.0 Dev is the full, non-distilled model designed for generation, editing, and adaptation to specialized needs. Its recommended configuration relies on 50 inference steps with classifier-free guidance set to 5.0. It also serves as a base for LoRA training or more comprehensive fine-tuning for a brand identity or photographic domain.

Fibo 2.0 Edit Turbo is a distilled version geared toward fast editing. It executes its modifications in four inference steps, compared to fifty for the Dev model's reference configuration. This reduction targets workflows where numerous variations must be produced, notably product catalogs and advertising assets adapted for multiple platforms.

The distinction is therefore not just about speed. Dev constitutes the general-purpose model on which developers can experiment and train adaptations, while Edit Turbo prioritizes low latency for editing operations. Bria claims an execution time of under two seconds for the model and under five seconds end-to-end on its own API. These measurements do not constitute a latency guarantee for all environments, especially third-party services. High Editing Results, to Be Interpreted According to Rankings

Bria highlights Fibo 2.0's results in Artificial Analysis evaluations. According to the AA-Image-Editing v2.1 leaderboard dated October 20, 2026, and reproduced by distribution platforms, Fibo 2.0 holds the first spot among open-weight models in image editing, with an Elo score of 1,087. The model card also presents it in second place among open-weight models in text-to-image generation.

These results concern two distinct categories. A high ranking in editing evaluates the quality of transformations applied to an existing image, whereas text-based generation relies on other criteria. Elo scores measure relative preferences within an evaluation arena; they can evolve as new comparisons are collected and do not guarantee perfect fidelity on every image.

Bria particularly emphasizes progress regarding text embedded in visuals, material textures, and human anatomy compared to previous versions. The company also highlights the preservation of elements that should not be altered. These capabilities are important for commercial uses, but their performance must be verified on the specific products and visual identities of each project. An Eight-Billion-Parameter Architecture

Fibo 2.0 is based on a