Skip to main content
Back

Bria's Fibo 2.0 Beats 33 Closed Models With One 8B Open Checkpoint

Bria's 8B Fibo 2.0 unifies image generation and editing, tops open-weights leaderboards, and ships trained only on licensed data.

  • Bria released Fibo 2.0, an 8B open-weights model unifying image generation and editing.
  • Ranks #1 open-weights on AA-Image-Editing and tied #1 on AA-Image-T2I.
  • Driven by structured JSON prompts (VGL) instead of loose text captions.
  • Trained only on licensed data; outputs are C2PA-signed with provenance.
  • Hosted API costs $32 per 1,000 images, undercutting Luma and Qwen Pro.
  • Weights are CC BY-NC 4.0; commercial use needs a paid Bria license.

Bria has released Fibo 2.0, an 8-billion-parameter open-weights model for text-to-image generation and instruction-based editing. One checkpoint handles five image workflows, accepts structured JSON prompts, and was trained on a corpus that Bria says consists entirely of licensed images.

In the published Artificial Analysis snapshot, Fibo 2.0 ranks first among open-weights models on AA-Image-Editing v2.1. Fibo 2.0 ranks second among open-weights models on text-to-image generation. Bria also says the editing model outranks 33 closed models. Elo rankings reflect comparative preferences and can move as votes accumulate, so teams should validate the model against their own prompts, source images, and quality criteria.

Elo chart showing Fibo 2.0 leading the image-editing benchmark
Fibo 2.0’s published result on AA-Image-Editing v2.1.

Five workflows share one checkpoint

Earlier FIBO releases focused on text-to-image generation. Fibo 2.0 combines generation, instruction editing, masked inpainting, multi-reference composition, and reference-guided generation in one deployment. Production teams can route several creative operations through the same model and infrastructure.

Mode Inputs Typical use
Generate Text prompt or structured scene Create an image from a described scene.
Edit Source image and instruction Change an object, material, color, text element, or layout.
Masked edit Source image, mask, and instruction Restrict changes to a selected region.
Multi-reference Up to four reference images Combine products, garments, subjects, or visual elements.
Inspire Reference image and prompt Borrow selected subject, style, or composition attributes for a new scene.

JSON turns prompts into a control surface

Fibo 2.0 uses Bria’s Visual GenAI Language, or VGL, to describe scenes through JSON fields for objects, lighting, composition, camera settings, style, and depth of field. The original FIBO training recipe used structured captions reaching roughly 1,000 words, giving the model explicit labels for visual attributes that prose prompts often leave ambiguous.

The open-source bria-vgl package expands short prompts and edit instructions into the schema expected by the model. Developers can store, inspect, version, and modify individual fields, which makes prompt changes easier to audit. Fixed seeds and settings can improve repeatability, although GPU kernels, dependency versions, and inference settings may still change the resulting pixels.