FIBO is a text-to-image foundation model developed by Bria AI, designed for deterministic image generation and enterprise-grade production. Named after the Fibonacci sequence to symbolize mathematical precision, the model introduces a JSON-native architecture aimed at eliminating the unpredictability of traditional prompting. By mapping creative intent to structured, machine-readable controls, FIBO enables users to achieve consistent, reproducible results suitable for professional workflows and automated pipelines.
The model is built on an 8-billion parameter architecture utilizing Diffusion Transformers (DiT) and flow-matching techniques. Its technical stack includes the SmolLM3-3B model as a text encoder and Wan 2.2 as the Variational Autoencoder (VAE). Notably, FIBO was trained exclusively on a dataset of 100% licensed and rights-cleared content from over 20 premium partners, ensuring its outputs are commercially safe and compliant with the EU AI Act.
Technical Architecture & VGL
A core innovation of the model is the Visual GenAI Language (VGL) paradigm. VGL replaces ambiguous natural language prompts with explicit JSON schemas that disentangle over 100 visual attributes, such as lighting, camera angle, focal length, and composition. This architecture allows for independent adjustments to specific elements without causing unintended cascading changes to the rest of the scene, providing a level of control similar to traditional rendering engines.
Operational Modes
FIBO supports three primary modes of operation to facilitate different creative needs:
- Generate: Automatically expands short natural language prompts into highly detailed, 1,000-word structured JSON schemas to produce high-fidelity images.
- Refine: Allows users to provide iterative instructions to update specific JSON attributes, enabling precise edits without prompt drift.
- Inspire: Extracts visual parameters and stylistic cues from a reference image to create variations while maintaining thematic or brand consistency.