FIBO Lite

FIBO Lite is the performance-optimized variant of the FIBO (Foundation for Image Background Optimization) model family developed by Bria AI. It is a specialized latent diffusion model designed for high-speed, enterprise-grade image generation and refinement. The model distinguishes itself by being one of the first JSON-native text-to-image systems, allowing users to interact with it using a structured format rather than traditional natural language prompts, which minimizes the unpredictability often referred to as "prompt roulette."

The architecture of FIBO Lite is based on a Diffusion Transformer (DiT) framework utilizing flow-matching techniques. It integrates the SmolLM3-3B model as its text encoder and the Wan 2.2 VAE. To achieve high inference speeds, FIBO Lite employs a two-stage distillation process: Classifier-Free Guidance (CFG) distillation and Shortcutting Flow Matching (SCFM). These optimizations enable the model to produce high-quality, high-fidelity images in as few as 4 to 8 inference steps, making it particularly effective for real-time applications and on-premise deployments.

A primary focus of FIBO Lite is legal and ethical compliance for commercial use. The model was trained exclusively on a dataset of over 100 million fully licensed images, ensuring that generated content is rights-cleared and compliant with regulatory standards like the EU AI Act. This "safe-for-commercial-use" foundation is coupled with a deterministic generation process where each image comes with a machine-readable JSON "recipe," ensuring auditability and reproducibility across professional workflows.

Prompting FIBO Lite involves the Visual GenAI Language (VGL) paradigm, which uses structured JSON schemas to define over 100 visual attributes. These include explicit parameters for lighting (e.g., "backlit," "soft morning light"), camera settings (e.g., "85mm lens," "f/1.8"), and composition. By modifying specific values within the JSON schema, users can perform targeted refinements—such as changing the color palette or camera angle—without causing "prompt drift" or unintended alterations to the rest of the scene.