SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

FIBO - Open-source image generation model, the first with native JSON support

Author: AI小集 · Original source: AI工具集

What is FIBO

FIBO is the first open-source natively JSON-supportingtext-to-imagemodel, trained specifically for long structured descriptions. The model was trained on over 100 million structured JSON descriptions (each about 1,000 words), enabling precise, repeatable control over lighting, composition, color, and camera parameters. FIBO supports three modes: generation, refinement, and inspiration, and has feature disentanglement capability, allowing individual adjustment of a certain attribute without disrupting the overall scene. FIBO uses 100% licensed data, ensuring compliance and legal transparency, and is suitable for professional workflows.

FIBO

FIBO's main features

  • Text-to-image generation: Generates high-quality images based on the text description input by the user.
  • Structured JSON prompts: Expands short text prompts into detailed structured JSON descriptions, including details such as lighting, composition, and color.
  • Iterative controllable generation: Supports generating images from short prompts, or multi-round refinement based on existing JSON prompts.
  • Feature disentanglement control: Individually adjusts a certain attribute (such as camera angle) without disrupting the overall scene.
  • Inspiration mode: Extracts structured prompts by inputting images, generates related images, and sparks creativity.
  • Enterprise-grade compliance: 100% uses licensed data, ensuring legal transparency and repeatability.
  • Production-grade integration: The model supports API interfaces, ComfyUI nodes, and local inference.

FIBO's technical principles

  • Architecture: Based on an 8B-parameter DiT architecture, using Flow Matching training.
  • Text encoder: Uses SmolLM3-3B, paired with the innovative DimFusion conditioning architecture, to enable efficient long-description training.
  • VAE: Uses Wan 2.2, responsible for image encoding and decoding.
  • VLM guidance: Expands short text prompts into detailed structured JSON prompts through a vision-language model (VLM).
  • Structured supervision: Trains using structured JSON descriptions, promoting feature disentanglement and avoiding prompt drift.
  • Data compliance: Trained on over 100 million licensed long structured JSON descriptions, ensuring data compliance.

FIBO's project address

  • GitHub repository：https://github.com/Bria-AI/FIBO
  • HuggingFace model library：https://huggingface.co/briaai/FIBO
  • Online experience Demo：https://huggingface.co/spaces/briaai/FIBO

FIBO's application scenarios

  • Professional design and creative workflows: Generates high-quality images for advertising, product design, and graphic design, supports rapid iteration and precise control, and improves creative efficiency.
  • Film, television, and entertainment: FIBO can generate concept art and scene designs for films, games, and animations, helping visual creation and accelerating the development process.
  • Education and training: The model can generate teaching images and virtual experiment scenes, assist educational content production, and improve the learning experience.
  • Scientific research: The model can transform scientific data into intuitive images, assisting research presentation and data visualization.
  • Healthcare and health: FIBO can generate medical diagrams and virtual surgery scenes, supporting medical teaching and surgical training.