The Vision Language Model (VLM)
Analyzes visual intent and outputs a structured JSON representation—geometry, lighting, composition, attributes—all explicit, all inspectable.
Open. Modular. Yours.
No Black Box. No Monolith. No Lock-In. Full visibility into every model. Open weights you can run anywhere. A modular architecture that lets you use exactly what you need.
The Problem with Monoliths
Most visual AI platforms are opaque bundles. You get everything or nothing-with no visibility into what's happening inside.
This creates risk:
Black box outputs
Can't explain, audit, or reproduce results
Forced bundles
Pay for capabilities you don't need
Vendor lock-In
Switching costs grow with every integration
Deployment constraints
Their cloud, their rules
See Inside the Model
Fibo isn't one model—it's two, working in sequence.
The foundation model
Renders the final image from that structured state. No ambiguity. No interpolation from vague prompts.
What this means:
You can inspect, modify, and version the intermediate state- before a single pixel is rendered.
No guessing. No "prompt and pray." Full observability.
Traditional models
Prompt in → image out (black box)
Can't see why you got what you got
Reproduce results? Good luck
Reproduce results? Good luck
Fibo
Prompt in → structured state → image out
Full visibility into every parameter
Deterministic: same state = same output
Deterministic: same state = same output
One Language, Multiple Models
The Fibo family shares a common control language. Only execution characteristics change. Same paradigm. Different problems solved. No paradigm shifts when requirements change.
Fibo generation
Full image creation from structured state
Optimized for: Quality, prompt adherence
Fibo edit
Targeted mutations to existing visual state
Optimized for: Precision, consistency
Fibo fast
Production-grade generation
Optimized for: Latency, throughput
Fibo edge
Lightweight on-device generation
Optimized for: Size, offline capability
Fibo tuned
Custom-trained on your data
Optimized for: Brand fidelity
Fibo brand
Style-locked generation
Optimized for: Visual consistency
Auxiliary models
RMBG-2.0 is Bria's segmentation model, designed to preserve information rather than discard it. Instead of binary masks, it outputs continuous alpha mattes—retaining fine-grained boundary detail.
Clean compositing. Reliable background replacement. Seamless integration into structured pipelines.
Part of the modular toolkit—use it standalone or as a building block in larger workflows.
Open source
Bria is committed to open, responsible AI development. Our foundation models are available on Hugging Face with full transparency—no black boxes, no hidden constraints.
What's available:
Model weights - Download and run Fibo models, RMBG-2.0, and more.
Source code - Full implementation, not just inference.
Training recipes - Reproduce, extend, customize
Documentation - Comprehensive guides and examples
Commercial + Open: Why It Matters
Open source builds trust. Commercial support builds production systems. With Bria, you get both:
Evaluate freely
Download models, run locally, test against your requirements
Self-serve
For startups under $1M ARR or $25M funding
Scale with support
Enterprise pricing and dedicated solutions
Community
1M+ downloads on Hugging Face
Active contributor community
Peer-reviewed research (NeurIPS, CVPR)