Bria uploaded Fibo-2.0-dev to Hugging Face on 18 October 2026 and Fibo-2.0-edit-turbo on 19 October. A Comfy-Org repack followed on the 20th and is the reason the model showed up in our tracker at all: Comfy-Org/Fibo-2.0 went from nothing to 70,682 downloads by this morning, a 24.5% day-over-day gain.
The model is interesting. The licence is more interesting, and the two pull in opposite directions.
What it is
FIBO 2.0 is an 8B-parameter flow-matching model that both generates and edits images from a single set of weights. The architecture, from Bria’s own card:
- Operates in the FLUX.2 VAE latent space
- Prompts encoded by Qwen3-VL-4B-Instruct
- A two-layer Perceiver resampler compresses the VLM output into 256–512 “gist” tokens
- Those feed a single-stream DiT with 44 blocks
- Weights are BF16
It ships as two checkpoints. Dev is the full, non-distilled base — classifier-free guidance, recommended at 50 steps and guidance scale 5.0, and intended as the fine-tuning starting point. Edit Turbo is distilled from Dev for few-step, low-latency editing and needs no guidance.
Bria claims FIBO 2.0 Dev ranks #1 in image editing and #2 in text-to-image among open-weights models on Artificial Analysis, as of 20 October 2026. That is the company’s own claim citing a third-party leaderboard snapshot; we have not independently reproduced it, and a one-day snapshot of a fast-moving arena is a weak foundation for a durable claim.
The actually novel part: prompts are JSON
FIBO was trained on structured JSON captions of up to 1,000+ words, describing lighting, composition, colour, camera and depth of field as named fields rather than prose. Bria calls this the Visual GenAI Language (VGL) paradigm. Negative prompts are JSON too.
This is a genuine departure. Most text-to-image prompting is an exercise in guessing which adjectives the captioner used. A structured caption makes the controls addressable: you change the lighting field rather than appending “cinematic lighting, volumetric, 8k” and hoping. For anyone building a tool on top of a generator — which is most of our readers who use these things in production — an editable, reusable prompt object is far more tractable than a string.
You are not expected to hand-write the JSON. The bria-vgl package generates it from a short prompt.
It requires a Gemini API key.
Read that in context. The selling point of a downloadable model is that it runs on your hardware, under your control, with your data. FIBO’s ergonomic path to its own prompt format routes through Google’s API. You can write the JSON by hand and stay local — the format is documented — but the convenience layer is a cloud dependency bolted to a local model.
The licence
Here is where the pitch and the product separate.
Bria’s provenance claim is the strongest in the field and it is not hedged: the model is trained “exclusively on licensed data”, curated from Bria’s roughly 1B-image dataset, with “100% licensed training data”, governance and attribution traceability. Bria’s commercial proposition has always been indemnified output for customers who cannot risk a training-data lawsuit, and FIBO 2.0 is that proposition as open weights.
Except:
The model is released under a CC BY-NC 4.0 license for non-commercial use.
Non-commercial. Commercial use requires a separate agreement with Bria. The Hugging Face metadata tags the licence as bria-fibo, and the API reports license: other on both repos — not Apache, not CC-BY, not any of the permissive licences the phrase “open weights” tends to imply.
Both repos are also gated (gated: auto on the API). You fill in a form and agree to the non-commercial conditions before you can download.
So the position is: the one image model whose training data is clean enough to use commercially without legal exposure is the one you are not licensed to use commercially. The provenance and the permission are decoupled. If you want both, you are back to a sales conversation — which, to be fair to Bria, is a coherent business model and not a bait-and-switch. It is just the opposite of what “trained on 100% licensed data, open weights” sounds like at a glance.
Third-party coverage has this wrong in both directions, which is why we checked. One inference host describes it as trained only on licensed data for commercially safe output — true of the output under a commercial agreement, misleading about the weights. A model-listing site gives the licence as CC-BY-NC-4.0 and says commercial use is not permitted — correct, and the one worth trusting.
Who it is for
Use it if: you are doing research, personal work, or evaluation; you want structured, editable prompt control; you want an editing model that is the same weights as your generator; or you are assessing whether to pay Bria for the commercial grant.
Do not reach for it if: you need to ship client work tomorrow on weights you control. For that, the permissively licensed Apache and MIT image models are less capable and more usable, which is the trade the whole field is currently making.
The pattern
This is the fourth model in two weeks where the advertised openness and the enforceable openness differ, and the gap has been in a different place each time: Voxtral 3 had no weights at all, GenXR-Bench’s “code available” link 404s, RLMuscle’s repo is real but carries no licence, and FIBO’s weights are real, licensed, downloadable and non-commercial.
The lesson is not that anyone is lying. It is that “open” has stopped carrying information, and the only reliable move is to read the licence field and try the download.
Sources
- briaai/Fibo-2.0-dev — Hugging Face
- briaai/Fibo-2.0-edit-turbo — Hugging Face
- Comfy-Org/Fibo-2.0 — Hugging Face
- Fibo 2.0: Bria’s 8B open-weights model edits and generates images — ComfyUI Wiki
- Introducing FIBO: Structured Control for Text-to-Image Generation — fal.ai