Bria AI has released Fibo 2.0, an 8-billion-parameter open-weights model that generates and edits images from structured, JSON-style captions. According to its Hugging Face model page, it ranks first among open-weights image-editing models and second among open-weights text-to-image models on Artificial Analysis as of October 20, 2026. The catch for builders: the weights are licensed for non-commercial use only, and the widely shared "$32 per 1,000 images" figure is a hosted-API price attributed to a third party.
This post covers what Fibo 2.0 is, what the leaderboard chart actually shows, how to run it, and which claims you can check yourself.
TL;DR: Fibo 2.0 at a glance
| Question | Answer |
|---|---|
| What is it? | 8B flow-matching image model that generates and edits |
| Variants | Fibo 2.0 Dev (full, fine-tunable) and Fibo 2.0 Edit Turbo (4-step distilled editor) |
| Rank claimed | #1 open-weights editing, #2 open-weights text-to-image (Artificial Analysis, Oct 20) |
| T2I score | 1038 Elo vs Qwen-Image-2.1 at 1041 |
| License | CC BY-NC 4.0, non-commercial; commercial needs a separate Bria agreement |
| Access | Gated on Hugging Face; Bria API, fal, Spaces, ComfyUI nodes |
| Price claim | $32 per 1,000 images via Bria API (attributed to Artificial Analysis) |
| Training data | Licensed data from Bria's roughly 1B-image dataset, per the model page |
What is new: structured captions, one model for generate and edit
Most image models take a short prompt. Fibo was trained on long structured captions, in some cases over 1,000 words, that spell out lighting, composition, color, camera and depth of field, built on Bria's Visual GenAI Language (VGL) and described in the FIBO technical paper. Editing is treated as a change to that structured description rather than a free-form instruction, which is why edits can stay local: change the jacket, keep the face.
The model page describes the architecture: a flow-matching model working in the FLUX.2 VAE latent space, with prompts encoded by Qwen3-VL-4B-Instruct, a two-layer Perceiver resampler compressing the result into 256 to 512 "gist" tokens, and a 44-block single-stream DiT. One set of weights handles edit, masked edit, multi-reference edits with up to four images, text-to-image, and an "Inspire" mode that borrows style or composition from a reference.
The upcoming LoRAs listed on the page, including multi-turn local editing, background removal, product shots and image expansion, are not out yet.
What the leaderboard chart shows
The model page reproduces the Artificial Analysis text-to-image board, filtered to open-weights models.
Read it carefully. Qwen-Image-2.1 is first at 1041 Elo and Fibo 2.0 second at 1038, a three-point gap whose error bars overlap, so "ties for the lead" is a fair description and "beats" is not. Third place, Ideogram 4.0 (Quality), is at 1025. These are blind preference votes in an arena, which measure what voters like on the prompts they saw, not accuracy on your use case. For the competitor in first place, see our guides to Qwen-Image-2.1's transparent-image generation and license and the 8-step Qwen-Image-2.1 Turbo checkpoint.
On editing, the model page claims first place among open weights but gives no numeric scores. A social summary of the Artificial Analysis posts says Fibo 2.0 ranks 14th in both the consumer and frontier editing categories overall (open and closed combined), which would place it well behind the best closed editors. We could not open those posts ourselves, so treat that as secondhand.
Bria's own example strip on the model page shows one input photo edited into a farm scene, a steampunk workshop and pixel art while the subject stays recognizable. Vendor examples are chosen to look good; test your own images.
Claimed vs verified
| Claim | Source | Checkable? |
|---|---|---|
| #1 open editing, #2 open T2I | Bria model page, Artificial Analysis chart | Chart verified on the page; check the live leaderboard |
| 1038 vs 1041 Elo | Chart image | Yes, on the chart; error bars overlap |
| $32 per 1,000 images on Bria API | Attributed to Artificial Analysis | We could not open the original post; a price directory lists older FIBO at $0.04 per image |
| Cheaper than Luma ($100) and Qwen ($75) per 1,000 | Social summary of Artificial Analysis | Unverified here |
| Trained only on licensed data | Bria model page | A Bria statement; not independently audited |
| Open source | Model page says open weights | Weights are open, but CC BY-NC 4.0 is not an open-source license in the usual sense |
That last row is worth pausing on. "Open weights" here means you can download and run them for non-commercial work. If you are building a product, you need Bria's commercial agreement, and the hosted API at the claimed price is the other route.
How to run Fibo 2.0
From the model pages:
- Accept the conditions on Hugging Face (the repository is gated).
- Install diffusers from the pull-request branch that adds
BriaFibo2Pipeline(PR 14991, not yet merged per the page), and installbria-vgl. - Load the pipeline in bfloat16 on a CUDA GPU, use sizes that are multiples of 16 near one megapixel, and pass a structured prompt produced by
bria-vgl. - Dev: about 50 steps with guidance scale 5.0. Edit Turbo: 4 steps, guidance scale 1.0, negative prompts have no effect, editing only and not meant for fine-tuning.
Two practical flags: bria-vgl, which writes the structured prompt from plain language, depends on a Gemini API key, so your prompts pass through another provider unless you author JSON yourself. And no inference provider currently hosts the model on Hugging Face, per the page.
Why structured prompts change the workflow
With a conventional image model, the prompt is a sentence and the model decides everything you did not say. If you ask for "a red jacket on the model" in an edit, the model may also shift the background, the skin tone or the lighting, because it has no explicit record of what to keep. Fibo's approach is to keep a written description of the whole image, with fields for subject, lighting, camera angle, color palette and depth of field, and to express an edit as a change to one field. The model was trained to treat that description as the source of truth, so untouched fields tend to stay put.
For a product team this matters in three ways. First, repeatability: the structured description can be stored, diffed and versioned, so a campaign's look can be regenerated or tweaked later. Second, auditability: you can see exactly what the model was told. Third, automation: a pipeline can change a single attribute across hundreds of images, such as swapping a color or moving the camera angle, without prompt engineering per image.
The trade-off is friction. Writing a thousand-word caption by hand is impractical, which is why Bria ships bria-vgl to expand a short instruction into structured form. That tool uses a Gemini API key, so there is a second model in the loop, with its own cost, latency and data handling. If you need a fully local pipeline, you will have to produce the structured captions yourself or with a local language model, and quality may differ from what Bria's examples show.
A short test plan before you commit
Leaderboards use someone else's prompts. To decide whether Fibo 2.0 fits your work:
- Collect 20 real images and the edits you actually need, such as recolor a product, remove an object, change a background, or place the same item in a new scene.
- Run each through Edit Turbo for speed and through Dev for quality, and record time per image and VRAM on your GPU.
- Run the same set through the model you use today and compare identity preservation, text rendering and how often an edit changes areas you did not ask it to touch.
- For generation, write five prompts in your brand style and test both short text and structured captions to see how much the structure helps.
- Confirm your use is covered by the license. If it is commercial, price the Bria agreement or the hosted API against what you pay now before you build anything.
Keep the sample small and honest: twenty images will tell you more about fit than a one-point Elo difference.
Fibo 2.0 vs the alternatives
| If you need | Consider | Why |
|---|---|---|
| Commercial use with open weights | FLUX.3 Klein open weights or Qwen-Image-2.1 | Check each license; Fibo 2.0's weights are non-commercial |
| Local, precise, region-level edits | Fibo 2.0 Dev or Edit Turbo | Structured captions and masked edit |
| Fast, cheap hosted editing | Krea 3 editing or Nano Banana 2.1 | Hosted prices and speed published |
| Research or fine-tuning | Fibo 2.0 Dev | Full, non-distilled base |
Prices and rankings in other posts come from their own sources and dates; compare like for like.
Questions people are asking
Is a non-commercial open model useful? Yes for research, prototyping, fine-tuning experiments and personal projects. It is not for shipping a paid product without a license.
Why does JSON prompting matter? Structured captions make attributes explicit, so an edit can change one field without rewriting the whole prompt. The cost is a prompt-authoring step, which bria-vgl automates with a Gemini key.
Does licensed training data make it safe? It lowers one kind of legal risk, according to Bria's own statement. It does not address misuse of edited images, an issue we covered in jailbreak tests of image models.
What is still unknown
- The original Artificial Analysis price and ranking posts, which we could not retrieve directly.
- Independent speed and VRAM figures for Dev and Edit Turbo.
- When the diffusers pipeline merges and when the promised LoRAs ship.
- Commercial license terms and pricing.
Related reading
- Qwen-Image-2.1 transparent images and license
- Qwen-Image-2.1 Turbo: 8-step open checkpoint
- FLUX.3 Klein open weights guide
- Krea 3 image editing vs Nano Banana Pro and GPT Image 2
- Nano Banana 2.1 half-price image model
- Open-source image editor Compositor
Rankings and licenses are as listed on the model pages on October 20, 2026 and may change.