FIBO 2.0
Bria's 8B image model that generates and edits images from structured JSON prompts, trained on licensed data. Non-commercial weights.
In short. FIBO 2.0 is an image model from the Israeli company Bria. One model can create a picture from a description or edit an existing photo by instruction, and it works best when its prompt is a detailed JSON blueprint of the scene, which a helper tool writes for you. Bria says it trained only on licensed images, but the free weights are for non-commercial use only.
Summary
Bria AI, a Tel Aviv company that sells visual generative AI trained on licensed data, published FIBO 2.0 on Hugging Face in October 2026. It is an 8-billion-parameter flow-matching model, and for the first time in the FIBO family one set of weights handles both text-to-image and editing. It works in the latent space of the FLUX.2 autoencoder, and a Qwen3-VL-4B-Instruct model encodes prompts and the context images used for editing. A single-stream transformer with 44 blocks does the drawing. The editing side covers instruction edits, masked edits and combining up to 4 reference images, and the text-to-image side offers five photographic style modes, such as Kodak Film and Hard Flash.
The model is built around structured prompts. It was trained on JSON captions of up to 1,000 words, written in what Bria calls Visual GenAI Language, which list lighting, composition, color, camera and every object as separate fields. You normally do not write them. Bria's bria-vgl package turns a short prompt, or an image plus an instruction, into the JSON, and its examples call Gemini as the language model, so full-quality use needs a hosted LLM or a substitute you test yourself. There are two checkpoints. Fibo 2.0 Dev is the full base, run with 50 steps and a guidance scale of 5.0, and a single-GPU LoRA trainer for editing and text-to-image comes with it. Fibo 2.0 Edit Turbo is distilled for few-step, guidance-free editing. Each repository holds a 16.3 GB transformer, an 8.9 GB text encoder and a small autoencoder, about 25.5 GB in all, and the pipeline needs Diffusers 0.42 or newer.
Bria says FIBO was trained only on licensed data, curated from its dataset of about one billion images, and its website names Getty Images, Alamy and Envato as sources. The weights are released under CC BY-NC 4.0 for non-commercial use, and commercial use needs a separate agreement with Bria. The download is gated on Hugging Face, with a form that asks for your intended use and is approved immediately. Bria says FIBO 2.0 ranked first for image editing and second for text-to-image among open-weights models on the Artificial Analysis Image Arena on 20 October 2026, which is the developer's claim and a snapshot. The model is a niche choice for now. In the 30 days to 11 November 2026, Hugging Face counted about 2,600 downloads of Dev and 3,300 of Edit Turbo.
Good for
- One model that both generates images and edits them by instruction, with masks and up to four references
- Work where you want to change one attribute, such as camera angle or lighting, without the rest of the scene drifting
- Research and fine-tuning on a base trained on licensed images, using the included LoRA trainer
- Non-commercial projects, evaluation and academic work
Watch out for
- The weights are non-commercial, and any commercial use needs an agreement with Bria
- Best results need the bria-vgl helper, which calls a language model such as Gemini to write the JSON prompt
- The download is gated and about 25.5 GB with its text encoder, and the pipeline needs a recent Diffusers
- The licensed-data statement is Bria's, and the open weights come with no output indemnification that we found
- The leaderboard ranking is Bria's claim and dated 20 October 2026
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of FIBO 2.0. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Fibo 2.0 Dev (base, 50 steps) BF16 | 8.13B | 26 GB | 29 GB | Gaming GPU |
| Fibo 2.0 Edit Turbo (few-step editing) BF16 | 8.13B | 26 GB | 29 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.
- Attribution or model name must be shown
See also
- Best open-source image and video generation models
- Krea 2: an open image model trained from scratch for variety
Frequently asked questions
- What is FIBO 2.0?
- FIBO 2.0 is an image model from the Israeli company Bria. One model can create a picture from a description or edit an existing photo by instruction, and it works best when its prompt is a detailed JSON blueprint of the scene, which a helper tool writes for you. Bria says it trained only on licensed images, but the free weights are for non-commercial use only.
- Can I use FIBO 2.0 commercially?
- Not without separate permission. FIBO 2.0 is released under the CC BY-NC, which allows research and personal use only.
- Is FIBO 2.0 open source?
- FIBO 2.0 is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run FIBO 2.0 on my own computer?
- Yes, on a good graphics card. The smallest variant, Fibo 2.0 Dev (base, 50 steps), needs about 29 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was FIBO 2.0 released?
- FIBO 2.0 was released by Bria AI on October 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Fibo-2.0-dev model card (Hugging Face) huggingface.co
- Fibo-2.0-edit-turbo model card (Hugging Face) huggingface.co
- Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions (arXiv 2511.06876) arxiv.org
- Bria AI website: licensed training data and licensing bria.ai
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Krea 2 Krea | 12.8B | – | Krea 2 Community (Conditional) | 41 GB Workstation | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU | |
| Z-Image Tongyi MAI | 6.15B | – | Apache 2.0 (Permissive) | 16 GB Laptop or phone |