Bria Fibo 2.0 Leads Open-Weights Image Editing

The top-ranked image editor whose weights anyone can download is, narrowly, one that most businesses cannot use without signing a contract first. Bria’s Fibo 2.0, released on 20 October, sits at the top of the open-weights models on Artificial Analysis’s image-editing leaderboard, but its model card is explicit: “The model is released under a CC BY-NC 4.0 license for non-commercial use. Commercial use is subject to a commercial agreement with BRIA.”

That trade-off is the point of Bria’s pitch. The company says “Like every FIBO model, it is trained exclusively on licensed data,” drawing on what it describes as its roughly one-billion-image dataset. It is betting that companies worried about where training images came from will pay for that assurance, while researchers and hobbyists get the weights to experiment with.

One model now edits and generates

Earlier Fibo releases split the work; this one does not. “For the first time, image editing and image generation live in a single unified model,” Bria says. The same 8-billion-parameter weights handle text-to-image, edits from a plain-language instruction, edits confined to a masked area, and compositions that combine up to four reference images.

There are two checkpoints. Fibo 2.0 Dev is “the full, non-distilled base model”, meant for fine-tuning and run with 50 steps by default. Fibo 2.0 Edit Turbo is “distilled for few-step, low-latency image editing” and needs only four steps, but it cannot generate images from scratch. Under the hood, Bria describes “an 8B-parameter flow-matching model operating in the latent space of the FLUX.2 VAE”, with prompts encoded by Alibaba’s Qwen3-VL-4B-Instruct.

Fibo’s distinguishing feature is still its structured prompts. A helper package, bria-vgl, turns a short request into a long JSON description of lighting, camera and objects, so a single attribute can be changed without the rest of the scene drifting. That helper “runs on Gemini”, the card notes, so even a local set-up calls Google’s model and needs a Gemini API key. Bria’s launch notes add 12 new aspect ratios, for 21 in total, and say existing API users have nothing to change in their code; the model is also on fal.

The lead over Qwen is within the margin of error

Artificial Analysis ranks models by blind human votes on pairs of outputs. When TSN checked the board on Wednesday morning, Fibo 2.0 had an Elo score of 1,087 from 4,819 samples, 21st of all models and first among open-weights ones. Alibaba’s Qwen-Image-2.1 was just behind on 1,083, and the two scores’ 95% confidence intervals overlap almost entirely, so the ranking separates them more firmly than the data does. The clearer gain is over Bria’s own previous editor, FIBO Edit, on 913.

Bria also says Fibo 2.0 “edits better than closed models from Ideogram, ByteDance and Black Forest Labs”. That holds for Ideogram 4.5, ByteDance’s Seedream 5.0 Lite and Black Forest Labs’ FLUX.2 Pro and Max on the same board, but ByteDance’s Seedream 5.0 Pro scores well above it, and the board’s leader, a version of OpenAI’s GPT Image 2.5, is more than 100 points clear. Bria also claims second place among open-weights models for text-to-image, which TSN has not checked.

Open image models have been arriving quickly this month. Black Forest Labs put FLUX.3 Image Klein into Hugging Face’s Diffusers library with editing and layout control in one model, and Krea launched its Krea 3 editing model, pitched on speed and price.

For now, running Fibo 2.0 locally means installing Diffusers from a development branch, because Bria’s pipeline is still waiting to be merged into the library (pull request 14991). Multi-turn editing, in which each edit builds on the last, heads Bria’s “coming soon” list, with no date attached.