Bria drops Fibo 2 image editor to take the open-weight crown

Bria Fibo 2 is a new open-weights model designed for text-guided image generation and editing. Think of it as a tool that lets you tweak images using plain natural language instructions instead of complex mask painting.

It just took the number one spot on the Artificial Analysis image editing leaderboard, edging out the previous open-source king Qwen Image 2.1 by a few points. Bria also built the entire model on 100% licensed training data.

What it can do

Fibo 2's standout trick is multi-image compositing. You can feed it up to four reference images in a single edit and combine their distinct elements.

In testing, you can take a subject from image one, put them in a dress from image two, match the pose from image three, and drop everything into the background of image four. It successfully tracked intricate clothing textures, though it occasionally tripped up on tricky anatomy like missing feet.

It also shines at aggressive restoration tasks. When fed heavily damaged black-and-white photos with folds and cracks, it repaired the structural damage, colorized the scene, and kept the facial identity intact.

Throw a photo with destroyed white balance, crushed shadows, and aggressive blue tint at it, and Fibo 2 can pull the lighting back to its original state accurately.

That said, it still struggles with extreme edge cases. Upscaling a heavily pixelated photo of Barack Obama produced a plausible face that did not quite match the real person. Complicated character sheets also resulted in inconsistent props and disproportionately short legs.

Hardware specs and license friction

On the hardware front, Fibo 2 is accessible. The INT8 quantized build weighs in at 8.4GB, slightly larger than Qwen Image 2.1 at 7.26GB, but comfortable enough for consumer desktop GPUs.

Bria also offers an API that supports 12 aspect ratios like 21:9 and 3:1, alongside curated style presets such as kodak_film.

The catch is the license. Unlike fully open alternatives, Fibo 2 is strictly non-commercial. Grabbing the weights requires filling out an application form with detailed background information, making it significantly more restrictive than Qwen.

Why it matters

If you are an indie researcher or creative exploring multi-image blending and photo restoration, Fibo 2 is easily one of the sharpest tools available right now. The 100% licensed data lineage also makes it safe for cleanroom experimentation.

However, we think the moat here is narrower than the leaderboard suggests. Its performance only matches or slightly leads Qwen Image 2.1, but the non-commercial license and gated approval process throw up major friction for builders. Unless you specifically need its multi-image blending, Qwen remains the more practical route for production workflows.