Bria drops Fibo 2.0 Edit Turbo for 4-step image editing

Fibo 2.0 Edit Turbo is an 8B-parameter open-weights image editing model released by Bria AI. In plain terms, it is designed for precise, surgical image modifications: altering clothes, swapping colors, inserting objects, adjusting lighting, or painting masks to fix specific spots without drifting the rest of the image.

As of October 20, 2026, it ranked #1 among open-weights models on the Artificial Analysis Image Arena for image editing.

What you can do with it

Anyone who has tried editing with AI knows the pain of scene drift: you ask to tweak a shirt color, and the person's face and background suddenly change. Fibo 2.0 Edit Turbo is built to eliminate that drift.

First, precise attribute modification. You can turn white sneakers into matte black while keeping the seams, silhouette, and scuffs completely untouched. Turn a summer landscape into a snowscape, and the building structures stay locked while snow settles naturally on roofs and lawns.

Second, seamless additions and removals. Prompt it to put a party hat on a cat, and the hat sits properly between the ears with matching lighting. Tell it to erase an object, and the void fills in seamlessly with context-aware textures.

Third, masked editing. You can paint a quick black-and-white mask over the source image. The model redraws only the white area, leaving every surrounding pixel intact.

Fourth, multi-reference edits supporting up to four images. You can pass a portrait and a photo of a dress, instructing the model to put that specific garment onto the subject. It accounts for body pose, folds, and ambient lighting rather than pasting a flat sticker.

How it delivers 4-step editing

Fibo 2.0 is an 8B flow-matching model operating in the FLUX.2 VAE latent space. It uses Qwen3-VL-4B to encode text and visual context, running through a 44-block single-stream DiT architecture.

Its spatial stability comes from Visual GenAI Language (VGL). Instead of relying on vague natural language, the model treats images as structured JSON blueprints with distinct fields for camera angles, lighting, background, and objects. Editing simply applies a clean diff to that blueprint. Bria provides the open-source `bria-vgl` package to translate everyday prompts into structured JSON on the fly.

The speed comes from distillation. Fibo 2.0 Edit Turbo cuts out classifier-free guidance entirely and converges in just 4 inference steps. Community benchmarks show a 16GB consumer GPU can process a 1000px image in around thirty seconds.

Specs, hardware, and getting started

The weights are hosted on Hugging Face under a CC BY-NC 4.0 non-commercial license. Personal experimentation, research, and local tinkering are open, while commercial deployment requires an agreement with Bria. The model was trained entirely on licensed data.

For local execution, DiffSynth-Studio supports memory offloading, making it possible to run on NVIDIA GPUs with as little as 8GB of VRAM by paging to system RAM. You can run it via Diffusers, ComfyUI nodes, or test it immediately on Hugging Face Spaces and fal.ai.

Why it matters

If you build automated design pipelines, manage e-commerce catalog variations, or need complete data privacy for client photos without sending files to commercial cloud APIs, Fibo 2.0 Edit Turbo is an immediate download. 4-step generation makes interactive local editing genuinely responsive instead of a waiting game.

We think Bria's combination of structured JSON disentanglement and distillation points where practical generative workflows are heading. The main friction is handling prompt formatting, but as an open-weights local tool, it sets a high bar for deterministic image manipulation.