Introduction

Bria launched Fibo 2.0 on October 20, an 8B-parameter open-weights image model that does text-to-image and instruction-based image editing in a single set of weights. You can run it today in three ways. The weights are on Hugging Face as Fibo-2.0-dev (full base model) and Fibo-2.0-edit-turbo (4-step editor), behind a gate that approves you automatically and limits you to non-commercial use. It's also on fal at $0.04 per image, and on Bria's own API and Bria Create app. Bria says it trained Fibo 2.0 only on licensed data, and it's claiming the top spot among open-weights models in Artificial Analysis's image-editing arena.

The structured-prompt crowd has been waiting for this one, so let's get into it ✨

What shipped

  • Fibo 2.0 Dev is the full, non-distilled base model. It runs with classifier-free guidance, handles both generation and editing, and is the checkpoint you fine-tune.
  • Fibo 2.0 Edit Turbo is distilled from Dev for editing only. It edits in 4 steps with no CFG (guidance_scale=1.0), and its card says it's not meant for fine-tuning.
  • Edit modes: plain instruction edits, masked edits (white marks what to regenerate), and multi-reference edits with up to 4 images, called <image_1>, <image_2> and so on in your instruction. Bria pitches these for virtual try-on, product placement, and style or appearance transfer.
  • Generate and Inspire: a short prompt gets expanded into a full structured caption before rendering. Inspire mode borrows the subject, style or composition of a reference image without the model ever seeing that image.
  • Bria's stated upgrades: better in-image text, texture, and anatomy (hands, faces, poses) than earlier FIBO releases.
  • Day-0 access: fal (bria/fibo-2/generate and bria/fibo-2/edit), each::labs, the Bria API and sandboxes, Bria Create, and two free Hugging Face Spaces (Fibo-2.0-T2I and Fibo-2.0-editing, which runs Edit Turbo).

Under the hood

Fibo's trademark is still VGL (Visual GenAI Language). Instead of a loose sentence, every image is described as structured JSON with separate fields for objects, lighting, composition, camera, style, and text to render. Generating means "render this blueprint," and editing means "apply this diff to the blueprint." That's why Fibo can change one thing, like the time of day or an outfit, without the rest of the scene drifting. You don't have to write the JSON yourself: Bria's MIT-licensed bria-vgl package (v0.8.0 on PyPI) turns plain prompts and edit instructions into VGL. Note that it runs on Gemini, so you need a GEMINI_API_KEY.

From the model card:

  • 8B-parameter flow-matching model working in the FLUX.2 VAE latent space (the diffusers PR lists the transformer at 8.1B)
  • Qwen3-VL-4B-Instruct encodes the prompt. For edits, context images go through both the VLM and the VAE.
  • A two-layer Perceiver resampler packs the VLM output into 256–512 "gist" tokens, which feed a single-stream DiT with 44 blocks. Five of those blocks also read VLM layers directly through cross-attention.
  • Trained at about 1 megapixel. Use sizes that are multiples of 16, such as 1024×1024, 1424×800 (16:9) or 1248×832 (3:2).
  • Training data was curated from Bria's ~1B-image licensed dataset and paired with long structured captions.
  • Each repo is about 25.5 GB of bf16 weights: about 16.3 GB for the transformer (two shards), 8.9 GB for the text encoder, and 0.34 GB for the VAE.

The leaderboard claim, in context

Bria calls Fibo 2.0 the #1 open-weights image-editing model on Artificial Analysis, with 1087 Elo from 4,819 blind human-preference votes, "ahead of 33 closed models from leading labs." The model card adds a #2 among open-weights models for text-to-image, as of October 20.

A friendly reality check: when we checked the public AA-Image-Editing v2.1 open-weights board shortly after launch, it hadn't listed Fibo 2.0 yet. Qwen-Image-2.1 was still on top at 1083 ±13. So the claimed lead is about four Elo points, within the usual confidence band. It's a genuinely strong debut, and a close race at the top of open-weights editing.

License: read this before you ship

This part matters, so here it is exactly as published:

  • The Hugging Face cards list the license as other / bria-fibo and link to Bria's Fibo 2.0 licensing page.
  • The card body says the model is released under Creative Commons Attribution–Non-Commercial (CC BY-NC 4.0) for non-commercial use, and commercial use requires a commercial agreement with Bria.
  • The access gate says the weights are "open source for non commercial use only." Before you download, you confirm that commercial use needs a separate agreement and that academic access is limited to non-commercial research and teaching.
  • The "clear commercial rights" in fal's and Bria's marketing applies to the hosted API route. For self-hosted commercial work, talk to Bria first.

How to run it today

On fal (fastest)

Both endpoints cost $0.04 per image. Generate takes a free-form prompt (or a vgl document), offers 21 aspect ratios from 5:1 to 1:5, and returns the image plus the VGL it rendered, so you can reuse or tweak the blueprint later. It also has four built-in styles: Kodak Film, Hard Flash, Hazy Dreamy, and Rough Realism. Edit takes 1–4 image_urls plus an instruction, with an optional mask_url for single-image edits.

import { fal } from "@fal-ai/client";

const result = await fal.subscribe("bria/fibo-2/generate", {
  input: {
    prompt: "An editorial still life on pale peach paper: a cobalt glass cylinder on the left, a cut blood orange in the center, and a folded cream linen napkin on the right. Afternoon light enters from the left, casting all three shadows toward the lower right. The glass refracts blue light onto the paper.",
    aspect_ratio: "3:2",
    seed: 5555
  }
});
console.log(result.data.image.url);

Here's a multi-reference instruction from fal's own example for the edit endpoint. Send the product shot as image 1 and a lighting reference as image 2:

Restage the serum bottle from <image_1> on a pale lavender stone plinth. Use <image_2> as a reference for the warm side lighting and peach background. Keep the bottle's shape, cap, label text, and colors consistent with <image_1>. Leave space to the right for campaign copy.

Locally with diffusers

BriaFibo2Pipeline is still in an open diffusers PR (#14991), so for now you install diffusers from the PR branch. Accept the gate on the model page and log in with your Hugging Face token first.

pip install -U "git+https://github.com/muhammadali-afridi/diffusers@ali/fibo2" torch torchvision transformers accelerate
pip install bria-vgl
export GEMINI_API_KEY="your-api-key"

This is the card's Edit Turbo quick start, merged into one script. Use 4 steps with guidance 1.0. The model also needs the two bundled fonts, because Fibo 2.0 "sees" any text in the caption drawn as a small image.

import json
import torch
from PIL import Image
from diffusers import BriaFibo2Pipeline
from huggingface_hub import hf_hub_download
from bria_vgl import modify
from bria_vgl.dialects.thales import EditImages

pipe = BriaFibo2Pipeline.from_pretrained("briaai/Fibo-2.0-edit-turbo", dtype=torch.bfloat16).to("cuda")
fonts = [hf_hub_download("briaai/Fibo-2.0-edit-turbo", name, subfolder="fonts") for name in ("Aileron-Regular.otf", "DejaVuSans.ttf")]

def vgl(scene) -> str:
    return json.dumps(scene.model_dump(exclude_none=True), ensure_ascii=False)

source = Image.open("photo.jpg").convert("RGB")
scene = modify(EditImages(prompt="replace the jacket with a red leather one", images=[source]), seed=0)
edited = pipe(prompt=vgl(scene), image=source, num_inference_steps=4, guidance_scale=1.0, fonts=fonts).images[0]
edited.save("edited.png")

For the full-quality Dev model, the card recommends num_inference_steps=50 and guidance_scale=5.0, with the negative-prompt JSON it provides. The Dev repo also includes a single-GPU LoRA trainer in fine_tuning/, which covers both editing and text-to-image. Bria hasn't published a VRAM figure yet, so plan for a large card or offloading.

No GPU? No problem

Try the free Spaces linked above, or use Bria's Generate and Edit sandboxes on platform.bria.ai.

What's next on Bria's roadmap

The card lists these as "coming soon," with no dates:

  • Multi-turn local-editing LoRA (chained edits that keep everything else pixel-stable)
  • Native background removal
  • Product Shot LoRA
  • "Boring Reality" style
  • Eraser by mask
  • Image expansion (outpainting)

For background: FIBO started in 2025 as a JSON structured-prompt text-to-image model, and Fibo-Edit-1.5 arrived in August 2026. Bria also says Fibo was named one of TIME's Best Inventions of 2026. Fibo 2.0 is the first release to put generation and editing in one model.

Original Source

Conclusion

Fibo 2.0 is easy to like. You get one 8B model for making and editing images, a JSON blueprint you can read and diff, up to four references, licensed-only training data, and pricing that's cheap enough to experiment with on fal. Just remember the split: the open weights are non-commercial unless you sign an agreement with Bria, while the hosted API is the route to commercial rights. For tinkerers and fine-tuners, it's a lovely new sandbox to play in.

—Aurelia ♡