Bria's FIBO 2.0 Matches Qwen on Open-Weights Image Benchmarks at Half the Price

Bria has released FIBO 2.0, an 8-billion-parameter image model that ranks first among open-weights models on AA-Image-Editing v2.1 and statistically ties Qwen-Image-2.1 for the open-weights lead on AA-Image-T2I v2.1. The external benchmark results also show gains over the original FIBO across every measured editing action and text-to-image capability.

The release contains two checkpoints: FIBO 2.0 Dev for text-to-image generation and FIBO 2.0 Edit Turbo, a distilled checkpoint optimized for faster editing. Open weights allow developers to download the model parameters, while the noncommercial license limits how those checkpoints may be used. Commercial deployment requires a separate agreement with Bria.

Top scores at a lower API price

Artificial Analysis tested FIBO 2.0 through Bria’s hosted API and lists generation at $32 per 1,000 images, or 3.2 cents per image. Its score in the cited comparison met or exceeded the three higher-priced services below. These results characterize the hosted implementation; local performance can vary with quantization, model revision, inference settings, and hardware.

Leaderboard positions and hosted prices can change as providers update their models and endpoints, so production estimates should use current pricing from the intended provider.

Strength shifts by task

FIBO’s strongest editing results come from enhancement, restoration, scene changes, and style changes. Composition, framing, and edits that require multi-step reasoning remain farther from the leading systems. Text-to-image performance follows a similar pattern: anatomy and reasoning rank relatively well, while text rendering and layout still lag despite posting the largest gains over the first generation.

JSON becomes the control surface

FIBO’s main architectural distinction is its structured prompting system. According to the repository, the model was trained on JSON captions that can exceed 1,000 words. Bria’s prompting pipeline can expand a short request into fields covering composition, lighting, color, camera settings, subjects, and other visual attributes before sending the structured description to the generator.

Bria’s paper describes this design as disentangled control: changing one field, such as camera angle, should reduce unintended changes elsewhere in the scene. Machine-readable prompts can also be stored in version control, compared as diffs, generated by agents, and modified one key at a time. Reproducible workflows should preserve the model revision, JSON prompt, random seed, scheduler, and inference settings because a structured prompt alone does not guarantee identical pixels.

An 8B diffusion transformer

FIBO uses a diffusion transformer trained with flow matching, a method that learns a path from noise to an image representation. SmolLM3-3B encodes the text, while the Wan 2.2 variational autoencoder converts between compressed image representations and pixels. Bria’s DimFusion conditioning system supports training with the unusually long structured captions.

Bria reports that this architecture narrows the prompt-alignment gap with proprietary systems such as Gemini 2.5 Flash Image despite its 8-billion-parameter size. Artificial Analysis provides separate evidence for the deployed model’s competitive editing and generation performance, although direct results can differ between the hosted service and downloaded checkpoints.

Bria says FIBO was trained exclusively on licensed data, which may help teams with procurement rules that restrict scraped training corpora. Commercial users still need to review the license scope, data documentation, output terms, and any indemnity provided for their deployment.

Three ways to run it

1. Hosted inference: Use the Bria API or fal when infrastructure management and predictable per-image pricing matter. Confirm current rates, limits, resolutions, and commercial terms with the selected provider.
2. Local research: Download the checkpoints from Hugging Face and follow the applicable noncommercial license. Full-precision memory requirements depend on the pipeline, resolution, and offloading configuration.
3. Apple silicon: Use the MLX build for a 4-bit Mac deployment. Quantization reduces memory use and can change output quality.

The Python example below loads the text-to-image checkpoint with Diffusers, assumes a CUDA device with bfloat16 support and sufficient memory, and fixes the seed for repeatable testing.

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "briaai/FIBO",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)

image = pipe(
    "a houseplant in a white pot, soft morning light",
    generator=generator,
).images[0]

image.save("fibo-output.png")

The separate editing checkpoint requires a source image and its corresponding editing inputs. Structured JSON users should follow the current model card instead of inventing field names, since the accepted schema and pipeline interface may change between revisions.

Pick it for control and provenance

Production evaluations should compare the hosted API and intended local checkpoint on representative prompts, then lock the model revision, prompt JSON, seed, scheduler, dependency versions, and inference settings. That process measures the actual deployment path and preserves reproducibility across releases.