A new generation of background removal at Bria targets challenging images
The RMBG-3.0 model adapts Fibo 2.0 for background removal. Bria reports gains on complex images, with commercial use subject to licensing.
The third generation of Bria’s background removal model takes a different approach. Built on Fibo 2.0, RMBG-3.0 uses a generative foundation to produce a transparency mask from a photograph. This mask identifies what should be kept, removed, or made partially transparent, then is applied to the original image. The resulting cutout can be placed in another composition.
The approach targets situations that complicate visual production: multiple people in a scene, fine hair, movement, or objects with openwork. According to the model card, mask generation takes four steps, following fine-tuning and distillation of Fibo 2.0. The company says it used exclusively licensed training data. Preserving detail while transparency remains a challenge
For a product catalog, the challenge extends beyond the outer contour. Background removal must preserve packaging text, avoid making an opaque bottle partially transparent, and remove the background between chair rungs or bicycle spokes. The model’s presentation emphasizes these cases, along with separating a product from its cast shadow and reflections.
The demonstrations also cover portraits, groups, and illustrations. On fal, the service returns a transparent PNG at the same dimensions as the input image. The mask supports intermediate transparency levels, useful for hair, fur, and motion blur, rather than being limited to a hard-edged cutout.
The model card acknowledges persistent difficulties with glass, veils, smoke, and splashes, which can become opaque. In cluttered or low-contrast scenes, grass or nearby props may remain attached to the subject. Some very fine structures are also handled better by Photoroom in Bria’s published comparison. An internal ranking based on 234 images
The claimed first-place finish comes from a vendor-run benchmark of 234 deliberately challenging images. Competitors include Photoroom, remove.bg, Ideogram, BiRefNet, and RMBG-2.0. Each system received the full-resolution images, using its default settings and producing one output per image. The evaluation combines blind voting by professional and nonprofessional users with comparisons performed by a vision-language model.
Under this protocol, RMBG-3.0 produces the best result or ties for first place on 185 images, approximately 79% of the total. That percentage describes its relative position among the systems tested, rather than a rate of perfect cutouts. The model also ranks last on five images, with failures including an empty mask or removal of the main subject. These remain results from Bria’s internal benchmark. Heavier deployment requirements and separate usage rights
Moving to a generative foundation increases hardware requirements. For local execution, Bria lists an eight-billion-parameter model and recommends at least 32 GB of GPU memory. The reported processing time is approximately six seconds per image on an Nvidia L40S, using the published reference configuration. The RMBG-2.0 version is still presented as better suited to low-latency requirements and smaller graphics cards.
Hosted access removes the need to manage that infrastructure. On fal, the listed price is $0.018 per image, or $1.80 for 100 cutouts. The model is also available through Bria’s API, while its weights are accessible on Hugging Face.
The “open source” label used in the announcement needs qualification: the model card presents these weights under the CC BY-NC 4.0 license, which restricts use to noncommercial purposes. Commercial use requires a separate agreement with Bria or access through its services and the fal offering licensed for that purpose. Licensed training data and permission to use the model are therefore two separate matters.