Bria FIBO 2: Image Controls and Commercial Licensing

Bria's FIBO 2 combines structured image controls with two distinct access paths: non-commercial public weights and a separate commercial offer. Choosing the model means assessing both the editing workflow and the terms of the intended use. The weights went up on Hugging Face on October 18 and 19, 2026 under Creative Commons Attribution-NonCommercial 4.0, with the model cards stating that commercial use is subject to a commercial agreement with Bria. The standard commercial self-hosted plans meter calls per model; Enterprise has a separately negotiated flat-price offer, and hosted API access is another path.

That split is the subject of this post, because it is the decision a creative-agent builder actually faces. The model's controls are unusual and worth understanding: it renders from a structured description Bria calls Visual Generation Language, so an agent can change one attribute of a scene and re-render with the same seed rather than rewriting a prompt and hoping. But the controls only matter if the deployment path is open, and the deployment path runs through a licence page, a set of API keys, a usage ledger and an attribution agent that has to run beside the model.

We covered the other open image model of the month, Black Forest Labs' FLUX 3 Klein, in its own guide; section 07 compares the two on operational terms rather than repeating that overview. Section 05 is the heart of this one.

Bria did not publish a dated launch post we could find, so the repository dates are reported as availability evidence rather than a confirmed announcement date. The FIBO 2.0 Dev repository was created on Hugging Face on October 18, 2026 and the Edit Turbo repository on October 19; both were last modified on October 25 and both are gated behind a form. Bria's documentation site carries a page headed with the words that FIBO 2 is live, and its licensing page for the model, read on October 26, offers FIBO 2 for self-hosting. Those timestamps identify the source records, without establishing a single launch day.

The cards describe an 8B-parameter flow-matching model working in the latent space of the FLUX.2 variational autoencoder, with prompts encoded by a 4B vision-language model and a 44-block diffusion transformer rendering the result. Both checkpoints total about 8.1 billion parameters in BF16 safetensors. Dev is the full, non-distilled base; Edit Turbo is distilled from it for few-step editing. Both cards cite rankings from Artificial Analysis dated October 20, first among open-weight editing models and second in text-to-image; those are Bria's citations of a third-party arena, and we did not read the leaderboard.

A detail that shows the product is still moving: the Dev card lists five photographic styles, while the API reference for the hosted generate endpoint lists four, and the self-hosting guide describes an installable FIBO 2 product, while a separately named FIBO 2 Turbo commercial selector is marked as coming soon. That selector does not establish the availability of the distinct Edit Turbo weight checkpoint. A buyer should expect the documentation to keep changing for a few weeks and read the page on the day, not the summary from the week before.

Bria's documentation describes the usual prompt-to-image loop as a black box: the prompt goes in, an image comes out, and nobody can see how the prompt was interpreted. Its answer is a JSON document that describes the scene explicitly, which it calls Visual Generation Language and which the FIBO models are trained to render from directly. In practice you still send a prompt or a reference image; a translation step turns that into the JSON, the image is rendered, and the response returns both. From there you can store the specification, change one field and re-render, or hand it to an agent.

Why that matters for an agent: the specification is addressable. Bria's schema has required fields for a short description, the background setting, the medium and lighting and aesthetics objects, and optional arrays for objects with normalised bounding boxes and for text to render with its own placement and colour. A person in the scene is an object with clothing, pose keypoints and expression fields. Bria's own example is that changing a shirt colour becomes a data edit rather than a masking exercise. An agent that can validate and patch JSON can therefore produce a controlled variation without a round of prompt engineering.

Two cautions, both from Bria's own pages. The claim that the same specification and seed give the same output is Bria's, and it holds for one model version on one deployment; the documentation notes a coming schema change to colour and clothing fields, so a stored specification is reproducible only while the schema it was written against is served. And a free-form prompt or edit instruction is expanded into the JSON by a separate language model. In the hosted API that is invisible; when self-hosting, you supply that model on your own account, and a free-form request without one simply stops.

The two repositories share an architecture and a licence and little else in intent. FIBO 2.0 Dev is the model you generate with and fine-tune: the card recommends 50 inference steps with a guidance scale of 5.0, supports text-to-image, instruction editing, masked editing where white marks the region to regenerate, multi-reference editing with up to four images, and a mode that generates from a reference plus a short prompt without copying the reference. The card links a single-GPU LoRA trainer and describes Dev as the base for adapters and full fine-tunes.

Edit Turbo is the production editor. It is distilled from Dev for four guidance-free steps, so the recommended settings are four steps at a guidance scale of 1.0, and three things stop working as a result: it does not generate from text at all, negative prompts have no effect, and the card says it is not intended for fine-tuning. What it keeps is the editing surface, including masks and four references. A team that edits product photography at volume wants Turbo; a team building a house style through adapters wants Dev and should expect to run Dev for the edits the adapter changes.

The failure case is choosing by name. Turbo sounds like the fast version of the same thing, and for generation it is not a version at all. A pipeline that sends a text prompt to Edit Turbo gets nothing useful back; a pipeline that fine-tunes Turbo is working against the card's advice. Both cards also list features marked as coming soon, such as a multi-turn editing adapter, background removal and image expansion, which are not available as of October 26 and should not be in a plan yet.

The licence deserves a careful reading because it has been misread in both directions. The rendered cards state that the model is released under a CC BY-NC 4.0 licence for non-commercial use and that commercial use is subject to a commercial agreement with Bria. The access gate repeats that the weights are open for non-commercial use only. The Hugging Face metadata, meanwhile, labels the licence as a custom Bria licence and links to the Creative Commons deed. The metadata label does not expand the permission stated in the cards. The Creative Commons deed summarises the legal code; the public licence and any separate Bria agreement govern the relevant use.

What non-commercial excludes is the ordinary business of an agency or a product team: generating images for a client, editing product photography for a store, rendering creative for a campaign. None of that is permitted under the weight licence alone. The public licence permits non-commercial uses subject to its conditions. Research, evaluation or prototyping labels do not automatically make a commercially motivated activity non-commercial. Assess the intended use before downloading or testing, and use an appropriate Bria commercial agreement where needed.

For the position of these terms against the field, our open-weight licence audit tracks the full range, from Apache 2.0 to revenue-capped and non-commercial grants. FIBO 2 sits at the restrictive end for the weights and resolves it with a commercial product, which the next section describes.

Bria's standard self-serve licence is a subscription in your billing account, one per model, that unlocks an installation path and a meter. Enterprise has separate terms, including advertised offline deployment and unlimited calls at a custom annual price. The documentation describes three steps. Buy the licence on the pricing page. Install the model's package, either through Bria's own command-line installer or manually, which exchanges one of your ordinary API keys for a short-lived registry credential; the weights come from Bria's own distribution, and the docs point out that no Hugging Face account is needed. Then load the model with an API key and run it, with every generation counted against the licence and shown in the console tagged as self-hosted.

The consequences of that design are operational. A cancelled or lapsed licence refuses new starts, although a running process continues until it restarts, and a card payment that fails but is being retried keeps the licence active. The licence and the hosted API plan are separate: buying one does not block the other, and an account with no API plan moves to pay-as-you-go. The self-hosting guide says inputs and outputs stay on your machine and only a small attribution vector is sent to Bria. The attribution page explains what that is: an irreversible embedding of each output, matched against licensed training data to pay data partners, and running the attribution agent is a condition of the commercial licence and of indemnity cover.

The hardware requirement, from the FIBO 2 self-hosting guide, is an NVIDIA GPU with 40 GB of memory, Python 3.11 or 3.12 and about 26 GB of disk for the weights, plus a language model of your own for free-form prompts. The failure case is a team that reads self-hosted as offline: the model needs network access to Bria's engine for licence checks, the weights CDN and attribution. The documentation reserves air-gapped and on-premises deployment, full precision and custom models for the enterprise path.

Bria prices the licence by call, so the question for a budget is what a call costs at each tier, before the GPU. Our arithmetic on the October 26 list prices, with the plan fully used: Starter at $499 for 25,000 calls is about two cents a call; Scale at $1,499 for 125,000 is about 1.2 cents; and any call above a plan is four cents. The Free tier is zero for the first 2,000 calls and four cents after, which the page says is billed on the month's invoice rather than blocked.

Three hypothetical workloads, our illustrations rather than any customer's bill. A studio that renders 1,500 product variants a month under $2M revenue pays no licence fee within the eligible Free allowance; its compute, prompt-model and operating costs remain separate. A small retailer's catalogue team at 20,000 edits a month pays $499 on Starter, about 2.5 cents each. An agency under $10M revenue at 200,000 images a month is above Scale's allowance by 75,000 calls: $1,499 plus $3,000 in overage, about 2.25 cents each. The page bills that overage rather than blocking it, and points larger companies and those wanting indemnity, an SLA or a custom model to Enterprise. Every one of those figures excludes the 40 GB GPU, the language model for free-form prompts and the engineer running it.

The failure case is the revenue threshold. Free is for companies under $2M in annual revenue and Starter and Scale for companies under $10M, confirmed at checkout. A team inside a larger company cannot self-serve at all; its only route is the enterprise contract with a flat annual price. That is not a surprise Bria hides, but it is one that a proof of concept on the Free tier will not surface until procurement asks.

The obvious comparison is Black Forest Labs' FLUX 3 Klein, which we covered on October 16 and whose licence permits free commercial use for organisations under $5 million in annual revenue, with larger firms licensing or using the API. The two models answer the same question differently. FLUX's threshold is on the weight licence itself: below it, download and deploy with no meter; above it, negotiate. FIBO's weight licence has no commercial grant at any revenue; the threshold lives in the self-hosted product and sets which self-serve tier you may use. Enterprise is advertised separately with unlimited calls and a custom annual price.

The practical consequences differ. A sub-$5M company shipping FLUX 3 Klein has no monthly fee and no call count to report. For FIBO 2, a company below $2M may qualify for the 2,000-call Free allowance; a company between $2M and $5M instead needs an eligible paid plan or an agreement. Standard self-hosted plans also require the key-based licensing and attribution workflow. Bria advertises licensed training data, IP warranties on outputs for the self-serve tiers and uncapped indemnity for Enterprise. These are Bria's stated offers, not a finding that competing providers offer no contractual protections. For a brand whose legal team asks where the training data came from, that is the product.

Our product-content comparison of Nano Banana and FLUX 3 covers output quality questions we are not qualified to settle here without running the models. For an eligible small company, the operational choice is between the revenue-limited grant described in our FLUX guide and Bria's standard keyed, metered plans. Larger or offline deployments need the applicable commercial agreement. Check both licences for the intended use.

FIBO 2 fits a team for whom two things are true: the structured specification is worth engineering around, and the provenance story is worth paying for. The first is true of agents that produce many controlled variations of a scene, such as requesting twelve colourways of a product. Editing one JSON field makes the requested change explicit, but each rendered image still needs checking for unintended changes to lighting, geometry or other details. The second is true of brands and publishers whose counsel asks about training data, because Bria's licensed-data claim and contractual protections can inform that review. Counsel still needs to assess the actual warranty or indemnity wording for the intended use.

It does not fit a team that wants offline weights with no vendor relationship, because the standard self-serve product contacts Bria for licensing and attribution. Offline deployment is advertised under a separate Enterprise agreement. It does not fit a company above $10M revenue that wants to self-serve, because only the enterprise contract is open to it. And it does not yet fit a pipeline that needs the coming-soon features, such as image expansion or background removal on the self-hosted model, which are listed but not available as of October 26.

Where it does fit, the deliverable changes. A pipeline using Bria's structured workflow can retain both the image and its scene specification. Keep, version and review that specification alongside the image; do not assume a bare checkpoint automatically returns it. That is the same shift we see with editable vector output from Ideogram and with content credentials generally: the asset is the structured record, and the pixels are a render of it. Our content-credentials validator guide covers checking the provenance marks that Bria says it attaches to every hosted output.