Fibo cuts image generation from 50 steps to 6

Fibo Generate 1.5 reduces the number of generation steps from 50 to 6, retains the existing API, and produces each image for $0.04, according to Bria.

An existing integration can switch to Fibo Generate 1.5 without any code changes. Bria is retaining the same endpoint, inputs, and response format, while replacing the model running behind the scenes with a distilled version that requires significantly less compute.

The main change concerns the number of steps needed to produce an image. The first version of FIBO used 50 steps, with a guidance mechanism that evaluated two paths at each stage. Bria therefore counted 100 model evaluations for a complete generation.

Fibo 1.5 operates in four to six steps and removes this guidance mechanism. With the recommended six-step configuration, the process requires only six model evaluations. This reduction is more substantial than the comparison between 50 and 6 alone suggests, but it does not translate directly into an equivalent end-to-end speed increase for users.

The total wait time also includes prompt expansion, queuing, model loading, safety checks, and delivery of the result. The documentation further states that generating a 4MP image can add approximately 30 seconds to the process. Bria has not yet published a detailed comparison of the complete processing times for FIBO and Fibo 1.5 under different conditions.

To produce this version, the company applied an initial distillation phase called DMD, designed to reduce the number of steps. A second phase, DMD-R, then refines the distilled model to improve realism. The goal is to avoid the loss of quality commonly associated with accelerated models.

Bria claims that realism and material rendering improve despite the reduced step count. The Fibo 1.5 model card mentions a blind human evaluation covering 247 pairwise comparisons. However, it does not specify the number of participants, the exact source of the prompts, or how the results were distributed across different image categories. No independent benchmark currently accompanies this comparison.

Fibo 1.5 does not introduce a new architecture. It retains the eight-billion-parameter diffusion model, SmolLM3-3B text encoder, DimFusion conditioning system, and Wan 2.2 decoder used by the original version. Its weight structure and control mechanisms also remain unchanged.

FIBO’s distinguishing feature is its use of structured JSON descriptions. A short request can be expanded into a document of more than 1,000 words describing the subjects, their positions, lighting, colors, composition, depth of field, and camera characteristics.

The “no prompt engineering” claim should therefore be understood from the user’s perspective. A user can begin with a short instruction, but the platform still converts it into a much more detailed description before generating the image.

Bria’s generation endpoint currently uses Gemini 2.5 Flash to perform this transformation. It returns both the image URL and the JSON description used to create it. Advanced users can also submit their own structured document and bypass this intermediate stage.

Three workflows remain available. Generate expands a short idea before creating an image. Refine takes an existing description and changes selected attributes based on a new instruction. Inspire analyzes a reference image to produce variations or combine it with another creative direction.

The API continues to accept a prompt, an image, a combination of both, or a structured description. It also retains its options for output format, aspect ratio, generation seed, moderation, and synchronous or asynchronous responses. Developers therefore do not need to change their request bodies or the way their applications retrieve results.

This technical compatibility does not mean that images will remain identical after the update. Changes to the model weights and sampling method may affect the style, details, or interpretation of the same request. Companies that rely on previously validated results should test their usual scenarios before replacing the earlier version in production.

Some historical parameters also lose their usefulness when the model is run locally. The model card states that negative prompts have no effect on Fibo 1.5 because the associated guidance mechanism has been removed. They can be omitted, even if an interface continues to display the field for compatibility reasons.

The distilled checkpoint is not intended for fine-tuning either. Bria recommends starting from the original FIBO model when developing a specialized version, then considering acceleration as a separate stage. The fine-tuning code released for the first version does not directly apply to Fibo 1.5.

Bria has announced a price of $0.04 per image. The announcement does not explain whether rates vary between the available output definitions, whether volume discounts apply, or which contractual terms cover that price. Third-party platforms distributing the model may also use their own pricing structures.

Fibo 1.5 is available through Bria’s sandbox, its API, fal.ai, ComfyUI components, and a checkpoint hosted on Hugging Face. The model card lists Fibo 1.5 as early as August 24, 2026, while Bria’s September 2 announcement marks its availability through the company’s existing commercial infrastructure.

Access to the checkpoint requires users to accept a license limiting free use to personal, academic, and nonprofit projects. Commercial applications must go through Bria or be covered by a separate agreement. Despite the “open source” wording used on the model card, this restriction does not meet the strict definition of software that can be freely used for any purpose.

Like the first version, Fibo 1.5 was reportedly trained exclusively on licensed material selected from a collection of approximately one billion images. Bria says these assets were reviewed for commercial rights, traceability, and regional compliance. These statements come from the company and are not accompanied by an external audit in the launch materials.

Fibo 1.5 is ultimately less a new generation of FIBO than an adaptation designed for production. It retains