FLUX.2 [dev] lands on Workers AI

Black Forest Labs has released FLUX.2 [dev], an open-weight image generation model, and it is now available on Cloudflare's Workers AI inference platform. The release follows recent advances in closed-source image generation from Google and OpenAI, giving developers an open alternative with comparable capabilities. Cloudflare has hosted FLUX.1 [schnell] since its early versions, making it one of the most popular models in the catalog, and now brings the licensed successor to the same infrastructure.

The Workers AI deployment of FLUX.2 [dev] supports multipart form data for input images — up to four 512x512 images — and can output images up to 4 megapixels. This format lets users send multiple image references alongside standard model parameters in a single request.

What FLUX.2 brings to the table

FLUX.2 [dev] demonstrates a stronger grasp of the physical world compared to its predecessor. It renders accurate hands, faces, fabrics, logos, and small objects — details that frequently trip up other models. The model also produces realistic lighting, shadows, reflections, and depth perception, which makes it suitable for use cases demanding high visual fidelity like creative photography, e-commerce product shots, marketing visuals, and interior design.

Beyond photorealistic output, FLUX.2 handles digital asset creation such as landing page layouts and infographics. It also understands multiple languages natively, so a prompt written in French can yield a polished French-language web page design without additional translation steps.

Keeping characters consistent

Generative image models typically struggle with "stochastic drift" — the tendency for an identity, product, or style to subtly change between generations. FLUX.2 addresses this with multi-reference image inputs, letting developers supply up to four reference images to maintain consistency across outputs. This enables workflows where a background, lighting, or pose changes while the subject's face or product design stays true to the source.

Real-world applications of this capability include:

  • Ad variations: Generating numerous advertisements with the same actor without facial drift between frames.
  • Product shots: Placing a product in different scenes — a beach, a city street, a studio table — while keeping the product itself pixel-accurate.
  • Editorial spreads: Producing fashion shoots where the model appears identical across every angle and shot.

Implementation details

In code, multi-reference image inputs are handled via multipart form-data uploads. The input images are binary data, and the output is a base64-encoded image. The same functionality is available through the Workers AI Binding interface.

Fine-grained control with JSON prompting

FLUX.2 extends controllability beyond text prompts. Developers can pass structured JSON as part of the multipart form input to specify layout, composition, and style details with precision. The model respects these structured instructions, following them exactly as provided.

The model also understands precise color specifications through hex codes. For example, asking the model to recolor accent lighting to a specific orange can be done by including the hex value #F48120 in the prompt. This level of control is particularly valuable for brand-accurate work where exact color matching matters.

FLUX.2 [dev] is available now on Workers AI. Developers can access the model through the developer docs or experiment with it in the multimodal playground. Further details on the model's capabilities are available in Black Forest Labs' launch post.