Black Forest Labs: FLUX.2 Pro

FLUX.2 Pro is an image generation and editing model from Black Forest Labs, released on 25 November 2025. It sits in the provider's high-end tier, aimed at frontier-level visual quality and reliability rather than at lightweight or high-volume draft work. If you already know the FLUX line, this is the Pro variant: the one Black Forest Labs positions for output where prompt adherence and visual consistency matter more than raw throughput.

The model accepts text and image inputs and returns images. Its context window is 46,864 tokens, with a maximum output of 42,177 tokens. Practical generation controls include aspect ratio, with the enum values 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9 and auto, and output format, either png or jpeg. You can request exactly one image per call (n has a minimum of 1 and a maximum of 1), and you can supply up to 8 input references. A seed parameter is supported. According to Black Forest Labs, the model is built for strong prompt adherence, stable lighting, sharp textures, and consistent character and style reproduction across multi-reference inputs.

Those capabilities point to a fairly specific set of jobs. Multi-reference input with consistent character and style reproduction suits iterative design work, where you feed earlier renders back in and want the subject to stay recognisable across variations. Stable lighting and sharp textures matter for product and marketing imagery, where a single odd shadow or soft edge means a re-run. Editing workflows benefit from the combination of image input and text instructions, since you can describe the change rather than rebuild the scene. The aspect ratio enum covers the common delivery shapes, from square social crops through 16:9 and 21:9 widescreen, so you can generate close to final framing instead of cropping afterwards.

You call it through api14's OpenAI-compatible API using the model id black-forest-labs/flux.2-pro. Because the interface follows the OpenAI schema, existing clients and SDKs work with a change of model id and base URL. The request parameters the model supports are seed, plus the image parameters aspect_ratio, output_format, n and input_references. Set aspect_ratio to one of the listed enum values, output_format to png or jpeg, and n to 1. Use input_references to attach between 0 and 8 reference images when you want the output conditioned on existing material, and pass a seed when you need a reproducible result.

Two constraints are worth planning around. Only one image comes back per request, so batches mean multiple calls rather than a single wide response. And the supported parameter list is short: seed is the only general sampling parameter exposed, with the rest of the control surface living in the image parameters. If your task is text rather than images, or you need a model that returns text alongside generated media, this is not the right endpoint. For image work where reference consistency and finish quality are the priority, it fits.