Black Forest Labs: FLUX 3 Image

FLUX.3 Image is an image generation and editing model from Black Forest Labs, released on 1 October 2026. It is the company's flagship model in that line, and on api14 it sits in the image category. If you have used earlier FLUX releases, the general shape of the workflow will be familiar: you send a prompt, optionally with reference images, and you get an image back.

The model accepts text and image input and produces image output only. It supports both text-to-image generation and multi-reference editing, where you can supply up to 10 input reference images alongside your prompt. The context window is 46,864 tokens, with a maximum output of 42,177. Output images are rendered at fixed resolution tiers rather than arbitrary sizes: 768, 1K, 1.5K, 2K or 4K. You also choose an aspect ratio from a fixed list covering 21:9 through to 9:21, plus an auto option, and the model returns a single image per request.

Those capabilities point to a few practical patterns. Text-to-image at a chosen resolution and aspect ratio covers straightforward asset generation, such as hero images, thumbnails or social crops, where you want a specific shape rather than whatever the model defaults to. The reference-image path is the more interesting one for editing work: passing several inputs lets you steer style, subject or composition, which suits tasks like restyling a product shot, keeping a character consistent across a set, or combining elements from more than one source image. The 4K tier is there when the output needs to hold up at larger sizes, while 768 is a reasonable choice for drafts or previews where you plan to iterate.

Calling it through api14 works the same way as any other model on the gateway, since the API is OpenAI-compatible. You use the model id "black-forest-labs/flux-3-image" in your request. The supported parameter is seed, which is useful when you want to reproduce or vary a result. Image-specific options are passed as image parameters: resolution, aspect_ratio, n (which accepts a minimum of 1 and a maximum of 1, so one image per call), and input_references, which takes between 0 and 10 images. If you are already sending OpenAI-style requests, you mainly need to add those image fields.

One thing to plan around is that n is capped at 1, so generating several variations means several calls, and the seed parameter is the practical way to keep those variations related. Resolution and aspect ratio are enums, not free values, so you pick from the listed tiers and ratios rather than specifying exact pixel dimensions. If your task is text generation, classification or anything that returns text rather than an image, a text model on api14 will fit better; this one is built for image output.