Black Forest Labs introduces FLUX.1 Kontext

What was announced

On 29 May 2025, Black Forest Labs introduced FLUX.1 Kontext, a model family built for in-context image generation and editing^[1]. The term describes how the model is used: instead of writing a prompt that describes an entire new picture, a user supplies an existing image together with an instruction — replace this object, keep that face, restyle this scene — and the model returns an edited image. The hard part is not producing a plausible picture but leaving everything the instruction did not mention untouched, so that the result still reads as the same photograph or the same character.

Three variants

Kontext came in three variants: Kontext [pro], Kontext [max], and an open-weight version, Kontext [dev]^[1]. The [dev] model is a 12B model^[1]. Parameter count is a rough measure of size — the number of learned numbers inside the network — and at 12 billion it is small enough for outsiders to run and adapt, not only for the company that trained it. The split between the closed [pro] and [max] versions and the open [dev] version determines who can use the model and how: closed variants are reached through an interface or a hosted product, while open weights are published so that others can download, run and build on them.

What it was built around

The suite was built around character consistency, local editing, style reference and interactive speed^[3]. Character consistency means the same person or character stays recognisably the same across successive edits and images. Local editing means a change is confined to the region the user points at. Style reference means one image can supply a look — colours, texture, rendering — that is then applied elsewhere. Interactive speed means the model answers quickly enough to be used in a back-and-forth loop rather than as a queued batch job.

Black Forest Labs claimed up to 8x faster inference than leading models such as GPT-Image^[1]. That figure is the company's own comparison, and speed numbers of this kind depend on hardware, resolution and how the measurement is set up; what it signals is that latency was treated as a design goal rather than an afterthought.

Launch, access and the Playground

Alongside the models, the company launched the BFL Playground at the same time, its own place to try them^[1]. Kontext [pro] and [max] were made available at KreaAI, Freepik, Lightricks, OpenArt and LeonardoAI, and through infrastructure partners FAL, Replicate, Runware, DataCrunch, TogetherAI and ComfyOrg^[3]. Those partners cover two different kinds of access: creative tools, where the model appears inside an editing product, and infrastructure providers, through which developers call the model directly. Kontext [dev], meanwhile, initially shipped in private beta^[3].

Open weights

Open weights for Kontext [dev] followed on 26 June 2025^[2]. Publishing weights changes the audience: besides people who call an interface, it includes researchers who want to inspect or fine-tune the model and developers who want to run it on their own hardware.

Why it matters

Image editing in practice is iterative: change something, look at the result, change it again. Whether a character survives an edit, whether a change stays where it was asked for, and whether the response arrives fast enough to keep working all decide whether such a model fits into a real workflow rather than only a demonstration. Kontext packaged one set of capabilities into three variants with different degrees of openness, so that hosted products, developers and open-weight users each got a workable form of the same thing^[1][3].