FLUX models: sharp text-to-image results and legible lettering

The short answer: FLUX is the family of models from the European AI company Black Forest Labs (BFL) that generate images from text and edit images. The team is made up of researchers who built models such as Stable Diffusion and Latent Diffusion. FLUX's known strengths are prompt adherence, photorealism and the ability to put legible text inside images. The family includes both models offered only through an API and open-weight models you can download and run on your own computer, and the two have different licences. This guide covers the versions, what each one is for, the licence differences and how to write a good FLUX prompt.

What is FLUX, and who makes it?

Black Forest Labs announced its launch and its first model family, FLUX.1, on the same day: 1 August 2024. According to the company's introduction, the team is behind work such as VQGAN, Latent Diffusion, Stable Diffusion, Stable Diffusion XL and Stable Video Diffusion, as well as the rectified flow transformer research. The company calls its approach of offering open models alongside paid product models “open core”.

Technically, FLUX.1 has a hybrid architecture of multimodal and parallel diffusion transformer blocks scaled to 12 billion parameters, and it is trained with a method called flow matching. BFL describes this method as a general, conceptually simple approach that includes diffusion as a special case.

The FLUX family: which version is for what?

FLUX.1: pro, dev and schnell

The first family came out with three models. FLUX.1 [pro] was the highest-quality version and was offered through the API. FLUX.1 [dev] was distilled from pro, with open weights but licensed for non-commercial use. FLUX.1 [schnell] (German for “fast”) was the fastest version, designed for local development and personal use, and openly available under the Apache 2.0 licence. According to BFL, all three supported a range of aspect ratios between 0.1 and 2.0 megapixels. Later came FLUX1.1 [pro], its Ultra mode with output of up to 4 megapixels, and tools such as FLUX.1 Fill for inpainting and extending images.

FLUX.1 Kontext: editing with words

Announced on 29 May 2025, Kontext brought FLUX into editing. The model takes both text and images and lets you change an image with a simple instruction. The core capabilities BFL lists are character consistency across scenes, local editing that changes only part of an image, style reference from an image, and low latency. Kontext [pro], and Kontext [max] with stronger prompt adherence and typography, were offered through the API, and an open-weight 12-billion-parameter Kontext [dev] was announced too. At the time of writing, BFL recommends FLUX.2 rather than Kontext for new projects.

FLUX.2: multi-reference and 4 megapixels

Released on 25 November 2025, FLUX.2 focused on real production work. According to the announcement, it can use up to 10 reference images at once, edit at up to 4 megapixels, produce complex typography, infographics and UI mock-ups with legible text, and follow multi-part, structured instructions better. The family consists of these members:

• [pro]: a balance of quality and speed at production scale.
• [flex]: exposes settings such as the number of steps and the guidance scale; strong at text and fine detail.
• [max]: the highest quality; it can search the web when needed to generate images based on current information.
• [dev]: a 32-billion-parameter open-weight model; commercial use needs a separate licence.
• [klein]: small, very fast models released in January 2026; the 4-billion-parameter version can run on consumer graphics cards with about 13 GB of video memory.

FLUX 3: image, video and sound in one model

Announced on 23 July 2026, FLUX 3 is BFL's multimodal foundation model that learns from images, video and audio together. The video side opened in early access; according to BFL it can generate up to 20 seconds of video with sound in a single generation. On the image side, according to BFL's documentation, FLUX 3 Image offers up to 10 reference images, output up to 4K and bounding boxes for deciding where elements go. On 15 October 2026 BFL released FLUX 3 Image Klein, an open-weight model with 5.6 billion parameters.

What makes FLUX stand out?

• Prompt adherence: when introducing FLUX.1, BFL highlighted visual quality, prompt following, size and aspect flexibility, typography and output diversity. With FLUX.2, adherence to multi-part instructions with compositional constraints was emphasised further.
• Text in images: work that involves text, such as posters, menus, infographics and UI mock-ups, is among the areas the FLUX.2 announcement describes as reliable in production.
• Exact colour control: the FLUX.2 documentation recommends giving brand colours as hex codes; for example, you can describe a vase as “a gradient starting with #02eb3c and finishing with #edfa3c”.
• Multi-reference: FLUX.2 and FLUX 3 Image can use up to 10 images together, which makes it easier to produce series with the same product or character.
• An open-weight ecosystem: open models can run on your own hardware and be fine-tuned to your style with methods such as LoRA.

A note: most comparisons BFL publishes are the company's own evaluations. The best test for your work is trying the same prompt on a few models.

Licences: what does open weights mean, and can I use it commercially?

“Open weights” means the model's trained parameters can be downloaded; it does not mean every use is allowed. In the FLUX family the licence depends on the model:

Model	How to access it	Licence
FLUX.1 [schnell]	Open weights	Apache 2.0
FLUX.1 [dev]	Open weights	Non-commercial use; separate licence for commercial use
FLUX.2 [dev]	Open weights	Non-commercial use; separate licence for commercial use
FLUX.2 [klein] 4B	Open weights and API	Apache 2.0
FLUX.2 [klein] 9B	Open weights and API	FLUX Non-Commercial License
FLUX 3 Image Klein	Open weights and API	FLUX Kommunity License
Pro, max, flex and FLUX 3 Image	API	API service terms

The FLUX 3 Image Klein licence allows free commercial use for individuals and companies with annual revenue under 5 million dollars, alongside research and non-commercial use; larger organisations have self-hosted licensing or the API. If you plan to use an open-weight model in a commercial project, read the licence text yourself; this guide is not legal advice.

How do you write a prompt for FLUX?

BFL's prompting guide suggests this order as a starting skeleton: subject, location, style, camera settings, lighting, colours, effect, additional elements. It isn't a strict formula; you can skip the parts that don't apply. The guide's other recommendations:

• Write in natural language: explain relationships in sentences: who holds what, what stands behind what, where the light comes from.
• Put text in quotation marks: quote the exact words that should appear in the image, and say where they go and how they look.
• Name the camera instead of writing “professional photo”: give the camera, lens or film stock and describe its effect on the image: “visible grain”, “a warm colour cast”.
• Turn negatives into positives: FLUX 3 Image has no negative prompt field. Write “an empty promenade” instead of “no crowds”.
• Refer to references by position: when you provide several images, call them “image 1” and “image 2” and say what each one is for.
• Change one detail at a time: start with a simple version, see what came out right and wrong, then adjust a single detail.

For a poster with text:

A vertical film poster showing a narrow city street with wet pavement after rain and neon signs. At the top, large bold white lettering reads “MIDNIGHT”; below it, in small letters, “Coming soon to cinemas”. Cinematic lighting, blue and purple tones, subtle film grain.

A product shot in a brand colour:

A matte ceramic water bottle standing on a light grey studio floor; the bottle's colour is #1F6F5C. Soft side light, a gentle shadow, 85mm lens, shallow depth of field.

Editing with references:

Place the backpack from image 1 on the back of the person walking along the forest trail in image 2. Keep the backpack's colour and logo exactly the same; match the light to the morning sun.

For automation and repeated work, the FLUX.2 documentation also shows writing the prompt in a structured form (JSON) split into fields such as subject, background, lighting, style, camera angle and composition. The example prompts in the documentation are in English; if a prompt in another language doesn't give you the result you want, try the same prompt in English and keep any text for the image verbatim in quotation marks.

Which FLUX model should I choose?

• Quick experiments and high volume: FLUX.2 [klein] or FLUX 3 Image Klein; both are very fast and have open weights.
• Running it on your own computer: the open-weight models. According to BFL, FLUX 3 Image Klein can run on graphics cards with 8 GB of video memory using its INT8 build.
• Posters, infographics, dense text: FLUX.2 [flex] or FLUX 3 Image; text and fine detail are where these versions stand out.
• Images based on current information: FLUX.2 [max], which works with web search, or FLUX 3 Image, which searches the web by default.
• A commercial product: the API models, or open models whose licence allows commercial use; check the licence table.

What are its limits?

• Reproducibility: FLUX 3 Image has no seed field, so the same prompt run twice can give different results. BFL recommends looking at more than one result before deciding a change didn't help.
• Time for high resolution: according to the documentation, 4K generation can take several minutes, and since changing the resolution counts as a new generation, the composition can change.
• Hardware: running open models locally needs a strong graphics card; large models such as FLUX.2 [dev] run on consumer hardware only through compressed versions.
• Licence restrictions: check the terms before using an image made with a non-commercially licensed model in commercial work.
• Where comparisons come from: most performance charts are BFL's own tests; verify them with your own examples.

Common mistakes

• Writing a keyword list: instead of “cat, window, sun, 4k, masterpiece”, describe the scene in one or two sentences. BFL's guide recommends writing details you could point to in the frame rather than filler words such as “masterpiece”.
• Giving text without quotation marks: if you don't quote the text for the image, the model may treat it as part of the description.
• A ratio that contradicts the prompt: if you choose a very wide frame but describe a narrow scene, the model fills the space itself; in a wide frame, say what goes on each side.
• Not reading the licence: “open weights” doesn't always mean “free for commercial use”.
• Generating drafts at the highest resolution: refine the prompt at a low resolution and take the final version at a large size.

Summary

• FLUX is Black Forest Labs' family of image models; the team includes the researchers behind Stable Diffusion.
• FLUX.1 came out in three versions (pro, dev, schnell); Kontext brought editing, and FLUX.2 brought multi-reference and 4 megapixels.
• FLUX 3 combines image, video and sound in one model; FLUX 3 Image offers 10 references and 4K output.
• Its strengths are prompt adherence, photorealism and legible text in images.
• Open-weight models have different licences: Apache 2.0, a non-commercial licence and a revenue-capped community licence.
• A good prompt uses natural language, quoted text, a named camera and light, and positive phrasing.