Flux.3 — Klein

Unlike Flux.2 Klein, Flux.3 Klein comes in a single size: 5.6 billion parameters. The choice lies elsewhere, between two forms of the same model. The distilled variant produces a finished image in 4 to 16 steps, without CFG. The base variant needs 50 steps at CFG 3.5 and is meant for fine-tuning and LoRA training. For image work, the distilled one is the right choice.

The text encoder is Qwen3-VL-4B — a language model that also understands images. That's more than a technical detail: a language model reads a prompt the way a person reads it, and that matters when prompting. The VAE is the same as in Flux.2. If you've set up Flux.2 Klein, it's already in your folder.

Besides text-to-image, the model can edit existing images, combine up to ten reference images and place image elements via bounding boxes. In its own comparisons, Black Forest Labs ranks Flux.3 Klein above Flux.2 Klein 9B. Those are the manufacturer's figures — my own comparisons follow below.

The measurements — all at 1024×1024, warm (models in memory), 16 steps:

There's no negative prompt. Anything unwanted is rephrased positively, as with Flux.2 Klein: “an empty beach” instead of “no people”. What's new lies elsewhere.

A prompt isn't parsed. It goes through a language model as a whole, and that model reads structure the way a person does. That's why Flux.3 Klein — like Flux.2 Klein before it — understands a freely structured JSON prompt directly: camera, light, elements, interactions, each in its own field. The language model forgives syntax errors and interprets custom field names correctly.

The advantage shows when checking the result. Every field becomes an item you can tick off against the image. And whatever gets its own field ends up in the image: the two interactions in the test prompt — the dune grass overlapping the edge of the deck, the sky reflected in the glass door — each had their own field, and both were rendered.

What JSON alone can't do is composition. “Sky across two thirds of the image, house low in the frame” wasn't rendered by any model, neither Flux.2 Klein nor Flux.3 Klein, not even with the highest priority in the field. A ranking raises importance; it doesn't set coordinates.

This is where Flux.3 Klein comes in. Each image element gets its own box with coordinates and its own description. The coordinates are relative, from 0 to 1000, and so apply to any image size — only the aspect ratio has to be fixed before you lay out the boxes, because the grid stretches with the format.

Black Forest Labs documents a fixed format for this: running text with markers, followed by a list of boxes. But it also works directly in JSON. The boxes simply sit as a bbox field on the elements:

Coordinates can be calculated by hand: pixels from the top divided by image height times 1000, pixels from the left divided by image width times 1000. It's easier in ComfyUI itself. In the Create Bounding Boxes node you draw the boxes on a canvas in the format of the output image, and FLUX.3 Layout Prompt appends them to the prompt in the right format. The boxes are then called obj_1, obj_2 and so on, and the axis-order error can no longer happen.

Flux.3 Klein can edit existing images by instruction and combine up to ten reference images, addressed in the prompt as <ref_image_0>, <ref_image_1> and so on. Black Forest Labs promises that the rest of the image is preserved during an edit. Their own examples show what that means: depending on the edit, between 68 and 94 percent of the pixels remain identical — not all of them.

For product images, series with recurring characters and objects that have to stay exact, this is the most exciting part of the model. My own tests will follow.

Black Forest Labs' documentation describes the API model. Some of it isn't available locally: via the API, the model can search the web on request to depict real things correctly; it offers fixed aspect ratios and resolution tiers up to 4K; it expands short prompts automatically; and after an edit it additionally restores untouched areas. Locally you set width and height directly, the upsampler is a separate model, and web search is missing entirely. Worth knowing if you rebuild examples from the BFL documentation locally.

Flux.3 Klein comes under a new license, the FLUX Kommunity License. It sits between Apache 2.0 and the previous non-commercial license. If you, together with affiliated companies, have less than 5 million US dollars in annual revenue, you may use the generated images commercially and run the model in production for that purpose, without a license fee. Above that, you need a license from Black Forest Labs. For freelancers and small studios, that's good news. For the marketing department of a large company, it means: license before going into production.

The license also requires generated images to be filtered or reviewed before distribution, and to be labeled as AI-generated where the law requires it. Circumventing the model's safety mechanisms is expressly prohibited. The text encoder is separately licensed under Apache 2.0.

Compared with Flux.2 Klein, one advantage is lost: Flux.2 Klein 4B is under Apache 2.0 and free to use without a revenue cap. More on this in 2.1.4 — Understanding license models. What governs is the current version on bfl.ai; guidance, not legal advice.