Black Forest Labs: FLUX.2 Klein 4B

FLUX.2 [klein] 4B is Black Forest Labs’ compact model for creating an image from a text description and editing images with visual references. Its four-billion-parameter rectified-flow transformer places generation and editing in one architecture. The named 4B edition is step-distilled, and the manufacturer’s reference example uses four inference steps.

The model’s role is an interactive visual draft or revision: explore a product scene, change a character’s clothing, combine a subject with a setting, or refine the appearance of an illustration. Begin with the event or composition that matters, then describe the subject, material, lighting and viewpoint. A concrete brief might place a matte blue ceramic cup on a pale oak table, with a side-lit close view that keeps its handle and glaze visible. The result is an image to inspect against those decisions.

Editing with a small reference set

The documented 4B profile accepts up to four input images. Assign a clear role to each: product appearance, character, setting or style. If one picture is the scene to edit, identify it and state both the change and the details to retain. For example, keep the cup’s silhouette and surface from the first image, use the room from the second, and replace only the table covering. Reference order and a precise change instruction give the revision a clearer target than asking for a general improvement.

The manufacturer distinguishes this compact reference budget from the larger FLUX.2 variants. It also states that [klein] has no prompt upsampling. Supply the important descriptive information yourself, including the relation between referenced objects, rather than relying on an automatic expansion of a short label. For a series, retain the same subject reference and visual requirements while changing one scene decision at a time.

Output settings and edition boundaries

The producer’s exact 4B profile lists JPEG, PNG and WebP output, with JPEG as the default. Explicit width and height values are documented with a minimum of 64 pixels each; the reference model-card example uses a 1024×1024 image. That example is a sampling configuration, not an assertion that every task must be square or that 1024 pixels is the model’s maximum. Choose the intended crop and display shape before preparing the visual brief, then compare fine lettering and small features at that size.

The published four-step example uses guidance_scale 1.0. Keep this distilled edition separate from FLUX.2 [klein] Base 4B, whose undistilled training signal is intended for adaptation, LoRA training and research. Base 4B and the 9B models are distinct choices, with different preparation and sampling considerations. The compact model’s standard controls also differ from [flex]’s adjustable steps and guidance; [max] has its own grounding-search capability.

Details to verify in the image

The exact model card says this model is not a factual information source. Written text inside an image can be inaccurate or distorted, and the generated scene may depart from the request. Prompt style has a substantial effect on adherence. Check spelling, quantities, labels, subject identity and the specific objects meant to remain; an attractive composition does not establish that a product feature or illustrated fact is correct. The maker also notes that statistical outputs can reflect training-data biases.

For a reviewed asset, keep the source references, final wording and selected model identity with the image. The release documentation distinguishes example pixel-level watermarking in reference code from C2PA provenance metadata applied in the manufacturer’s hosted output workflow. Retaining returned provenance helps explain an asset’s origin; it is separate from checking whether the requested visual change succeeded.