FLUX Kontext: image editing, model choices, and practical prompts

What FLUX Kontext does - and what its name means

FLUX Kontext is built around contextual image editing: a supplied image provides visual information, and a written instruction specifies how that image should change. Rather than describing an entire scene from scratch, you can ask for a different background, a revised object color, a new visual treatment, or replacement wording within a design. The family also supports image generation.

The important distinction is between making a new picture and transforming an existing one. For an edit, the source carries information that would be difficult to recreate in a prompt: a person’s appearance, an object’s shape, a composition, or a particular arrangement of elements. Kontext is designed to retain relevant context while applying the requested transformation.

The full family name is FLUX.1 Kontext. It is not another spelling of FLUX.2, and settings, prices, or licensing attached to another FLUX model should not automatically be applied to it.

• Choose an editing workflow when the source image already contains the subject or layout you need.
• Choose generation when there is no existing image to preserve.
• Use conventional retouching when an exact pixel-level change matters more than a generated interpretation.

Think of Kontext as an instruction-driven transformation tool, not a guarantee that everything outside the requested change will remain identical. That distinction matters for product photographs, portraits, and designs containing approved text.

How context editing works: specify the change and the invariants

A useful Kontext instruction has two parts: the change you want and the properties that must remain stable. A prompt that says only “make this more professional” leaves the model to interpret the background, lighting, styling, and composition. A narrower instruction gives you a result that is easier to evaluate.

Product example: “Replace the tabletop with pale gray stone. Keep the bottle shape, cap, label wording, camera angle, and framing unchanged.” This identifies the editable region conceptually and names the details that define the product.

Portrait example: “Change the background to a plain warm-gray studio backdrop. Preserve the person’s facial features, hairstyle, clothing, pose, and expression.” Review the face afterward; a preservation instruction is a constraint, not proof of identity fidelity.

Text-edit example: “Replace ‘SPRING SALE’ with ‘AUTUMN SALE’. Keep the lettering style, placement, background, and other text unchanged.” Kontext is designed for manipulating embedded text, but spelling and layout still need inspection.

Start with one transformation. If you request a new background, different clothes, altered lighting, and a new pose simultaneously, it becomes harder to identify which instruction caused an unwanted change. Separate edits also let you stop at the last acceptable version.

Keep an original copy alongside each accepted result. Repeated edits should be judged against the original, not just the previous output; small shifts in facial features, proportions, or typography can otherwise become easy to miss.

A six-step workflow for reliable FLUX Kontext edits

1. Define the deliverable. Decide whether you need a product variation, portrait edit, revised headline, or illustrated scene. List the details that would make the result unusable if changed.
2. Select the exact model and service. Check that the interface names FLUX Kontext rather than simply FLUX. Review its upload requirements, output settings, billing terms, and applicable license before submitting work.
3. Prepare the source image. Use an image where the subject and relevant details are visible. Retain the original file separately. If a tiny label is central to the task, inspect its legibility before asking the model to preserve it.
4. Write one bounded instruction. State the requested change first, then the elements to retain. For example: “Make the jacket dark green. Keep its cut, fabric texture, buttons, the person’s pose, and the background unchanged.”
5. Review the whole output. Compare the edited region and the supposedly unchanged areas. Check faces, hands, logos, shadows, object counts, text, and crop. An attractive image can still fail the brief.
6. Iterate from a clean checkpoint. Save acceptable versions. If the subject drifts, return to the original or the last sound edit rather than adding more corrective instructions to a damaged result.

For a repeatable process, keep the source, prompt, model name, and accepted output together. This makes later revisions easier and prevents confusion when a platform offers several FLUX variants.

Finish with ordinary production checks: dimensions, readability at the intended display size, and any required manual cleanup. Generative editing does not replace approval of the final asset.

FLUX Kontext Pro, Dev, and alternative models compared

When choosing between FLUX Kontext Pro and FLUX Kontext Dev, match the exact model offering to your access, deployment, and licensing requirements. Do not treat a “Dev” label as a universal promise of free commercial use. Model terms and the rules of the service delivering it are separate considerations.

If your shortlist includes flux 2 or midjourney 7, define the task before comparing outputs: modifying an existing product photograph is a different requirement from creating a new campaign concept.

As of September 2026, FLUX.2 Pro editing starts at $0.045 per megapixel. That is a FLUX.2 price, not a Kontext quote. Likewise, the non-commercial designation for FLUX.2 Dev should not be used to infer Kontext Dev’s license. Check Black Forest Labs’ pricing for the exact offering you intend to use.

Limits, edit costs, and commercial-use checks

Kontext’s emphasis on subject consistency does not remove the usual failure modes of generated images. A face may shift, a product edge may change, lettering may become inaccurate, or an object may appear in the wrong position. Treat “keep unchanged” instructions as requirements to verify, particularly when the image represents a real person or a purchasable item.

• Identity: compare facial proportions, hairline, distinguishing features, and expression.
• Products: check shape, proportions, closures, labels, and visible accessories.
• Text: inspect every character, number, punctuation mark, and line break.
• Scene structure: check object counts, overlap, perspective, reflections, and shadows.

Do not assume every Kontext interface has the same resolution ceiling, reference-image allowance, or export options. Those are practical selection criteria: check them before building a workflow around a particular service.

Budget for accepted outputs, not just individual generations. If a billed attempt costs C and a usable image takes four attempts, the generation cost is 4 × C before any manual finishing. A subscription or credit package can introduce different limits, so inspect what an edit actually consumes.

Before commercial publication, check the exact model license and service terms through the applicable offering and Black Forest Labs’ licensing information. Separately confirm your rights to uploaded photographs, trademarks, and recognizable likenesses. A model license does not grant permission to use somebody else’s source material.

Useful Kontext tasks and when to choose another editor

Product-scene variations: change a setting while keeping the item recognizable. Start with background or surface replacements, then compare the product against the original. Reject outputs that alter packaging, proportions, or claims printed on a label.

Portrait background changes: request a different environment without changing the person’s pose or expression. This is useful when the background is the problem, but it requires a separate identity check before delivery.

Illustration and storyboard revisions: alter a scene’s setting or visual treatment while attempting to retain the main character. Keep a reference sheet or original frame visible during review so that clothing, silhouette, and distinguishing features do not gradually drift.

Embedded headline revisions: change a short phrase within an existing design. Specify the exact replacement in quotation marks and list what must remain. If the wording is legally sensitive or must fit an exact typographic system, manually setting the final text is the safer finishing step.

Choose a conventional photo editor for exact crops, precise color corrections, or layout work that must be reproducible without generated variation. Pict.AI is one option for general photo work: it offers an AI photo editor app for iPhone and Android, plus a website with guides and free image tools. It is not a substitute for checking which model a particular task requires.

The practical rule is simple: use Kontext when the desired change benefits from image understanding and generation; use deterministic editing when exactness is the primary requirement.