Flux Kontext Prompt Guide: Edit Images While Preserving What Matters

Try Flux Kontext

Flux Kontext prompts work best when they describe a revision to an existing picture. If you already like a character's face, a product's shape, or a scene's composition, your instruction should identify the change and the details worth keeping. A beautiful description of an entirely different scene leaves those preservation decisions open.

This guide walks through a reference-based editing workflow on Anime AI Studio, with copyable prompts for common revisions. The examples are starting instructions, not verified output samples or promises of exact preservation. Keep your original image so every result has a clear comparison point.

The Flux Kontext prompt formula

Use four parts: target, change, preservation, and visual integration. The target identifies an object already visible. The change states the new result. Preservation names the features you want retained. Integration describes how the revised area should fit the existing picture.

Change the jacket worn by the person in the center to a forest-green
canvas jacket. Preserve their face, hairstyle, expression, pose, hands,
and the existing framing. Match the jacket's shadows to the soft light
coming from the window on the left.

This gives the editor a smaller task than “make this a fashionable portrait.” It also makes review easier: you can check the jacket color, face, pose, and light separately. You do not need to repeat everything visible in the source. Describe the details that determine whether the edit succeeds.

Avoid mixing incompatible directions. “Keep the exact pose” and “turn the person toward the camera” request different outcomes. If you want a new pose, acknowledge that change and preserve other features instead.

How to edit an image step by step

1. Open the Flux Kontext image editor and select the reference-image workflow available in the generator.
2. Upload a clear source image. Choose a version where the area you want changed is visible and sufficiently large.
3. Write one edit using the four-part formula. Start with a single object or attribute.
4. Check the available settings and any displayed generation cost before submitting. Use only controls actually shown by the interface.
5. Compare the output against the source at both full-frame size and a close crop of the edited area.
6. Save the accepted result and its prompt. For another edit, deliberately choose whether to upload this result or return to the original.

Do not treat every new prompt as a continuation of remembered conversation. The uploaded image is your reliable visual context. If you need the last output as a reference, supply it through the image input rather than assuming the generator remembers it.

Background replacement without redesigning the subject

A background edit has two jobs: preserve the foreground and make the new environment plausible. Specify where the subject stands and how light reaches them. Otherwise, an attractive background can still feel pasted behind the person.

Replace the plain gray background with a quiet railway platform at dusk.
Keep the person in exactly the same position and size within the frame.
Preserve their facial features, hair, clothing, and expression. Place the
platform behind them with a believable horizon and soft amber station
lights. Keep the subject clearly separated from the background.

Check the outline around hair, ears, and shoulders. Look for leftover pieces of the old background, newly invented accessories, or a horizon that cuts awkwardly through the subject. If the identity changes, shorten the environmental description and repeat the specific facial details that matter.

For a cleaner profile illustration, ask for a restrained background instead of a detailed city. A broad color field gives fewer competing shapes and leaves room for cropping later.

Clothing and accessory prompts

Identify garments through position and color, especially when an image contains multiple people. “Change her coat” can be ambiguous; “change the red coat worn by the person on the right” is easier to interpret.

Replace the red scarf around the central person's neck with a navy knitted
scarf. Keep its approximate length and drape. Preserve the coat, face,
hair, posture, and background. Show soft knitted fabric that follows the
existing folds and lighting.

If the scarf overlaps hair or a hand, inspect those boundaries. Clothing changes can affect nearby anatomy because the model must reconstruct partially covered regions. A smaller accessory edit is often a useful first exercise before a complete wardrobe replacement.

For a full outfit change, name the essential pieces and their relationship. Avoid an inventory of tiny decorations until the basic silhouette works. Save an accepted silhouette before adding buckles, trim, or embroidery.

Lighting changes with clear limits

Lighting affects the entire frame, so it is a broader edit than replacing a scarf. Decide whether you want a small color adjustment or a new time of day. Those are different revision sizes.

Keep the scene, camera angle, subject pose, and all objects in place.
Change the lighting to soft late-afternoon sunlight from the right.
Use warm highlights and gentle shadows while preserving natural skin
tones and detail in the dark clothing.

Check whether the light direction agrees across the face, ground, and background. If you need strict composition preservation, avoid simultaneously asking for rain, night lighting, a new lens, and a different environment. Split those into separate experiments and return to the original when comparing alternatives.

Converting a photo into an illustrated look

Style transfer changes many pixels even when the subject remains recognizable. Explain which identity cues should survive and what the desired illustration actually looks like. “Anime” alone leaves line weight, shading, and proportion undecided.

Render this portrait as a clean anime illustration with fine dark outlines,
two-tone cel shading, and a restrained pastel palette. Preserve the
person's short wavy hair, round glasses, head angle, clothing colors,
and friendly closed-mouth expression. Keep the original crop and simple
background arrangement.

Compare distinctive features rather than expecting photographic equivalence. Glasses shape, hair part, eyebrow position, and face proportions are useful checkpoints. If the result becomes too generic, reduce exaggerated style instructions and describe those cues more precisely.

For a new scene without a source image, the Seedream 5 prompt guide covers composition-first generation. That workflow gives you more freedom to invent a layout before editing it.

Replacing short text in an image

Text edits need exact wording and a clear target. Quote the old and new strings and specify the intended placement. Treat successful spelling as something to verify, even when the surrounding artwork looks convincing.

Replace the words "OPEN TODAY" on the front window with "NIGHT CAFE".
Keep the lettering in the same window area with a similar simple bold
style and warm cream color. Preserve the window frame, reflections,
building, and all other visible text.

Read every letter at full resolution. Check that the old words did not remain as a shadow or duplicate line. If precise typography is essential, generate or edit the artwork with an empty sign and add the final lettering in a dedicated design editor.

Diagnose the result before rewriting the prompt

Change one hypothesis at a time. If a background swap fails because the person becomes smaller, revise the framing instruction before adding more scenery. This produces useful information about the next attempt instead of making each prompt a completely new experiment.

Plan a short revision sequence

Suppose you need a portrait with a new background, scarf, and illustrated finish. First test the background on the original. Then use the accepted result for the scarf. Finally, test the style conversion while comparing identity with the original portrait.

Keep intermediate files with descriptive names such as portrait-background, portrait-scarf, and portrait-illustration. Repeated edits can accumulate changes you did not request. If the face or composition deteriorates, return to the strongest earlier image instead of attempting to repair every new defect.

Before downloading the final choice, check the full frame, the target region, and the preserved regions. A successful edit satisfies the requested change while keeping the visual anchors you selected. Try that process in Flux Kontext, starting with one modest revision you can judge clearly.

FAQ

How long should a Flux Kontext prompt be?

Use enough detail to identify the edit and preservation rules. A few specific sentences are usually a better starting point than a long list of adjectives. Add detail only when it resolves a visible problem.

Can a prompt guarantee that a face stays identical?

No. Preservation instructions express your goal, but generated edits can alter identity details. Compare against the original and keep the source available for retries.

Should I describe the entire reference image?

Usually you can focus on the target and the features that must remain. Describe more context when multiple similar objects make the target unclear.

What should I do when several edits fail together?

Separate the request into smaller revisions. Test the most important change first, review it, and use an accepted intermediate image for the next step.