Flux Kontext Editing Guide: Edit Images With Instructions

FLUX.1 Kontext is the in-context editing model from Black Forest Labs, the team behind the Flux image models. You hand it an image and a plain-language instruction such as 'change the jacket to dark green', and it edits that part while keeping the rest of the picture as it was. Because each edit can feed into the next, Kontext is the most practical way in the Flux family to keep a character or product consistent across a series. This guide covers how to phrase instructions, how to chain edits without drift, and which settings matter.

What Kontext does differently

Most text-to-image models start from noise every time, so a 'small change' means regenerating the whole image and hoping the rest survives. Kontext works in context: it reads your input image together with your instruction and produces a new image that respects both. In practice that gives you:

• Local edits. Change one object, colour, piece of text or background element and the rest of the frame stays put.
• Style transfer with structure kept. Ask for 'the same scene as a watercolour illustration' and the composition survives the change of medium.
• Character consistency. The same face, hair and outfit carry across new poses, settings and expressions, because the model is looking at the reference rather than re-imagining it from text.
• Iterative editing. Each output can be the input for the next edit, so you build a result step by step.

Kontext comes in three variants. Kontext [dev] has open weights and runs locally or on open hosts; Kontext [pro] and Kontext [max] are hosted API models, with max aimed at the hardest prompts and at typography. The same instructions work across all three, and max usually follows complex requests more faithfully. For how the wider family fits together, see Flux models explained.

How to write an edit instruction

A Kontext prompt is an instruction, not a scene description. The difference matters. If you describe the whole image again, the model treats it as a request to redraw everything, and you lose the stability you came for. These habits hold up:

1. Name what changes and what stays. 'Change the car to red, keep everything else the same' beats 'red car'. The second half is not filler; it tells the model the scope.
2. Be specific about the target. If there are two people, say 'the woman on the left'. If there is text, quote it: replace the sign text with 'OPEN LATE'.
3. One change per edit. Two or three related tweaks in one instruction usually work, but a long list is where things go wrong. Chain edits instead.
4. Use verbs the model understands. Change, replace, remove, add, make, turn into, move. Avoid metaphors.
5. For style changes, name the medium and what to preserve. 'Convert to a pencil sketch while keeping the exact composition and the person’s face.'

Avoid negative-only prompts such as 'no hat'. Say what should be there instead: 'remove the hat and show her short dark hair'.

Keeping a character consistent across edits

This is where Kontext earns its keep. The workflow that holds up best:

1. Make a clean master image. Generate or photograph your character in neutral light, facing the camera, in the outfit you want to keep. Fix anything wrong now, because every later image inherits it.
2. Edit from the master, not from the last edit, when you can. Each generation adds small changes. Ten edits in a row and the face has drifted. Branching from the master keeps the drift to one step.
3. Refer to the character explicitly. 'Keep this exact person' or 'same woman, same outfit' in every instruction reminds the model what is fixed.
4. Change one dimension at a time. New pose first, then new setting, then new lighting. Changing all three at once gives the model licence to reinterpret.
5. Lock the seed when you want comparable results. Changing only the instruction with the seed fixed makes side-by-side variants easier to judge.

Expect the occasional edit to shift a jawline or a hair parting. When that happens, go back a step and rephrase rather than piling a corrective edit on top. The same discipline applies in other tools; our Midjourney consistent characters guide covers the reference-image approach there.

Settings at a glance

Hosted playgrounds (the BFL playground, Replicate, fal, Krea, Freepik and others) expose slightly different controls and labels. These are the ones you will meet most often.

Label names vary by host, so look for the control that does the job rather than the exact word.

Try it

Upload a portrait or product shot, then use this instruction.

Change the background to a quiet seaside promenade at golden hour. Keep the person exactly the same, including face, hair, outfit and pose, and match the new warm lighting on their skin and clothes.

Swap the background phrase first and notice how the lighting clause keeps the subject believable. Then try a second edit on the result: 'add a light grey linen scarf, keep everything else the same'. If the face shifts, return to the first image and rephrase.

Common problems and fixes

• The whole image changed. You described the scene instead of the edit. Rewrite it as an instruction and add 'keep everything else the same'.
• The edit was ignored. Make the target more specific, nudge guidance up slightly, or split the instruction in two.
• Text came out wrong. Quote the exact words, keep them short, and use Kontext [max] if you have access. Very long text is still unreliable.
• Soft output after several edits. Each pass loses a little detail. Branch from the master image, or finish with an upscale step.
• Edges or new areas look invented. You changed the aspect ratio. Match the input ratio, or crop afterwards.