FLUX.1 Kontext: Editing Images by Instruction

What FLUX.1 Kontext actually is

Most image models take text in and give an image back. Kontext takes text and an image in, which is what "in-context" means here: your photo is part of the prompt, not a starting point it is free to discard. That framing is the whole point. The model is asked to change the one thing you named and to leave the rest of the frame - the faces, the product, the layout, the lighting - recognisably intact.

In practice that turns editing into a conversation. You upload a photo, type a sentence, look at the result, then type the next sentence. There is no mask to paint, no layer stack, and no strength dial to guess at. Black Forest Labs, the German lab behind the FLUX family, released the first Kontext models in 2025 and positioned them squarely at this iterative editing loop rather than at one-shot art generation - although Kontext can generate from text alone too.

Instruction editing vs inpainting vs image-to-image

These three get mixed up constantly, and picking the wrong one is the usual reason an edit comes out badly. Instruction editing is the generalist: it is fastest to use and best when you can name the change in words. Inpainting is the specialist: you mark an exact region, so it is the safer choice when the edit has to stay inside a precise boundary - a logo on a shirt, one face in a crowd. Image-to-image is the transformer: it re-imagines the whole frame toward a prompt with a strength slider, which is what you want for a full style change and exactly what you do not want for a small fix.

On img.now those three map to the AI Magic Edit tool (instruction), the object remover (masked) and image to image (whole-frame restyle).

Where it is strong

Keeping a subject consistent. The headline capability: run the same character or product through several edits - new background, new outfit, new angle - and it stays identifiably the same thing. That is what makes it useful for a product line or a set of marketing shots rather than one lucky image.

Local changes without a mask. Recolour an object, swap a background, change weather or time of day, add or delete an item, restyle clothing. You name the target in words and the model works out where it is.

Style transfer that survives. Push a photo into an illustration, a 3D render or a painting while the composition and the subject's identity hold.

Editing text that is already in the picture. Signage, labels and packaging copy can be swapped, which older editing models handled badly.

Where it struggles (worth knowing before you start)

Drift across many passes. Every edit is a redraw, so quality and likeness degrade slowly as you chain edits. Do the big structural change first, keep the chain short, and re-upload the original rather than the fifth generation.

Compound instructions. "Make the car red, add a dog, change the sky and remove the fence" usually lands one or two of the four. Split it up and run one sentence at a time.

Small typography. Text can be edited but fine print, exact fonts and brand marks are still unreliable - set real copy over the image in a design tool afterwards.

Precise geometry. Exact object counts, hands, reflections and engineering-accurate detail remain hard for every model in this class, Kontext included.

How to write an instruction it will follow

Name one change, name the thing precisely, and say what should stay. "Change the wall behind the sofa to warm cream, keep the sofa and the lamp as they are" beats "nicer living room" every time. Use the words a person would use to point at the object - "the red mug on the left", not "the object in the mid-ground".

Describe the destination, not the process. "Make it look like an overcast afternoon" works; "reduce the saturation and add a cool LUT" does not, because the model paints results, not adjustment layers. If a result strays, do not add more words - shorten the sentence and run it again, since each pass varies. And start from a clean, well-lit, reasonably large source: an edit cannot recover detail the original never had.

Change the car colour to matte black, keep everything else the same
Replace the cloudy sky with a clear blue sky
Remove the poster from the wall and leave the brickwork clean
Put the product on a marble countertop in soft daylight
Same character, now standing on a beach at sunset

Our prompt-writing guide goes deeper on phrasing, and the image-to-image guide covers when a whole-frame restyle beats a targeted edit.

The three Kontext variants

Black Forest Labs ships Kontext in tiers rather than as a single model. Which one a given site or app is using changes the price, the quality and - importantly - what you are allowed to do with the output.

Model line-ups move fast. Check Black Forest Labs directly for the current variants, per-image pricing and licence wording before you build on any of them.