Flux Kontext - AI Image Generation and Editing Studio

Prompt

Model

Aspect Ratio

Output Format

Safety Tolerance: 2

Control content filtering sensitivity (0=strictest, higher=more permissive)

Auto-translate to English

Enhance Prompt with AI

Your generated image will appear here

About Flux Kontext

Flux Kontext takes an image and an instruction and returns the same image after the change. The composition, the subject and the light stay where they are; only what you describe moves. This page runs the Pro and Max tiers in two modes: text to image for something new, image editing when the frame already exists.

The model is built for instruction-following edits rather than composition. Give it a photograph of a jacket and ask for the fabric to become suede, and the garment keeps its cut, its seams and its place in the frame. That is the opposite of the Nano Banana page, where several references are combined into a new scene; here you supply one image and the job is to leave the rest alone. Two tiers are offered, Pro and Max, and Pro is selected by default.

The controls are narrow on purpose. Six aspect ratios cover ultrawide at 21:9 through vertical at 9:16, output is JPEG by default with PNG available, and safety tolerance runs 0 to 2 in editing mode and 0 to 6 in text to image, where 0 is the strictest setting. Two switches matter for prompt writing: automatic translation to English, on by default, and prompt upsampling, which expands a short instruction before it is sent. Uploads accept PNG, JPG, JPEG and WEBP up to 10 MB, and the file is stored before the request goes out.

This is not a stylization engine. There is no version selector, no variety slider and no art-direction dial, so if the goal is a look rather than a correction, Midjourney is the better page. What you get instead is feedback you can rely on: progress reports with a remaining-time estimate while the job runs, and a refund when a generation fails. Each job also stores its prompt, ratio and settings alongside the result, so a correction that worked can be reconstructed exactly on the next photo.

Instruction wording does more work here than any control. Name the object and the attribute that should change in one clause, then stop: the jacket's fabric, the colour of one wall, the surface under a product. Two changes in a single sentence are where edits come back uneven, because the model may apply the first and soften the second, and the result gives you no way to tell which part failed. If a photo needs two corrections, run two edits and compare each one against the frame you started from.

The two modes are better treated as a two-step workflow than as alternatives. Text to image needs no upload and allows the wider safety range, so it can produce the base frame; image editing then reworks that frame inside the tighter range, and each mode is its own run with its own charge. Settle the composition in text mode first, then edit, so each step carries a single decision. Choose PNG when the result will be edited again, since a later JPEG save compresses the pixels a second time, and JPEG when the file goes straight into a page or to a client.

How Flux Kontext works

1
Choose the mode
Text to image builds a new picture from the prompt. Image editing rewrites one you already have, and it will not submit without an upload.
2
Upload the frame and name one change
Describe the object and the change, not the whole scene. Anything you leave out of the instruction is what the model is meant to preserve, so resist adding detail that is already visible.
3
Pick Pro or Max, set the ratio, generate
Choose one of the six ratios, JPEG or PNG, and a safety tolerance. The run streams progress and an estimate, and the button shows the cost before you commit.

What Flux Kontext is best for

Product retouching
Change the surface a product stands on, or the colour of one panel, without re-shooting the item or rebuilding the set around it.
Seasonal variants
Recolour a garment or swap a prop for a holiday version while the model, the lighting and the framing stay identical across the set.
Photo clean-up
Remove an object, replace a background or change the material on one surface of an interior shot, with the perspective still holding together.
Sketch to render
Feed a line drawing or a blockout and describe the finished material. The layout you drew survives the render, which is the part that is hard to get elsewhere.

Flux Kontext at a glance

Model: FLUX.1 Kontext Pro (default) or Max
Modes: Text to image, image editing
Input image: One file, PNG/JPG/JPEG/WEBP, up to 10 MB
Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
Safety tolerance: 0 to 2 in editing mode, 0 to 6 in text to image
Output: JPEG by default, PNG available

Flux Kontext FAQ

What is the difference between Pro and Max?: Both appear in the model dropdown and follow the same workflow, and Pro is selected by default. Use Pro while you are still wording the instruction and switch to Max for the pass you intend to keep.
Will the rest of the image change?: The premise of the model is that it should not. Describe a single change and the subject, framing and light are meant to survive it. If you want a new scene assembled from several references, the Nano Banana page is the better fit.
What does safety tolerance do?: It controls how strict content filtering is, with 0 as the strictest setting. The slider stops at 2 in image editing and goes up to 6 in text to image, and the workspace starts it at 2.
Can I write the prompt in Chinese?: Yes. Automatic translation to English is on by default and runs before the request is sent, which matters because the model follows English instructions more reliably. Turn it off if your prompt is already in English.
How is a run billed, and what if it fails?: The cost is shown on the generate button before the request goes out, and a failed generation refunds the credits automatically. While the job runs you can follow the progress bar and the estimated time remaining in the workspace.
How many changes should one instruction make?: One. Name the object and the single attribute that should change, and leave everything else out, because the parts you do not mention are the parts the model is asked to preserve. Two changes in one sentence tend to come back applied unevenly. Run them as separate edits instead, and keep the same source file for both so you can see what actually moved. Each run shows its cost on the button before it is sent.

About Flux Kontext

How Flux Kontext works

Choose the mode

Text to image builds a new picture from the prompt. Image editing rewrites one you already have, and it will not submit without an upload.

Upload the frame and name one change

Describe the object and the change, not the whole scene. Anything you leave out of the instruction is what the model is meant to preserve, so resist adding detail that is already visible.

Pick Pro or Max, set the ratio, generate

Choose one of the six ratios, JPEG or PNG, and a safety tolerance. The run streams progress and an estimate, and the button shows the cost before you commit.

What Flux Kontext is best for

Product retouching

Change the surface a product stands on, or the colour of one panel, without re-shooting the item or rebuilding the set around it.

Seasonal variants

Recolour a garment or swap a prop for a holiday version while the model, the lighting and the framing stay identical across the set.

Photo clean-up

Remove an object, replace a background or change the material on one surface of an interior shot, with the perspective still holding together.

Sketch to render

Feed a line drawing or a blockout and describe the finished material. The layout you drew survives the render, which is the part that is hard to get elsewhere.

Flux Kontext FAQ

What is the difference between Pro and Max?

Both appear in the model dropdown and follow the same workflow, and Pro is selected by default. Use Pro while you are still wording the instruction and switch to Max for the pass you intend to keep.

Will the rest of the image change?

The premise of the model is that it should not. Describe a single change and the subject, framing and light are meant to survive it. If you want a new scene assembled from several references, the Nano Banana page is the better fit.

What does safety tolerance do?

It controls how strict content filtering is, with 0 as the strictest setting. The slider stops at 2 in image editing and goes up to 6 in text to image, and the workspace starts it at 2.

Can I write the prompt in Chinese?

Yes. Automatic translation to English is on by default and runs before the request is sent, which matters because the model follows English instructions more reliably. Turn it off if your prompt is already in English.

How is a run billed, and what if it fails?

The cost is shown on the generate button before the request goes out, and a failed generation refunds the credits automatically. While the job runs you can follow the progress bar and the estimated time remaining in the workspace.

How many changes should one instruction make?

One. Name the object and the single attribute that should change, and leave everything else out, because the parts you do not mention are the parts the model is asked to preserve. Two changes in one sentence tend to come back applied unevenly. Run them as separate edits instead, and keep the same source file for both so you can see what actually moved. Each run shows its cost on the button before it is sent.