Editing an Image by Describing the Change (Instead of Rolling Again)

You generate an image. It's almost right, good composition, good light, but the shirt is the wrong colour.

So you press generate again. And you get a different picture, where the shirt is still wrong and now the composition is worse too.

That's the trap. Generating is a dice roll; editing is not. Here's how to do the second thing.

---

Regenerating is not editing

When you run the same prompt again you get a fresh, unrelated image. The parts you liked were never guaranteed: they were luck, and the next roll spends that luck on something else.

Editing keeps the image and changes the one thing you named. It's cheaper, it's faster, and it's the only way to converge on a result rather than circling it.

---

The models that edit

• Flux Kontext Dev, 20 credits. Purpose-built for instruction editing. Attach an image, say what to change, and it changes that and leaves the rest alone. This is the default for edits.
• Nano Banana, 20 credits. Very good at keeping the subject intact while swapping surroundings or attributes.
• GPT Image 2, 10–200 credits. The one to reach for when the instruction is complicated or involves reasoning about the picture. Expensive; save it for edits the cheaper models refuse.
• Seedream 4.5 / 5.0, solid on product edits, especially surfaces and packaging.

Compare 20 credits for an edit against 125 for a Nano Banana Pro re-roll that gives you a different picture anyway.

---

Writing an edit instruction

Three rules, and they're all the same rule really: say less.

One change at a time. "Make the shirt navy" works. "Make the shirt navy, add a window, change it to evening, remove the mug" produces a mess where possibly none of the four happened properly. Chain them instead: run the first, feed the result back in, run the second.

Imperative, not descriptive. "Change the wall behind her to warm terracotta": not "a woman in front of a warm terracotta wall." The second reads as a fresh prompt and you'll get a fresh picture.

Name the thing precisely. "The mug on the left", "the text at the top", "her jacket." Vague targets get vague edits.

---

What edits well

• Colours, clothing, walls, packaging, props
• Backgrounds and settings, same subject, somewhere else
• Time of day and weather, often the single highest-value edit, because it changes the whole mood for 20 credits
• Adding or removing an object
• Style and medium, photograph to illustration, same composition

What struggles

• Small text. Editing text usually mangles it. See getting readable text into an image.
• Faces, if you need the person unchanged. Edits near a face tend to shift it slightly. Check identity carefully.
• Fine detail in a busy area, hands, jewellery, patterns.
• Anything requiring the model to understand what's behind an object. It invents, and sometimes badly.

---

Sometimes a photo tool is the cheaper answer

Before writing an edit instruction, check whether one of the photo tools already does the job, they're purpose-built, cheaper and more predictable:

You want to	Use	Cost
Remove an object cleanly	Object remover	20
Cut the subject out	Background remover	10
Put it in a described scene	Background change	20
Clean up skin or surfaces	Photo retouch	20
Change the aspect ratio	Image resize	20

Reach for an editing model when the change is creative rather than mechanical.

---

The chain that works

1. Generate cheaply until the composition is right
2. Edit the colour
3. Edit the background
4. Retouch
5. Enhance, once, at the end

Five deliberate steps, each cheap, each keeping what the last one earned. That's the difference between a workflow and a slot machine.

---

Try it

Take an image you've already generated, attach it on Flux Kontext Dev, and write one instruction: "change the background to a dark slate wall." Twenty credits, and the subject comes back the same.

Then: image-to-image · the photo tools