Image-to-Image: When to Give the Model a Starting Picture

There's a small image icon at the left of the prompt bar. It's the difference between asking the model to invent something and asking it to work from something.

Most people never touch it, and then wonder why their product never looks like their product.

---

What changes when you attach one

Text-to-image starts from nothing. You describe, the model invents. Every run gives you a different thing.

Image-to-image starts from your picture. The model keeps its structure, composition and subject, and applies whatever you describe on top.

That's the whole difference, and it's why the same prompt behaves so differently with and without a reference. Without one, "make it warmer and more editorial" produces a new picture that's warm and editorial. With one, it produces your picture, warmer and more editorial.

---

When a reference is the right answer

You have a real product. This is the big one. No prompt describes your bottle accurately enough for a model to draw your bottle. A photograph does. Attach it and describe the scene you want it in.

You need the same person or character twice. Text alone will not give you the same face in two runs. A reference will get you close, and adding it to every generation in a set keeps them recognisably related.

You like the composition and want a different look. Same layout, different season, different palette, different medium.

A previous generation was nearly right. Feed the output back in as the input and push it one step further. This is the loop that gets you from 80% to 95%.

You're working from a sketch. A rough drawing plus a description of the finished style is far more controllable than describing the drawing in words.

---

When to skip it

• You're exploring and don't know what you want yet: a reference narrows the space, and early on you want it wide.
• The thing you want has nothing to do with any picture you have.
• You want genuine variety. References make outputs converge; that's the point of them, and it's a problem when you're brainstorming.

---

Describe the change, not the picture

The most common mistake with a reference attached is re-describing everything that's already in it.

Doesn't work: "A white ceramic bottle with a gold cap on a marble surface with soft light", when the bottle is already in the reference. You're competing with the picture.

Works: "Place it on a dark slate surface with dramatic side lighting."

Everything you don't mention is inherited. Say only what should be different.

---

Which models hold a reference well

• Flux Kontext Dev (20 credits), built for exactly this. Give it an image and an instruction and it changes what you named and leaves the rest.
• Nano Banana / Nano Banana Pro (20 / 75–125), strong at keeping a subject recognisable while changing its surroundings.
• Seedream 4.5 and 5.0 (20 / 30–60), reliable at product work; the reference survives.
• GPT Image 2 (10–200), follows complicated multi-part instructions about the reference better than most.

If a reference is being ignored, the model is usually the reason before the prompt is.

---

The habit that makes this pay

Generate a subject you like once, cheaply. Then use it as the reference for everything else.

One good bottle shot becomes the bottle in a lifestyle scene, on a coloured background, in a flat lay and in a seasonal campaign, all recognisably the same bottle, because they all came from it.

Trying to re-describe it each time gives you five different bottles, which is the single most common way product image sets go wrong.

---

Try it

Open the image generator, attach any photo you have, and write only "in soft morning light on a pale wooden table." Twenty credits on Seedream, and you'll see immediately how much of your picture survives.

Then: editing an image you already have · prompt writing