Image to Image AI: How It Works, How to Use It, and What Each Model Costs

Eight models in the Image studio accept a picture as input. The cheapest render costs 10 credits and the dearest costs 55, and the 55-credit one is not the best model on the list. It is GPT Image 1.5 with its quality dial turned up, which multiplies that model's own base price by 5.5.

Nothing in the picker tells you that. It also does not tell you that two of the cheaper models hand you a 2K render at their flat price while a dearer one charges exactly double for the same jump. Those two facts are worth more than any prompt trick, so they come first.

Image to image vs text to image

Text to image starts from nothing. Your words are the whole input and the model invents every pixel, which means each attempt re-rolls the composition, the pose, the framing, the lot. Get a shot you like on attempt four and attempt five will be a different shot.

Image to image starts from a picture you already have. The model reads that image, then applies the change you described on top of it. The room, the product, the face, the layout you uploaded stay recognisable instead of being reinvented.

So the choice is not about quality. It is about whether the thing already exists:

• Nothing exists yet and you are describing it into being. Text to image.
• Something exists and needs to change. Image to image. Repaint a photo as an illustration, swap a background, drop a real product into a new setting, clean up an artefact, push forty product shots into one consistent look.

The mechanics are otherwise identical. Pick a model, write a prompt, press generate. Image to image adds exactly one step at the front: attach the reference.

The eight models, and what one render costs

Cheapest first, in credits per image:

• Nano Banana — 10. One output size only.
• GPT Image 1.5 — 10 at medium quality, 55 at high. Square output only.
• Seedream 5.0 Lite — 14, flat. Renders at 2K.
• GPT Image 2 — 15 at 1K, 26 at 2K.
• Seedream 4.5 — 17, flat. Renders at 2K.
• Seedream 5.0 Pro — 18 at 1K, 36 at 2K.
• Nano Banana 2 — 20 at 1K, 30 at 2K.
• Nano Banana Pro — 45, and 2K costs the same 45.

Quantity multiplies all of that. You can ask for 1, 2 or 4 renders in one submission and every one of them is charged at full price; there is no batch rate. The number on the generate button is computed by the same code that actually deducts from your balance, so what you see there is what you pay.

Output size is where the money goes, and it is not distributed sensibly

Read that list again by size rather than by name and the ranking scrambles.

Seedream 5.0 Lite gives you a 2K image for 14 credits. It is flat-priced: there is no upgrade to buy, because its base tier is already the big one. Seedream 4.5 works the same way at 17.

Seedream 5.0 Pro charges 18 for 1K and 36 for 2K. So the Pro model's 2K render costs about two and a half times the Lite model's 2K render. Sometimes it earns that. On a background swap or a style pass it usually does not, and I would run Lite first every time.

Nano Banana Pro is the only model here where the bigger output is free. 45 credits at 1K, 45 credits at 2K. If you have chosen Nano Banana Pro and left the size on 1K, you have thrown away half the pixels you already paid for.

GPT Image 1.5's high tier is the worst deal on the page. 10 credits to 55 is a 5.5x multiplier, which lands it above Nano Banana Pro's flat rate. There are specific jobs where its high tier is the right call, mostly ones involving readable text inside the image. Reaching for it out of habit is how you burn a pack of credits in an afternoon.

Two limits belong with that list, because both of them will surprise you at the wrong moment. GPT Image 1.5 only renders square, so every other aspect ratio in the picker disappears when you select it. Nano Banana has a single output size and no upgrade path at all.

How to use image to image

1. Attach the reference. JPEG, PNG, WebP, GIF, AVIF and HEIC/HEIF are accepted, up to 3 MB per file. SVG is refused outright. If your source is a 12 MB export, downscale it first; the models work from a resized copy anyway.
2. Pick the model before you write the prompt. The aspect ratio and size controls change depending on what you selected, and so does the price on the button.
3. Describe the change, not the picture. This is the step everyone gets wrong. More on it below.
4. Set the output size. On Nano Banana Pro, take 2K; it is free. On Seedream 5.0 Pro, decide whether you actually need the 36-credit version.
5. Generate one, not four. Find the prompt at quantity 1. Once a prompt works, run it at 4 and keep the best.

Prompts that keep the original picture

The failure mode of image-to-image prompting is re-describing what you uploaded. The model can already see it. Every word you spend describing the subject is an invitation to redraw the subject, and it will accept.

Compare:

Bad: A red vintage car on a coastal road at sunset, shot on film, cinematic, highly detailed.

Good: Repaint this as a 1970s gouache travel poster. Flat colour blocks, visible paper grain, no photographic texture. Keep the car's position, angle and proportions exactly as they are.

Four patterns that hold up across all eight models:

Restyle

Convert this photograph into a pen-and-ink illustration with cross-hatched shadows. Line weight heavier on the foreground. Keep the framing and the subject's pose unchanged. White background where the sky is.

Background swap

Replace the background behind the subject with an overcast beach at low tide. Match the existing lighting direction, which comes from the upper left. Do not alter the subject, their clothing, or the edges of their hair.

Product placement

Place this bottle on a wet slate surface with shallow standing water under it. Single soft light from the right, one reflection. Keep the label text, the cap and the bottle's proportions identical to the reference.

Repair

Remove the power lines crossing the upper third of the image. Reconstruct the sky behind them to match the existing gradient and cloud texture. Change nothing else.

The common thread is a sentence that says what must not change. Half of a good image-to-image prompt is a lock list, and almost nobody writes one.

More than one reference image

Six of the eight models take up to 10 reference images in a single request. GPT Image 1.5 and Nano Banana take up to 16.

That is what makes combination prompts work: hand the model a subject in one picture and a jacket, a room or a colour palette in another, then say how the two should meet. Name them by position, because the model has no other way to tell them apart.

Put the person from the first image into the room from the second. Keep their face, hair and clothing exactly as in the first image. Match the second image's lighting: warm lamp light from the right, deep shadow on the left wall.

Vague multi-reference prompts fail in a particular way. Ask it to "combine these" and you tend to get an average of both inputs rather than one placed inside the other.

When image to image is the wrong tool

It is not a precision editor. If you need one pixel-exact region changed and everything else byte-identical, use an editor. The model re-renders the whole frame every time, so small unrequested drift in skin tone, grain and edges is normal rather than a bug.

It is also poor at arithmetic on the image. Asking to remove one of five objects, or to change the number of something, is the class of request that comes back wrong most often. Describe the state you want instead of the operation you want performed.

And if the reference is a low-resolution crop, no model on this list will invent detail that is convincing at 2K. Start from the best source you have.

Before you press generate

• Did you pick the model before writing the prompt, or after?
• Does the prompt say what must stay the same?
• Did you describe the change, or re-describe the photo?
• On Nano Banana Pro, is the size set to 2K? It costs nothing extra.
• Do you actually need Seedream 5.0 Pro's 36-credit render, or would Lite's 14-credit 2K do?
• Quantity 1 for the first attempt.

---

Model pages with the full picture on each family: Seedream, GPT Image, Nano Banana. Everything the platform sells, in one list: all models. Prompt-level detail on the newest of them: how to use Nano Banana 2 and Nano Banana 2 prompts. Credit packs and plans: pricing. Or go attach a picture and find out: open the Image studio.