Image to image AI is an image generator that starts from a picture instead of a blank canvas. You give it two things: an image and a sentence describing what should change. It gives you back a new image that follows the sentence and, if the prompt is good, keeps the parts of the original you care about.
That sounds simple, and with current models it mostly is. But there are two very different technologies hiding under the same name, a few prompt habits that make a big difference, and some limits worth knowing before you spend credits. This guide covers all three.
What image to image AI actually does
A model that edits images reads your upload the way it reads a prompt: as information about subject, layout, lighting and color. It then generates a new picture that satisfies your instruction while staying consistent with that information.
The important word is generates. Even when only the background changes, the model usually redraws the whole frame. That is why a face can come back slightly different, or why a logo on a T-shirt can lose a letter. Good models keep these drifts small. None of them make them impossible.
Typical jobs:
- Restyling: a photo becomes a watercolor, a cartoon, a pencil sketch or a 3D render.
- Background swaps: the same person or product, somewhere else.
- Targeted edits: new clothes, a different season, a cleared table, a changed wall color.
- Concept work: a rough sketch or a phone photo of a room turned into a finished-looking design.
Two generations of image to image
Classic img2img and denoising strength
If you used Stable Diffusion around 2022 to 2024, you know "img2img" as a slider. The pipeline added random noise to your picture, then denoised it back toward your prompt. The denoising strength (usually 0 to 1) decided how much noise was added:
- Around 0.2 to 0.4: colors and textures shift, the layout stays. Good for light restyles.
- Around 0.5 to 0.7: shapes start to move, faces change, new details appear.
- Above 0.8: your image is little more than a color hint, and you get mostly a new picture.
The model never "understood" an instruction like "keep her face". It only knew how far to wander from the pixels you gave it. So you were always trading likeness against change. Want a strong style? You lose the face. Want the face? You get a weak style. Fixing that took extra tools like ControlNet, inpainting masks and face-restoration passes.
Instruction-following image editors
Newer models, such as Google's Nano Banana 2 (Gemini 3.1 Flash Image) and OpenAI's GPT Image, work differently. They read the image and the text together and treat the prompt as an instruction. "Change the background to a beach and keep her face and dress the same" is something they can actually follow, because the model knows which part of the picture is the face and which is the background.
There is no denoising slider. You control the amount of change with words. That is both easier and less predictable: the same prompt can give two slightly different results, and vague words give vague edits.
Image to image vs. text to image vs. old img2img
| Text to image | Classic img2img (Stable Diffusion) | Instruction-based image to image | |
|---|---|---|---|
| Starts from | A prompt only | An image + prompt + denoising strength | An image + prompt |
| Keeps your subject | No, it invents one | Only at low strength | Yes, if the prompt says so |
| How you control change | Prompt | Strength slider | Wording of the prompt |
| Best for | Ideas from nothing | Light restyles, local workflows | Edits to a photo you already have |
| Typical failure | Subject is never the same twice | Faces melt at high strength | Small details drift, text changes |
If you have nothing to start from, use text to image. If you want your photo, changed, use image to image.
Prompts that work: say what changes and what stays
The single most useful habit is to write two halves in every prompt:
- What to change, as concretely as you can.
- What to keep, naming the things you would be upset to lose.
Models protect what you name. They feel free to change what you don't. Here are prompts that follow that pattern.
Restyle a portrait
Turn this photo into a soft watercolor painting on textured paper. Keep her face, hairstyle, expression and the blue scarf exactly the same.Swap a background
Replace the background with a quiet beach at sunset with warm side light. Keep the man, his clothes and his pose unchanged, and match the lighting on him to the new scene.Product shot for a store
Place this bottle on a white marble counter in a bright kitchen, with soft morning light from the left. Keep the bottle, the label text and the cap exactly as they are.Change one item of clothing
Change his gray hoodie to a navy blazer over a white shirt. Keep his face, hair, glasses and the background the same.Season or time of day
Make this street scene look like a snowy winter evening with lit shop windows. Keep the buildings, the street layout and the parked cars in the same positions.Room makeover
Restyle this living room in a Japandi style: light oak, linen, low furniture, warm neutral colors. Keep the windows, the walls and the camera angle the same.Sketch to finished concept
Turn this pencil sketch of a sneaker into a clean product render on a light gray background. Keep the shape, the sole pattern and the proportions from the sketch.A few habits that help:
- One main change per prompt. "New background, new outfit, new hair and make it anime" asks for four edits at once. Do them in steps.
- Name the medium or the light. "Oil painting with thick brushstrokes" beats "artistic". "Overcast daylight" beats "nice lighting".
- Don't re-describe the whole photo. The model can see it. Spend your words on the change and the things to protect.
- Keep text short and quote it. If you need words in the image, put them in quotes, for example: a sign that reads "OPEN". Long text still often comes out wrong.
- Iterate on the prompt, not the photo. If the result is close, tweak one phrase and run again before you switch images.
Standard, Pro or HD: which model to pick
On the image to image AI tool you choose one of three tiers per image. They take the same upload and the same prompt.
Standard: Nano Banana 2 at 2K, 5 credits. This is the daily free tier. Without an account you get one free Standard run a day; after you sign in you get 10 free credits a day, which is two runs. Free results are JPGs up to 2048 px with a small watermark; runs paid with credits have no watermark. Start here. For most restyles, background swaps and outfit changes it is all you need.
Pro: GPT Image 2.5 at 2K, 6 credits. A different model family, not just a bigger version. Switch when Standard ignores part of a long instruction, when the layout is complex, or when you simply want a second opinion on the same prompt. Running the same prompt on both and keeping the better one is a reasonable habit for important images.
HD: Nano Banana 2 at 4K, 8 credits. The same model as Standard at four times the pixels. Pick it when the result is going to print, to a large screen, or will be cropped hard afterwards. It doesn't follow prompts any better than Standard, it just gives you more resolution to work with.
A simple rule: try Standard first, move to Pro if the edit is wrong, move to HD if the edit is right but too small.
Limits you should know about
- Small details drift. Logos, jewelry, fingers and fine patterns can change slightly, even when you ask the model to keep them. Zoom in before you publish.
- Small faces are fragile. A face that fills the frame is kept well. Five faces in a group shot, each a few hundred pixels wide, may each shift a little.
- Text is risky. Short labels often survive. Paragraphs, small print and non-Latin scripts are where models still invent letters.
- It is not a pixel editor. If you need the rest of the image to be exactly identical, use a masked tool such as remove object from image, which copies unbrushed pixels back from the original, or the AI image extender, which pastes your photo back into the center after filling the edges.
- Counting and physics. "Add three cups" can give two or four. Reflections and shadows are usually plausible, not always correct.
- Consent and rights. Only edit photos you have the right to use, and don't create misleading images of real people.
- Prompt length. The tool accepts up to 1,500 characters. Longer is rarely better; clearer is.
When a one-click tool is the better choice
Some edits are so common that a fixed, tested prompt beats writing your own. For those, use the dedicated tools:
- Photo to cartoon, photo to painting (oil or watercolor) and photo to sketch for restyles that keep faces and framing.
- Photo restoration to clear scratches, creases and stains from old prints without colorizing them.
- AI image extender to change the aspect ratio of a photo.
- Remove object from image to erase something you brush over.
For anything else, describe it. Upload a photo on ImageToImage AI, write what should change and what should stay, and compare the result with your original before you download it.