Image-to-Image AI: Working From Pictures, Not Just Words
How to give reference images a clear job, write edit instructions that change only what you intend, and keep identity and geometry stable across images.
4 min read · Updated September 2026
What image-to-image means
Image-to-image covers any workflow where you give the model a picture as well as words. The picture can play different roles: a composition to follow, a subject to keep, a style to borrow, or an existing image to change. Being clear about which role your image plays is the most important factor in getting a good result.
There are two broad approaches:
- Generating with references: you supply one or more images as guidance and describe a new image. The output is a fresh generation informed by the references.
- Editing by instruction: you supply an image and describe a change. The goal is to keep most of the image and alter a specific part.
Using reference images well
A reference image communicates things that are hard to put into words: the exact shape of a product, a character's face, a room layout, a color palette. Several models can take multiple references at once. In OpenDream AI, Nano Banana 2 accepts up to 14 reference images, Seedream 5 Pro up to 10 and FLUX.2 up to 8.
More references are not automatically better. Each one adds information the model has to reconcile. Good practice:
- Choose clean references. A well-lit, uncluttered image of the subject beats a busy snapshot.
- Give each reference a job. One for the character, one for the setting, one for the palette.
- Say what each reference is for. "Use the first image for the character's face and outfit, the second for the lighting and color palette."
- Avoid conflicts. Two references with opposite lighting will pull the result in two directions.
Using the first image as the character and the second as the environment, show her sitting on the stone steps of the courtyard at sunset, keeping her outfit and hairstyle exactly the same, warm low sun from the left
Editing by instruction
Instruction-based editing works like briefing a retoucher. You describe the change in plain language and, ideally, what should stay the same.
Effective edit instructions are:
- Specific about the target: "the blue mug on the right" rather than "the mug".
- Specific about the result: "replace it with a clear glass of water" rather than "change it".
- Explicit about preservation: "keep the lighting, framing and everything else unchanged".
- Singular: one change per instruction produces cleaner results than a list of five.
Change the background to a softly blurred botanical greenhouse, keep the person, pose, clothing and the lighting on the face unchanged
Typical edits include replacing or removing an object, changing the background, restyling the whole image, changing clothing, adjusting lighting, cleaning up a product photo and creating a variation.
Keeping identity consistent
Keeping a person or character recognizable across images is one of the most requested tasks and one of the most delicate. Techniques that help:
- Start from a strong reference. A clear front or three-quarter view with neutral lighting works best.
- Describe the defining features in words too. "Round glasses, freckles, short curly auburn hair" reinforces what the reference shows.
- Change fewer things at once. A new pose, a new outfit and a new setting in one step is harder than doing them in sequence.
- Reuse the best output as the next reference. Once you have a good new image, it can anchor the following one.
Only use images of people you have the right to use, and never use image-to-image to place real people in situations they have not agreed to.
Keeping geometry consistent
Products, buildings and vehicles need their shapes preserved. Models may subtly change proportions, add details or simplify logos and labels.
- Specify what must not change: "keep the bottle shape, cap and label position exactly".
- Change the environment, not the object: backgrounds, lighting and props are easier to alter than the object itself.
- Compare at full size: check edges, handles, seams and small print before using the result.
What works well
- Swapping or cleaning up backgrounds.
- Changing the time of day or the lighting mood.
- Restyling a photo as an illustration.
- Removing distracting objects.
- Changing clothing or the color of an item.
- Producing variations of a composition you already like.
- Placing a consistent character into new scenes.
What tends to be difficult
- Small text and logos: they may be simplified or reshaped. Recreate critical marks separately if accuracy matters.
- Exact counts and positions: "move the third chair two feet left" is less reliable than "remove the chair on the far left".
- Large structural changes with everything else identical: turning a subject to a new viewpoint changes much of the image, so some drift is expected.
- Hands interacting with objects: check grips and fingers carefully.
- Many simultaneous edits: split them into separate steps.
A reliable image-to-image workflow
- Decide the role of each input image: subject, setting, style or edit target.
- Choose a model that supports the number of references you need.
- Write a prompt or instruction that names what changes and what stays.
- Generate, then compare the result with the original at full size.
- Make further changes one at a time, building from the best result.
- Keep earlier versions so you can step back if a change goes wrong.
You can try reference-based generation with the image-to-image tool, and targeted changes in the editor, where each edit is saved as a new version rather than overwriting the original.
Keep learning
Put it into practice
Try what you just read in the Create Studio.
Start creating