AI image generator from image: a new picture from yours

An AI image generator from image takes a picture you already have, reads it alongside a short instruction, and paints a new picture that keeps what you asked it to keep. Give it a photo of your product and write “the same bottle on a wet stone ledge at night, rim light from behind”, and you get that bottle in that scene. Give it a selfie and write “same person, same jacket, standing on a rooftop at golden hour”, and the face comes along. The skill is in the instruction: say what to keep from the picture, then say only what changes. This guide covers how that works, the jobs it does well, how to write the instruction, and how to use several pictures at once.

What an image-to-image generator actually does

A text-only generator starts from nothing and builds a picture from your words. An image-to-image generator starts from your words AND one or more pictures, which it treats as a reference. The current generation of picture models does this by understanding the picture, in roughly the way they understand text: they see “a green glass bottle with a cream label” or “a woman with short grey hair in a denim jacket”, and they can hold on to that while they redraw everything around it.

That is a different thing from the older approach many guides still describe, where you set a “strength” slider and the model blurs and repaints your photo by some percentage. Instruction-based models do not need the slider. You tell them in words what stays and what goes.

What you get back is always a new picture. The model redraws the whole frame, so small details of the original (a logo, the pattern on a fabric, the exact lettering on a label) can drift. Expect to roll a few takes and keep the one that holds.

Six jobs it does well

Job What you give it What you write
New scene for a product One clean product photo “This bottle on a marble counter, morning window light”
Same person, new setting A clear photo of the face “Same woman, same coat, at a night market, neon from the left”
Change the style Any picture “This scene as a watercolour illustration, soft paper texture”
Change one thing The picture to edit “Same photo, but make the sky overcast and remove the car”
Combine pictures Two to six pictures “The man from picture 1 holding the mug from picture 2 in the kitchen from picture 3”
Variations for a set One finished picture “Same character, same outfit, now seen from behind, walking away”

The last two are where image-to-image earns its keep for creators. A carousel, a thumbnail series or the stills for a short video all need the same face, product or place in every frame, and a reference picture holds that far better than any description you could type.

How to make an image from an image, step by step

1. Start with a clean reference

The model can only keep what it can see. A product photo works best on a plain background with nothing else in frame. A face works best looking roughly toward the camera, in even light, at a decent size. If your picture has three things in it and you want one, crop it to the one first. A busy reference gives a busy answer.

2. Say what to keep, in plain words

Name the thing and say what to hold on to: “keep her face, hair and glasses”, “keep the bottle’s shape, colour and label”. Models follow this much more reliably than they guess it. If the reference has something you do NOT want (a background, a second person, a watermark), say so: “ignore the background of the photo”.

3. Write only what changes

Then describe the new picture the way you would describe any shot: where the subject is, what it is doing, the setting, the light, the framing and the style. Keep it to what is different. Re-describing the face in words fights the reference instead of helping it.

A good instruction reads like this:

Keep the woman’s face, short grey hair and denim jacket from the photo. She stands at a rooftop railing at golden hour, city behind her out of focus, warm light from the right, medium shot, shot on a phone.

4. Pick the shape before you generate

Aspect ratio is part of the request. A 9:16 picture for a reel, 1:1 for a feed post, 16:9 for a thumbnail. Choosing it afterwards means cropping away part of what you made.

5. Roll a few takes and compare against the reference

Each run starts from fresh noise, so the same instruction gives different results. Make three or four, put them beside the original, and check the details that matter for your job: the label, the eye colour, the cut of the collar. Keep the best and throw the rest away.

6. Use the best result as the next reference

Once you have a picture you like, it becomes the reference for the next one. This is how a set of pictures stays consistent: each new picture is anchored to one you already approved, and you never go back to describing the subject from scratch. The AI character generator guide takes this further with a character sheet, which is the same idea made repeatable.

Combining several pictures in one

Most current picture models take more than one reference. That opens up a different kind of job: putting together things that were never photographed together. A person and a product. A product and a location. A character and a prop.

Three rules make it work:

  • Number the pictures in the instruction. “The man from the first picture, the mug from the second” is much clearer to the model than “the man and the mug”.
  • Give each picture one job. One picture for the face, one for the product, one for the place. A picture that is supposed to supply two things usually supplies one of them badly.
  • Keep the count low. Two or three references give the model room to follow each one. Six can work, but every extra picture is one more thing it can blend wrongly.

When the manual way is enough

You do not always need a dedicated tool for this.

  • A one-off edit. Many general chat assistants and phone photo apps now accept a picture and an instruction. For a single fun edit or a quick background swap, that is the fastest route.
  • A real retouch. Removing a stray hair, fixing skin, straightening a horizon: a photo editor changes only the pixels you touch, while an image generator redraws the whole frame. For a photo that must stay a true photo, edit it instead.
  • Anything that must be exact. A legal label, a precise logo, a technical diagram. A generator redraws fine lettering and small marks, and they can come back wrong.

Image-to-image starts to pay off when you need a set: the same subject in five scenes, the stills for a short video, a week of posts that look like they belong together.

Doing it in Reel Studio

Reel Studio is built for that set. Its picture models include Nano Banana, Seedream, GPT Image, FLUX and others, and every one of them takes reference pictures, from four to six depending on the model. The way it works:

  1. Upload your picture to the bank, or use any picture you made earlier. Everything you make lands there with its prompt and settings.
  2. Attach it as a reference by dragging it from the bank into the composer, along with any other pictures the shot needs.
  3. Write the instruction, or write a rough line and have the prompt assistant rewrite it for the model you picked. You keep only the edits you like.
  4. Pick the model and the shape. Each model shows only the settings it takes, and the price in credits for those exact settings is on the Generate button before you press it. A take that fails gives its credits back.
  5. Remix a result to load its model, settings, prompt and references back into the composer, change one line, and roll again.

Because the result lands in the same bank, the next step is short. Attach it to the next picture as a reference, or hand it to a video model as the first frame of a clip. The guide to image to video AI picks up from there, and how to make an AI video shows how a set of stills becomes a short film shot by shot. If you are choosing between two of the strongest picture models for reference work, Nano Banana vs Seedream compares them on the same jobs.

A free account starts with 40 credits and no card, enough for a few pictures and a first clip. If you would rather write the instruction somewhere free first, the AI image prompt generator turns a rough idea into a full picture prompt with no account.

Common questions

Why did the face change even though I attached a photo? Usually the instruction re-described the face, or the reference was small, side-on or badly lit. Say “keep the face from the photo”, describe nothing about it, and use a clearer reference.

Can I make the result look less like the original? Yes. Say what to keep more narrowly (“keep only the colour palette and the mood”), and the model will take more freedom with everything else.

Does it work for drawings and illustrations? Yes. A sketch can become a finished illustration or a photo-style picture, and a photo can become a drawing. Say which direction you want in the instruction.

Will the text on my product come out right? Often, with the models that are strong at lettering, but check every take. Short, large text holds best; small print is where it drifts.

← All articles