IMAGE TO IMAGE

Start from a picture you already have

Upload an image or reuse one of your own results, describe what should change, and let the model do only that. Six editing models, multiple reference images, and full control over resolution and format.

Start editing

What image to image does

Image to Image takes a picture you supply and changes it according to a written instruction. It covers style transfer, relighting, wardrobe and set changes, and any edit where the composition should survive but the content should shift.

The single biggest factor in the result is how you write the prompt. An instruction that names only the change preserves the frame. An instruction that re-describes the whole picture gives the model licence to rebuild it.

The image to image form, control by control

Six controls, and the source and prompt fields account for most of the outcome.

Input images section with upload, image URL and previous result options

The input section. Upload a file, paste a URL, or pick a previous result.

Supply the source image

Three routes get an image into the form. Upload a file from disk, paste an image URL, or select one of your own previous completed tasks. That third option is what makes chaining practical, since a Text to Image result can go straight into an edit without a download and re-upload.

Models that accept multiple reference images let you add more than one. Use that when the edit needs to borrow from a second picture, such as applying the look of one frame to the content of another.

Choose an editing model

The image to image roster is not the same as the text to image one. Seedream 5.0 Lite and FLUX.2 Klein 9B appear here as editing models, while MidJourney and Krea do not.

Editing models differ mainly in how much of the original they preserve. Some hold composition tightly and change only what you named, others reinterpret the frame more freely.

Model dropdown listing the image to image editing models in the workspace

The editing model roster, which differs from the text to image list.

The prompt field. Describe the change, not the whole picture.

Describe the change

This is the part people get wrong most often. The prompt should describe the transformation you want, not re-describe the entire image. The source already carries the subject, the composition and the framing.

Write "change the lighting to late afternoon, warm and low from the left" rather than restating the whole scene with new lighting attached. Re-describing everything invites the model to rebuild parts you wanted untouched.

Set the output resolution

Resolution here controls what the edit produces, independent of the source file size. A large source does not automatically give a large result.

For work heading to delivery, generate at the resolution you need or plan a separate upscaling pass. Editing repeatedly at low resolution and enlarging at the end compounds softness.

Resolution options in the image to image form

Resolution options for the generated result.

Output format selector offering PNG and JPEG for the edited image

Output format. PNG for further editing, JPEG for review.

Choose the output format

PNG is the right choice when the result is going into another editing pass, because repeated JPEG compression accumulates artefacts across a chain of edits.

JPEG is fine for review and sharing, where file size matters more than preserving every pixel for the next operation.

Save a face for later use

If an edit produces a face you want to reuse, saving it as an actor or character makes it available to later generations rather than leaving you to find the frame again.

For work that needs the same person across many shots from the start, Character Generation handles that directly from reference photos.

Save actor and save character options in the image to image form

Save actor and save character, carried over from the generation forms.

Models available for image to image

Six editing models. The meaningful difference between them is how much of the original they hold on to.

Model	Best for	Notes
Nano Banana 2 Lite Default	Everyday edits	The default editing model. Fast enough to try several phrasings of an instruction.
Nano Banana Pro Ultra	Final quality edits	The highest quality editing option in the Nano Banana family.
GPT Image 2	Instruction following	Strong where the edit is a precise instruction about a specific part of the frame.
Seedream 5.0 Lite	Style and mood changes	ByteDance editing model, useful for shifting the overall look of a frame.
FLUX.2 Klein 9B	Detail preservation	FLUX editing, and the option to try when other models soften texture you need.
HiDream O1 Image	Stylised reinterpretation	Leans further from the source than the others, which suits illustrative work.

This roster differs from the text to image list. Seedream 5.0 Lite and FLUX.2 Klein 9B appear here as editing models, while MidJourney 8.0 and the Krea models are offered for generating from a prompt alone.

Settings reference

Every control in the image to image form and what it changes.

Source image

The picture being edited. Comes from an upload, a pasted image URL, or any of your previous completed tasks.

Upload, URL or previous result

Additional references

Extra images used alongside the source, on models that accept more than one. Useful for transferring a look from one frame to another.

Model dependent

Prompt

The transformation to apply. Describe only what should change, since the source already supplies everything else.

Up to 2,500 characters on most models

Aspect ratio

The frame shape of the result. Changing it from the source ratio means the model has to invent or discard edge content.

1:1, 16:9, 9:16, 4:3 and others

Resolution

Output size of the edit, set independently of the source file. Generate at delivery resolution or plan a separate upscaling pass.

Selectable in the form

Output format

File type of the result. PNG avoids compression loss across a chain of edits, JPEG produces smaller files.

PNG or JPEG

Writing edit instructions

Treat the prompt as a note to a retoucher rather than a description of the finished picture. "Replace the overcast sky with a clear blue sky and warm the light on the building to match" names the change and its consequence. It also tells the model that the building itself stays.

Compare that with "a building under a clear blue sky with warm light". That phrasing describes a picture rather than an edit, and it gives the model no reason to preserve the building already in frame.

When an edit needs several unrelated changes, run them as separate passes. Each pass has one instruction, and you can stop at whichever result works instead of discarding a combined attempt where one of four changes went wrong.

Model tutorials

Longer coverage of the editing models, with the cases where each one preserves or reinterprets the source.

Nano Banana Pro 2 image-to-image

Instruction phrasing and how much of the source each variant preserves.

Seedream 5.0 Pro

Where the Seedream family sits for editing and style changes.

FLUX.2 image generation guide

Detail handling across the FLUX.2 models, including the Klein editing variant.

Image to image questions

Should the prompt describe the whole image or only the change?

Only the change. The source image already carries the subject, composition and framing, so re-describing all of it invites the model to rebuild parts you wanted left alone. Name the transformation and nothing else.

Can I use a previous generation as the input?

Yes. The input section lets you pick any of your previous completed tasks as the source, so a text to image result can move straight into an edit without downloading and re-uploading it.

Can I use more than one reference image?

On models that support multiple references, yes. This is how you transfer the look of one frame onto the content of another. Models that accept only a single input will show just the one slot.

Why are the models here different from text to image?

Editing and generating from scratch are different tasks, and the rosters reflect that. Seedream 5.0 Lite and FLUX.2 Klein 9B appear as editing models, while MidJourney 8.0 and the Krea models are offered for generation from a prompt alone.

The edit changed parts of the image I wanted to keep. What now?

Two things usually fix it. Tighten the prompt so it names only the change, and try a model that preserves the source more closely, such as FLUX.2 Klein 9B. Reinterpreting the whole frame is expected behaviour on the more stylised models.

Does a large source image give a large result?

No. Output resolution is set in the form independently of the source file. If the result is heading for delivery, generate at the size you need or add an upscaling pass afterwards.

Which output format should I pick between edits?

PNG. Repeated JPEG compression accumulates artefacts across a chain of edits, which becomes visible after several passes. Switch to JPEG only for the final review or share.

Other ways to generate images

Image to Image is one of seven generation types in the image workspace.

Image & Edit

Speech & Voice

Music & Sound Effects