Improve an existing image
The image enhancer takes attachments you already have and produces more polished versions of them. It is a batch tool: select several images and each is processed as its own request.
Open it at Media → AI improve image. The screen is titled AI generate professional photos based on your images.
Before your first run
Set Update existing image references? to No, create a new media library entry while you are learning the tool.
For users with site media-reference permission, the default is Yes, update references and keep the original. Supported references are updated in resumable batches and both images are retained. See Replace the original or create a new attachment.
The form
Choose a predefined style
| Style | Effect |
|---|---|
| Convert into a studio photograph, in a scene | A styled studio treatment with a setting |
| Convert into a clean studio photograph (no background elements) | A plain studio treatment — the default |
| Add a cinematic look | A graded, filmic treatment |
| No preset, only based on your prompt below | Only what your prompt says |
Prompt. Optional with a preset, required with No preset. Up to 3,100 characters, with a counter. This describes the change, not the whole picture: make the background a rich wooden surface is the example the screen gives.
Choose a model. Select Nano Banana 2 (Google Gemini), Nano Banana Pro (Google Gemini), ChatGPT Sunburst (OpenAI) or ChatGPT Flare (OpenAI). Nano Banana 2 is the default. Choose Resolution, then set the aspect ratio under Image settings. OpenAI models also show Quality (OpenAI), defaulting to High. Their resolution labels update with the exact pixel dimensions when you change the aspect ratio. See Image formats and dimensions.
For OpenAI editing, the original and any reference images must be valid PNG, JPEG or WebP files of 2 MB or less each. The selected original is the primary image to edit; references guide the result.
Update existing image references? Two options, covered below.
Select your reference images. A searchable list of your media library. Tick the images to process; a counter reports how many are selected. Each selected image is processed separately and charged separately.
Optional reference images. Up to three additional images that inform the result without being processed. Each must be a valid image attachment of 2 MB or less.
How the reference numbering works
This is the part worth understanding before writing a prompt.
The image being processed is always Image #1. The optional references are Image #2, Image #3 and Image #4, in slot order.
That lets you write prompts that refer to them:
Relight Image #1 using the lighting and mood of Image #2.
Place the product from Image #1 on a surface like the one in Image #3.
Optional references are the way to carry a consistent look across a batch: supply one reference that defines the style, and every processed image is pushed towards it.
Per-image prompts
Once images are selected, you can add an additional prompt for this image, appended to the base prompt for that image only.
Use the base prompt for what applies to the whole batch and per-image prompts for the exceptions. If a per-image prompt makes the combined prompt exceed the character limit, the tool tells you which image is over.
Running it
Use the estimate-first button. The confirmation covers the whole selection, so it scales with how many images you ticked and which profile you chose.
The tool then processes the images and reports progress. When it finishes it may report that some images could not be processed — read that summary rather than assuming every image succeeded.
What happens to metadata
When replacing the original, the new file keeps the original's title, alternative text, caption and description. Accessibility work already done is preserved.
When creating a new entry, the new attachment gets its own metadata and the original is untouched.
Using it well
Batch by treatment, not by convenience. Images that need the same style and the same prompt belong in one run. Mixing product shots and location photography in one batch produces inconsistent results.
Test on one image first. Run a single representative image, look at it full size, then run the batch. One test charge is cheaper than twenty wrong ones.
Be specific about what to keep. Enhancement can alter details you cared about — a label, a texture, a colour. If something must not change, say so in the prompt.
Watch out for products. An enhanced product image must still show the product a customer receives. Improving lighting is fine; improving the product is misrepresentation. See Review generated and edited images.
Cost
One charge per selected image, based on the model, resolution, OpenAI quality, complete prompt and input images. Estimate the complete selection before confirming. See What image operations cost.
Related
- Choose the source image — what makes a good input.
- Replace the original or create a new attachment — the decision that deletes files.
- Troubleshoot image tool errors.