Prepare suitable image formats and dimensions
Decisions about size and shape are cheapest before generating and most expensive after publishing. This page covers what the tools offer and what to choose.
Aspect ratio
The generator offers ten ratios, defaulting to 16:9:
| Ratio | Typical use |
|---|---|
| 21:9 | Wide hero band |
| 16:9 | Hero, featured image, video-shaped slot |
| 3:2, 4:3 | Conventional landscape photography |
| 1:1 | Cards, thumbnails, social, catalogue grids |
| 5:4 | Slightly tall landscape |
| 4:5, 2:3, 3:4 | Portrait, product on mobile |
| 9:16 | Full-screen mobile, story format |
Choose the ratio the image will actually occupy. Generating 16:9 for a square card means cropping, and a crop taken from the centre often removes the subject.
If the same image must work in two shapes, generate it twice at the two ratios rather than cropping one into the other.
Resolution
Generation and improvement only offer output sizes up to 2560 pixels wide or high.
| Model | Resolution choices | Quality choice |
|---|---|---|
| Nano Banana 2 (Google Gemini) | 1K; 2K for aspect ratios that fit the limit | Managed by the model |
| Nano Banana Pro (Google Gemini) | 1K; 2K for aspect ratios that fit the limit | Managed by the model |
| ChatGPT Sunburst (OpenAI) | 1K, 2K, with exact pixel dimensions | Low, Medium, High, XHigh, Max |
| ChatGPT Flare (OpenAI) | 1K, 2K, with exact pixel dimensions | Low, Medium, High, XHigh, Max |
The default is Nano Banana 2 at 1K. OpenAI quality defaults to High. Higher OpenAI quality costs more independently of resolution. Nano Banana Pro has the same base image cost at 1K and 2K.
The resolution selector shows the actual pixel dimensions. For Google models, 2K exceeds the limit at 9:16, 16:9 and 21:9, so those ratios only offer 1K. Other available Google ratios support 2K. When changing the model or ratio makes your selected size unavailable, the selector switches to 1K. Check the new size and estimate before confirming.
OpenAI uses sizes adapted to the selected ratio while keeping both edges within 2560 pixels. For example:
| Ratio | 1K | 2K |
|---|---|---|
| 16:9 | 1280 × 720 | 2560 × 1440 |
| 9:16 | 720 × 1280 | 1440 × 2560 |
| 1:1 | 1024 × 1024 | 2048 × 2048 |
The screen marks OpenAI resolutions above 3,686,400 pixels as experimental, including square 2K output. This is an area threshold; these choices still meet the 2560-pixel edge limit. Max is an OpenAI quality setting and does not increase the selected dimensions.
Choose the smallest size that covers the intended display. Use one image to assess the selected model and quality before running a batch. Increasing resolution also increases download size and may reach your site's upload limit.
What WordPress does afterwards
Once an image is in the media library it is an ordinary attachment, and WordPress generates the registered thumbnail sizes from it in the usual way.
Two consequences:
- Feed the largest sensible original. Downscaled variants come from it.
- Existing optimisation applies. Any compression or format-conversion plugin you run processes generated images the same as any other upload.
Transparency
Background removal produces a transparent result when you leave the background colour empty, and a flat result when you pick a colour.
Practical implications of transparency:
- File size. A transparent image of a photographic subject is often larger than the JPEG it came from. A page full of them is measurably heavier.
- Downstream conversion. Anything that flattens images later will produce a black or white box where the transparency was. Check your optimisation pipeline before committing to a catalogue of transparent images.
- Email and print. Both frequently flatten. A solid fill is safer there.
For catalogue work where the background is always the same, a solid colour fill avoids all of this. See Remove an image background.
Page weight
Generated images are photographic in character and compress like photographs. On a page with several of them, check the total weight before publishing, particularly on mobile.
The usual measures apply — they are not specific to POM AI:
- serve appropriately sized variants rather than the full-size file;
- keep hero images to the size they display at;
- lazy-load images below the fold;
- let your optimisation layer do its job.
Cropping and editing afterwards
Generated and edited images can be cropped, rotated and scaled with WordPress's own image editor, and with any other tool.
Editing after the fact is free, so a small composition problem is better solved with a crop than with another generation. A subject in the wrong place, however, needs regenerating with a better composition instruction — see Write a useful image prompt.
Before you generate
Answer these first; each one is expensive to change later:
- Where will this image appear, and at what displayed size?
- What shape is that slot?
- Does it need space for text, and on which side?
- Will it also be used somewhere with a different shape?
- Which model, resolution and quality meet the intended display size?
Two minutes on these questions saves repeated generations.
Import limits
Generated image imports use the smaller of the site upload limit and 25 MiB. URL downloads are stopped at that limit, and oversized inline images are rejected. If a provider result is too large, request a smaller image or ask the administrator to review the site upload limit.