Aspect ratios for AI images: how to choose and set them

Why the ratio comes first

The aspect ratio is the relationship between an image's width and height, written as two numbers such as 3:2 or 16:9. It is the first decision a photographer makes, because it decides what the frame can hold. In image generation it matters even more: the model composes the whole scene for the frame you give it, so changing the ratio afterwards means cropping or regenerating. Choose it before you refine the rest of the prompt.

Ratio is not resolution

A ratio says nothing about the number of pixels. Midjourney's documentation makes this point directly: aspect ratio is not the same as image dimensions, and the final size depends on the version and upscaler. A 3:2 image can be small or large. Decide the shape from where the image will be seen and how the subject should sit in it, then deal with resolution as a separate question.

Square, 1:1

The square is balanced and static, with no dominant direction. It suits single objects, symmetrical compositions and centred subjects, and it is the frame many social platforms have long used for posts and profile images. It is also the default in Midjourney, whose documentation notes that images start as squares. A square is a poor fit for landscapes and group scenes, where the subject naturally spreads sideways.

Classic photo, 3:2 and 2:3

3:2 is the traditional ratio of still photography, inherited from 35 mm film, and many cameras still use it. It feels natural for a photographic prompt because it is the shape people associate with photographs. Horizontal 3:2 suits streets, interiors and landscapes; vertical 2:3 suits portraits, doorways and tall buildings, and Midjourney's documentation mentions it as common in printed photography and picture frames.

4:3 and 5:4

4:3 was the original television ratio and is still used by many tablets and by most compact digital cameras. It is slightly squarer than 3:2, which leaves more room above and below the subject. 5:4 is squarer again and is common in large and medium format photography and for 8 by 10 inch prints. Both suit interiors and portraits where you want a little breathing room without the strong horizontal pull of a wide format.

Widescreen, 16:9

16:9 is the standard ratio for television and online video; YouTube's help centre describes it as the standard aspect ratio on a computer. Its wide frame suits landscapes, cityscapes, cars in motion and cinematic scenes with a subject placed to one side. It is a strong choice for banners, presentation slides and video thumbnails. It is weaker for single standing figures, which end up small in a wide frame.

Vertical, 9:16

9:16 is the same shape turned on its side: the full screen of a phone held upright. Midjourney's documentation describes it as common for mobile content on social media, and YouTube notes that vertical 9:16 videos can be shown with padding on computer screens. It suits standing figures, tall architecture, waterfalls and anything with a strong vertical line. Compose for it deliberately; a horizontal scene squeezed into 9:16 usually loses its subject to empty sky and floor.

How composition changes with the frame

A wider frame gives more room for setting and for the rule of thirds, with the subject off-centre and space for the eye to travel. A taller frame emphasises height and invites layering from foreground to background. Mention the placement in the prompt: "the car in the right third of the frame, the empty road leading in from the left" reads very differently in 16:9 and in 1:1. When you change the ratio, reread the composition sentences and adjust them.

Setting the ratio in Midjourney

Midjourney's documentation lists the --ar parameter, or its long form --aspect, added at the end of the prompt with a space before the dashes, for example "quiet harbour at dawn --ar 16:9". The numbers must be whole: the documentation says decimals are not accepted and suggests 139:100 instead of 1.39:1, and a pixel size such as 1920:1080 is simplified to 16:9. It also notes that some ratios change slightly when upscaling and that extremely wide or tall ratios are experimental.

Setting the size with OpenAI's image models

OpenAI's image generation documentation describes the size in pixels rather than as a ratio. At the time of writing it lists 1024x1024 as square, 1536x1024 as landscape and 1024x1536 as portrait among the recommended sizes, with an automatic option. For its newer models it also describes custom sizes written as width by height, where both edges must be multiples of 16 and the ratio between the long and short edges must stay within 3:1. Note that 1536x1024 is itself a 3:2 shape.

Stable Diffusion and other tools

In the Hugging Face diffusers library, the Stable Diffusion text-to-image pipeline accepts height and width arguments in pixels. The principle is the same everywhere: decide the shape first, then set whatever the tool asks for, pixels or ratio. When a tool offers only a few sizes, pick the closest one and crop afterwards, leaving margin in the composition so the crop does not cut into the subject.

Plan for more than one format

If an image will be used in several places, say a wide web banner and a vertical story, generating one image and cropping it both ways rarely works. Either compose with generous empty space around a central subject so both crops survive, or generate each format separately from the same prompt with the composition sentences adapted. The second option usually gives better results and costs only a little more time.

A quick way to choose

Ask where the image will be seen and what shape the subject has. A single object or a centred symmetrical scene suits a square. A photograph meant to feel like a photograph suits 3:2 or 2:3. A wide landscape or a video frame suits 16:9. A full-screen phone image or a tall subject suits 9:16. When nothing points clearly one way, 3:2 is a safe photographic starting point that crops well to other shapes.

Where LarpGPT fits

LarpGPT is a library of photo prompts for travel, architecture, cars and lifestyle scenes, with composition written into each prompt alongside subject, lighting and lens. You copy a prompt, adapt its framing to the ratio you need, and set that ratio in the image generator you already use. LarpGPT does not generate images, and results vary between attempts.

Further reading

• Midjourney documentation: Aspect Ratio (archived copy, May 2026)
• OpenAI API documentation: Image generation
• Hugging Face diffusers: Stable Diffusion text-to-image pipeline
• YouTube Help: Video resolution and aspect ratios
• Aspect ratio (image) (Wikipedia)