# AI social media graphics: sizes, a brand look and one image in many shapes

> To make social graphics with AI, pick the aspect ratio first, paste the same brand brief into every prompt, and pass your best image back in as a reference to get the other shapes. Gemini, GPT Image and Qwen-Image document their sizes, and each new shape is a new image whose words you check again.

Mindket editors · Published 2026-11-06 · Updated 2026-11-06 · https://mindket.com/learn/ai-design/ai-social-graphics

## In short

- Choose the ratio before the prompt: square, vertical or wide.
- Keep one brand brief and paste it into every prompt.
- Attach the approved image as a reference and change one thing per round.
- Check the logo, the colors and every word again at each new size.

## Pick the shape before you write the prompt

Decide where the image will sit first, because the shape decides the layout. Wikipedia's [aspect ratio page](https://en.wikipedia.org/wiki/Aspect_ratio_%28image%29) says the 1:1 square is often seen on platforms like Instagram, which accepted only 1:1 uploads until August 2015. Vertical video was popularized by Snapchat, Instagram and YouTube. Widescreen is the HDTV standard, and the [Thumbnail page](https://en.wikipedia.org/wiki/Thumbnail) says YouTube recommends 1280x720 for a thumbnail, with a minimum width of 640 pixels.

Each model has its own list of shapes. Gemini makes a square unless your input images or the `aspect_ratio` setting say otherwise. Nano Banana 2 and Pro add 1:4, 4:1, 1:8 and 8:1, which the cookbook calls ultrawide, panoramic and extreme panoramic. Qwen-Image's [model card](https://huggingface.co/Qwen/Qwen-Image) lists seven sizes. OpenAI's [generation guide](https://developers.openai.com/api/docs/guides/image-generation) takes a custom size: both sides a multiple of 16, a ratio from 1:3 to 3:1 and no edge over 3840 pixels. Its slide example uses 1536x864.

Platforms change their pixel sizes and file limits. Read the help page of the one you post to before you export. The table sticks to what the image makers and Wikipedia document.

| Ratio | Where it fits, per Wikipedia | Gemini, 1K output | Qwen-Image model card |
| --- | --- | --- | --- |
| 1:1 | Square images, often seen on platforms like Instagram | 1024x1024, the default | 1328x1328 |
| 9:16 | Vertical video, popularized by Snapchat, Instagram and YouTube | 768x1344 | 928x1664 |
| 16:9 | The HDTV standard, and the shape YouTube recommends for thumbnails | 1344x768 | 1664x928 |
| 3:2 | Still photography | 1248x832 | 1584x1056 |
| 21:9 | Ultrawide, popular for gaming and desktop monitors | 1536x672 | Not in its table |

## Why the second image stops matching the first

It is a known limit, and OpenAI lists it. Its [generation guide](https://developers.openai.com/api/docs/guides/image-generation) says the model "may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations."

The [prompting guide](https://developers.openai.com/api/docs/guides/image-prompting) answers with repetition. Its children's book example first defines a character: a green hooded tunic, brown boots, a kind expression. A second prompt continues the story with the same character, lists the tunic, the face and the palette again, and ends with a ban on redesigning it. A brand works the same way, with your colors, type and logo rule in place of the tunic.

The prompt below is OpenAI's own, from its "Keep a character consistent" section.

> Pass the previous output as the next edit input, request one change, and repeat the details to preserve.
>
> OpenAI, image prompting guide, [OpenAI API docs: Image prompting](https://developers.openai.com/api/docs/guides/image-prompting) (read 2026-11-06)

````text
Continue the children's book story using the same character.

Scene:
The same young forest hero is gently helping a frightened squirrel
out of a fallen tree after a winter storm.
The character kneels beside the squirrel, offering reassurance.

Character Consistency:
- Same green hooded tunic
- Same facial features, proportions, and color palette
- Same gentle, heroic personality

Style:
Children's book watercolor illustration,
soft lighting, snowy forest environment,
warm and comforting mood.

Constraints:
- Do not redesign the character
- No text
- No watermarks
````

- The **Scene** changes with every image, as a new post will.
- The **Character Consistency** list is what you copy word for word each time. For a brand it holds the colors, the type and the logo.
- The **Constraints** repeat the exclusions, so a no-text rule does not quietly lapse in the third image.

## A brand brief to paste into every prompt

Write the brief once. Put it ahead of each post's own details. OpenAI's fundamentals say to organize a complex prompt as scene, subject, details and constraints in labeled sections, and to name colors, materials, lighting and the visual medium.

The brief below is ours, built on those two rules. Field & Flour is OpenAI's example bakery from [lesson 1](https://mindket.com/learn/ai-design/ai-logo-design). The colors, type and layout are placeholders for yours.

````text
BRAND BRIEF (paste at the top of every prompt)
Brand: Field & Flour, a local bakery.
Look: warm, simple, flat shapes, no gradients, plenty of empty space.
Colors: cream background (#F6EFE3), wheat gold (#D9A441), dark brown text (#3B2A20).
Type: bold rounded sans-serif headlines, plain sans-serif for small text.
Logo: use image 1 exactly as it is, once, in the bottom-right corner. Do not redraw it.
Constraints: only the words in quotes below. No other text, no watermarks, no extra logos.

THIS POST (change every time)
Format: 1:1 square.
Text: "Rye bread, Saturdays only"
Scene: a dark rye loaf on a wooden board in soft morning light, with empty space on the left for the text.
````

- The **Logo** line gives the attached file a number and a role. OpenAI's fundamentals say to identify each input by number and purpose.
- The **Constraints** line borrows the exclusions of OpenAI's Thread ad prompt: no extra text, no watermarks, no unrelated logos.
- Keep **Colors** to a short list, and compare what comes back with your own swatches. A model can drift from a hex code.
- The **Text** line is the quoted copy from [the text lesson](https://mindket.com/learn/ai-design/text-in-ai-images), so each post has its words in a note first.

## One master image, many sizes

Make the master in the shape with the most text or detail, then ask for the others from it. Gemini's cookbook shows the pattern in a chat. Each call after the first passes `previous_interaction_id`, and the last one adds an `aspect_ratio` setting to bring the scene back as a wide shot. The call below is from its chat-mode example, where the earlier turns made a toy fox figurine.

In a chat app, ask for the new shape in the same conversation and restate what must stay, because a detail you do not repeat is a detail the next image can drop. [Edit images with AI](https://mindket.com/learn/ai-images-video/edit-images-with-ai) has the app steps for sending a picture back with one change.

````python
turn5 = client.interactions.create(
    model=MODEL_ID,
    input="Bring it back to the bedroom in an ultra-wide panoramic shot",
    previous_interaction_id=turn4.id,
    generation_config={"image_generation_config": {"aspect_ratio": "16:9"}},
    response_modalities=["text", "image"],
)
````

1. Pick the master: the shape with the most text or detail.
2. Attach the master and the logo, and give each a number and a role.
3. Name the new ratio in the prompt, and in the size or ratio setting if the app has one.
4. Say what stays (logo, colors, wording) and what changes (the layout for the new shape).
5. Move the words to where the new shape has room, and describe that position in the prompt.
6. Open the result at full size and check it against the master. Then repeat for the next shape.

> If a logo must be exact, place the original file over the result in a design app. OpenAI's guide gives the same advice for any part of an image that must stay pixel-identical.

## A looping GIF for a post

OpenAI's GPT Image 2.6 models, Flare and Sunburst, can return an animated GIF, and the guide's lantern example uses `gpt-image-2.6-sunburst`. Set `output_format` to `gif` on the generation or edit endpoint, and send still reference images to the edit endpoint to guide the animation, for example to move a mascot you already have. A [community post](https://community.openai.com/t/images-upgraded-introducing-chatgpt-images-2-6/1412901) says ChatGPT Images 2.6 turns a prompt or an image into a GIF with a speed you can adjust.

For a smooth loop, the prompting guide says to give the subject a starting pose to return to and to say whether the camera should move. Its lantern example, below, does both.

**The lantern GIF in OpenAI's guide**

- **9** frames, which is the default (Source: [OpenAI API docs: Image generation](https://developers.openai.com/api/docs/guides/image-generation), read 2026-11-06)
- **200 ms** per frame, which is the default (Source: [OpenAI API docs: Image generation](https://developers.openai.com/api/docs/guides/image-generation), read 2026-11-06)
- **1.8 s** for one full loop at those settings (Source: [OpenAI API docs: Image generation](https://developers.openai.com/api/docs/guides/image-generation), read 2026-11-06)

````text
A little monkey in an olive-green hooded coat stands in a snowy pine forest at dusk. A wooden cabin glows in the background. The monkey holds a glowing amber paper lantern at chest height, slowly lifts it, looks up and blinks, then lowers it back to its starting pose. Keep the lantern in its hands and the camera still. Make the motion loop smoothly.
````

- The `size` setting is the size of each frame. When a request goes over the model's limits, the guide says to reduce `frame_count`, the frame dimensions or `n`.
- GIF output does not support streaming, partial images, custom `output_compression` or `prompt_rewriter`.
- Decode the `b64_json` value in the response as base64 and save the bytes with a `.gif` extension.

## Common mistakes

- Choosing the shape last. A layout built for a square rarely survives a tall crop with its words in the right place.
- Typing "same style as before" and nothing else, although OpenAI's guide says the phrase carries some context and the critical constraints still need restating once the result drifts.
- Changing the colors, the text and the layout in one round, then not knowing which change broke the look.
- Treating the generated colors as your brand colors. Compare them with your swatches.
- Pushing key words to an image's edge, where a crop or an app's own labels can cover them.
- Trusting the first export. Open each size at full size and read every word against your note.

## Try it yourself

Use a real or pretend small business. Keep everything in one chat, so the master stays in reach.

- Write the brand brief with two colors, a type style and the logo rule. Keep the logo file from lesson 1 next to it.
- Make a square post with one short quoted line.
- Attach that post and the logo, and ask for a tall story and a wide banner, restating what must stay.
- Put the three images side by side at full size and mark each difference in logo, color and spelling.
- Optional: make a looping GIF of the loaf or the logo with OpenAI's pattern, a starting pose and a still camera.

> You are done when the three shapes read as one brand and every word matches your note.

## Next: put a product in a scene

A social image often shows a product on a shelf or an app on a phone. [The next lesson](https://mindket.com/learn/ai-design/ai-mockups) puts a product or a screen into a scene with image editing models.

## Check yourself

1. You want a vertical story image. Which ratio do you ask for?

   A. 16:9
   B. 1:1
   C. 9:16

   Answer: C. Wikipedia names 9:16 as the vertical video shape that Snapchat, Instagram and YouTube popularized.

2. What does OpenAI's generation guide say about a custom image size?

   A. Any pixel count works
   B. Only three fixed sizes exist
   C. The ratio must be exactly 16:9
   D. Both sides must be multiples of 16, with a ratio from 1:3 to 3:1

   Answer: D. The guide asks for multiples of 16 on both sides and a ratio between 1:3 and 3:1, with no edge over 3840 pixels.

3. Your second image drifts from the look of the first. What does OpenAI's guide suggest?

   A. Restate the details to preserve and request one change
   B. Ask for "same style as before" and nothing more
   C. Switch to a different model each time

   Answer: A. The guide says to request one change and repeat the details to preserve, and to restate critical constraints when the result drifts.

4. A GIF is made with frame_count 9 and frame_duration_ms 200. How long is one loop?

   A. 0.9 seconds
   B. 1.8 seconds
   C. 9 seconds

   Answer: B. Nine frames of 200 milliseconds make 1,800 milliseconds, which OpenAI's guide gives as 1.8 seconds.

## Frequently asked questions

### What size should AI social media graphics be?

The ratio is the shape and the pixel size is a setting, so decide the shape first and the export size second. Take the export size from the platform's help page, since platforms change it. If a model's largest output is smaller, enlarge the result in an image editor and check the text again.

### How do I make the same character appear in every post?

Attach a reference image of the character each time and repeat its defining details in the prompt. According to OpenAI's community post, ChatGPT can save reference images of yourself, pets and characters, so you do not upload them again. In the API, each edit call takes its reference images as input.

### Which AI image generator has the most size options?

On the pages we read, Gemini lists 14 ratios for Nano Banana 2 and Pro, Qwen-Image's model card shows seven sizes, and OpenAI accepts any size that follows its size rule. More options do not mean better results. Run your brand brief through two of them and compare the logo and the colors.

### How long can an AI-made GIF be?

OpenAI's settings allow at most sixteen frames of up to 2000 milliseconds each, which is 32 seconds on paper. The frame size, the frame count and the number of animations also have to fit the model's pixel budget, so a bigger frame may mean fewer frames.

## Sources

- [OpenAI API docs: Image prompting](https://developers.openai.com/api/docs/guides/image-prompting) (read 2026-11-06)
- [OpenAI API docs: Image generation](https://developers.openai.com/api/docs/guides/image-generation) (read 2026-11-06)
- [OpenAI Developer Community: Images, upgraded, introducing ChatGPT Images 2.6](https://community.openai.com/t/images-upgraded-introducing-chatgpt-images-2-6/1412901) (read 2026-11-06)
- [Google Gemini cookbook: Get started with Nano Banana](https://github.com/google-gemini/cookbook/blob/main/quickstarts/Get_Started_Nano_Banana.ipynb) (read 2026-11-06)
- [Hugging Face: Qwen/Qwen-Image model card](https://huggingface.co/Qwen/Qwen-Image) (read 2026-11-06)
- [Wikipedia: Aspect ratio (image)](https://en.wikipedia.org/wiki/Aspect_ratio_%28image%29) (read 2026-11-06)
- [Wikipedia: Thumbnail](https://en.wikipedia.org/wiki/Thumbnail) (read 2026-11-06)

[Browse skills](https://mindket.com/skills)