AI Image Prompts for Launch Visuals: OG Images, Gallery Shots and Thumbnails

Every launch needs a small stack of images: a logo, a gallery of screenshots, a share image for when someone pastes your link, and a few cards for social. Most builders leave them for the last evening before launch day, and that is when the temptation is strongest to fill the gaps with AI art.

This post is about where AI image prompts actually help a product launch and where they quietly hurt it. It covers what each visual needs (with the real limits of our own launch form, since those are the numbers I can vouch for), why real screenshots still win the gallery, and six prompts for Nano Banana, GPT Image and Midjourney that you can copy and change.

A disclosure first: I also built Prompt Generator, a free tool that writes prompts for a specific AI model, and it launched on Nick Launches this week. It comes up once below, where it is useful. Everything else here works without it.

What each launch visual has to do

Each image has one job, and the job decides the size, the crop and whether AI belongs in it at all.

The logo

On Nick Launches the logo has to be square, 512x512 or larger, in PNG, JPG, WebP or GIF, and up to 2 MB. A logo that is not square gets cropped to a square from the center, so a wide wordmark loses its ends. SVG is not accepted, because an SVG file can carry scripts.

The logo renders small: somewhere between 56 and 128 pixels on the site, and we store it at no more than 512 pixels wide. That size kills fine detail. A mark that reads at 56 pixels is a simple shape with strong contrast. If you only have a wordmark, put the first letter or a symbol on a solid square for the launch and keep the wordmark for your site.

This is the one visual I would not generate. A logo is something you need to own, redraw and use for years, and an image model hands you a raster file you cannot edit cleanly. Use AI to explore directions if you are stuck, then draw the one you pick.

The gallery

The launch form takes up to five screenshots, 16:9 at 1920x1080 or larger, up to 10 MB each. Anything that is not 16:9 is cropped from the center, so a tall phone screenshot loses most of its height. The form says it plainly: put phone screenshots on a 16:9 background instead.

Two more details change how you should plan the gallery. First, the first screenshot is also your share image: when someone posts your Nick Launches page on X or in Slack, that is the picture they see. Second, uploads are re-encoded as WebP and capped at 1600 pixels wide, because nothing on the page renders wider. Big text survives that. Tiny interface labels at the edge of a 4K capture do not.

If you add a demo video, a YouTube or Vimeo link plays inside the gallery as its first slide, and the player shows the video's own thumbnail, which is the last image on the table above. (Making the video is a separate post.)

The share image for your own site

Your homepage needs its own og:image for the moment someone pastes your URL anywhere. 1200x630 is the size to make. Keep the important part away from the edges, because some apps crop the card to a slightly different shape, and keep any text large enough to read in a preview the size of a thumbnail.

Check that the URL in the tag actually resolves. A missing or broken og:image turns every share into a bare link, and nobody tells you.

Social cards and thumbnails

These are the loosest. A launch day post on X or LinkedIn can carry a 16:9 card, a square, or a 4:5 portrait that takes more of the feed. A video thumbnail is 16:9; YouTube asks for 1280x720 or larger. The rule that matters is the same as for the gallery: someone glancing at it should see what the product is.

Real screenshots beat AI art in the gallery

I look at every launch on Nick Launches before it goes live, so I have seen a lot of galleries. The ones that work are honest: the product doing its main job, cropped so the important part is big, with a line of text that says what is happening.

The ones that do not work are generated scenes of glowing dashboards and floating charts that are not the product. A visitor opens the gallery to decide whether to click through, and a picture of an interface that does not exist answers nothing. It also shows. Image models redraw interfaces from scratch, and the text inside them comes out half readable: buttons with invented words, charts with numbers that do not add up.

So the gallery should be your real product. That still leaves a lot of room for AI.

Where AI images earn their place

1. Backgrounds behind screenshots. A 16:9 canvas with a soft gradient, a texture or a desk scene, with your real screenshot placed on top. This is the fix for phone screenshots, and it makes desktop shots look finished.
2. Mockup scenes. An empty laptop on a desk or a phone on a table, with a blank screen you fill with your own screenshot in a design tool. You keep the real UI and get the context.
3. Hero art. The background of your OG image, the top of your landing page, the header of your launch day post. Mood and color, no interface.
4. Thumbnails. A strong background and one subject for a video or a blog post, with the title added by you or by a model that handles text well.
5. Social cards. A launch day announcement card, a "we are live" card, a card for a feature you want to call out.

What the five have in common: the AI draws the frame, and the product inside it is real.

How each model wants the prompt

The three models people reach for at launch read prompts differently, and a prompt written for one often disappoints in another.

Nano Banana (Google's Gemini image model) reads full sentences. Describe the scene the way you would brief a photographer: subject, setting, light, camera. Its strength for launches is editing: you can upload an image you already have and ask for a change, and it keeps the rest. The Nano Banana photo prompt guide↗ (opens in a new tab) goes deeper on the structure.

GPT Image works best with a plain description and the exact words you want in the picture in quotes. It is the one I would start with when the image has text on it, like an OG card with your headline.

Midjourney likes short comma-separated phrases followed by parameters: --ar for the aspect ratio, --style raw for less of its own look, --no for things to leave out. It is strong on mood and texture, and I would not rely on it for exact words.

Six prompts to copy

Replace the parts in brackets, keep the structure. Each prompt leaves room for your real product or your own text.

1. A 16:9 background for phone screenshots (Midjourney)

soft gradient background, warm orange fading into deep violet, subtle film grain, gentle light falloff from the top left, empty center, clean and minimal, no objects, no text --ar 16:9 --style raw --no text, letters, logo, device

Export at 1920x1080 or larger, then place two or three phone screenshots side by side in the middle. Swap the colors for your brand's.

2. An empty laptop scene for a mockup (Nano Banana)

A silver laptop open on a light oak desk, photographed straight on at eye level so the screen is a flat rectangle facing the camera. The screen is evenly lit plain white with nothing on it. Morning daylight from a window on the left, a small potted plant and a ceramic mug at the edge of the frame, soft shadows, shallow depth of field on the background. Landscape 16:9 photo, realistic, no text anywhere.

The straight-on angle matters: it lets you drop your real screenshot onto the white screen in Figma, Canva or Keynote without warping it.

3. An OG image with your headline (GPT Image)

A wide 1200x630 social share image. Left half: the headline "[Your product]: [what it does in five words]" in a bold, clean sans-serif, dark gray on an off-white background, with a smaller line under it reading "[yourdomain.com]". Right half: an abstract shape made of soft overlapping circles in [brand color] and pale yellow. Generous margins on every side, nothing important near the edges, flat design, no other text.

If the model only offers fixed sizes, generate the landscape option and crop to 1200x630. Read every letter before you ship it.

4. A launch day card (GPT Image)

A square social media card announcing a product launch. A large rounded rectangle in [brand color] fills the center with the text "Live today" in white, heavy sans-serif, and below it "[Product name]" in a lighter weight. Background: a subtle paper texture in warm cream. A few small confetti shapes in two colors around the edges. Clean, flat, plenty of empty space, no other words.

Change "square" to "vertical 4:5" if the card is for a feed where portrait takes more room.

5. Hero art for a landing page or a thumbnail (Midjourney)

isometric miniature workshop on a floating island, tiny desk with glowing screen, shelves of small boxes, warm evening light, pastel palette, soft shadows, tilt-shift look, lots of empty sky above for a headline --ar 16:9 --stylize 150 --no text, letters, watermark

The empty sky is deliberate: that is where your title goes. For an exact 1200x630 OG crop, Midjourney also accepts --ar 40:21.

6. Change the background of a screenshot you already made (Nano Banana)

Keep the screenshot in the center exactly as it is, including every word and number on it. Replace only the gray background around it with a soft gradient from pale mint to light sky blue, with a faint shadow under the screenshot so it looks like it sits slightly above the background. Do not crop, resize or redraw the screenshot.

Upload your existing 16:9 image with this one. Check the result closely: if anything inside the screenshot changed, do the swap by hand instead and use the model only for the background.

If you would rather start from one line ("an OG image for a habit tracker, green and calm") than from these templates, Prompt Generator's image prompt generator↗ (opens in a new tab) writes the prompt for whichever of these models you pick, in that model's syntax. It is free for 10 prompts a day without an account.

Putting the images together

The prompts give you the parts. The assembly is quick work in any design tool:

1. Start from the frame. A 1920x1080 artboard for gallery shots, 1200x630 for the OG image.
2. Drop in the AI background, then your real screenshot. Make the screenshot big. Crop to the part that shows the job being done, not the whole window.
3. Add one line of text. "Paste a URL, get a report" says more than a logo. Type it yourself so every word is right.
4. Export and check the crop. Look at each image at the size it will actually appear: a gallery thumbnail, a link preview in a chat app. If the text is not readable there, it is too small.
5. Order the gallery. The strongest image goes first, because it is also the share image.

Before launch day

• Logo: square, 512x512 or larger, readable at 56 pixels.
• Gallery: up to five 16:9 images at 1920x1080 or larger, real product in every one, the best one first.
• Share image: 1200x630 on your own site, and you have pasted your URL into a chat to see the preview.
• Social cards: one for launch day, in the shape of the feed you post to.
• Text in every image read letter by letter.

When the images are ready, the rest of the kit is the copy, and Will Ballentine's post on what happens after launch has the full list. Then pick a launch week.

FAQ

Can I use AI-generated images as my launch screenshots?

Use them around your screenshots, not instead of them. A visitor opens the gallery to see the product, and a generated interface is a picture of something that does not exist, usually with garbled text in it. Real screenshots on an AI-made background give you both: the product as it is, in a frame that looks finished.

What size should a launch OG image be?

1200x630 is the safe size for your own site's og:image; it renders cleanly on the big social and chat apps. On Nick Launches you do not upload a separate one: the first gallery screenshot is the share image, so make that one a 16:9 image at 1920x1080 or larger.

Which image model is best for launch visuals with text on them?

GPT Image is the one I would start with when the words have to be exact, because it follows quoted text closely. Nano Banana is strong at editing an image you already have. Midjourney is the pick for mood and backgrounds, and I would add any text afterwards in a design tool rather than ask it for words.

Do I need a paid plan to write these prompts?

No. The prompts in this post are free to copy. If you want them rewritten for your product, Prompt Generator gives 10 free prompts a day without an account and 20 with a free one.