AI images · prompts to copy

AI images — ChatGPT, Gemini & Grok

You don't need Midjourney or complicated tools — today you can create and edit images right in ChatGPT, Gemini or Grok. This page has ready-made prompts from the basics to advanced, what works and how to ask for edits properly. Plus how to bring an image to life in Grok.

🟢ChatGPT

✓ The best at text in images (logos, lettering).

✓ Understands long context and can keep the “style of a scene”.

✗ Stricter filters (celebrities, trademarks, violence).

In ChatGPT, write “Create an image…” or use the 🖼 button.

🔵Gemini (Nano Banana)

✓ Photorealism, light and composition at a pro level.

✓ Fast, handles series of variations.

✓ Takes up to 14 reference images at once.

✗ Every image carries an invisible SynthID watermark, so its AI origin can be detected.

In Gemini, write “Create an image…” or upload an image and describe the edit.

⚫Grok Imagine

✓ Creates an image from a description and edits an uploaded one, takes up to 5 references at once.

✓ Turns a finished image into a short video.

✗ Only in paid SuperGrok (since March 2026).

✗ Don't edit photos of real people; authorities are investigating Grok over deepfakes.

In Grok, open Imagine and write a description, or upload an image and describe the edit.

Want images at home on your own graphics card? You can run models like Flux, SDXL or Qwen-Image in ComfyUI, more in AI video → ComfyUI.

🎨 Create an image

How to write: describe the scene in sentences, not a pile of words. Say what the image is for, and write what you want in it, not what shouldn't be there. Replace the yellow parts in square brackets with your own.
A photorealistic scene GeminiGPTGrokCreating
Create a photorealistic image for [a website hero section].

Scene: [a young woman in a linen shirt typing on a laptop and smiling over a cup of coffee].
Setting: [a small café with a big window, in the morning].
Shot: [a medium shot from a slightly low angle, the woman on the left, free space on the right for a headline].
Light and colours: [soft morning light from the window, warm tones, a blurred background].
Style: a real photograph, natural skin and textures.
Aspect ratio: [16:9]. No text, logos or watermarks.
Illustration / digital art GPTGeminiGrokCreating
Create an illustration for [the header image of an article about programming].

Subject: [a robot planting flowers into a motherboard].
Style: [a flat vector illustration with a subtle grainy texture], one style across the whole image.
Palette: [pastel, mostly blue and mint, one warm accent].
Composition: [simple, the subject in the middle, lots of free space around it].
Aspect ratio: [4:5].
Clean surfaces, no text or signature.
A product photo on white GeminiGPTGrokCreating
Create a product photo for an online shop.

Product: [a brown leather wallet, with the stitching and the leather grain visible].
Background: pure white (#ffffff), just a soft contact shadow under the product.
Light: studio, two softboxes from the front left and from above, no blown highlights.
Angle: 3/4 from above, the product in the middle, free space around it.
Style: a real photograph, sharp details, true-to-life material colours.
Aspect ratio: 1:1. Only the product itself, no props, text or watermarks.

✏ Edit an image (upload it and describe the change)

How to do it: In ChatGPT, Gemini or Grok, just upload the image (the paperclip / image icon) and describe the edit. The AI generates a new version. Tip: write what should change, and separately list what should stay the same. Change one thing at a time. Edit photos of other people only with their consent.
Remove an object from a photo GPTGeminiGrokEdit
Remove only [the car in the background on the left] from the image.

In its place, fill in the background that naturally continues there ([the pavement and the wall of the house]).
Leave everything else as it is: the main person, their face, pose and clothes, the light, shadows, colours and crop.
A transparent background (PNG) GPTEdit
Remove the background and keep only [the product].

The background must be truly transparent; save the image as a PNG with transparency. A checkerboard drawn into the image is not transparency.
Edges: clean, [keep fine details like hair or fur].
Don't change [the product] itself: its shape, colours or details.
Changing the style / atmosphere GeminiGPTGrokEdit
Redraw the image in the style of [watercolour on rough paper].

Keep the composition, the number of people and objects, their poses and expressions.
Palette: [warm earthy tones].
Atmosphere: [calm, early evening].
Faces and hands should stay natural and whole. Add text only if it was in the original image.
Extending the surroundings (outpaint) GeminiGPTGrokEdit
Extend the image [horizontally] to a [16:9] format.

In the new part: [the beach continuing and a blue sea on the right].
Follow the original light, perspective and style.
The main person stays the same size and in the same place.
Don't add new people, lettering or logos.
Portrait retouch (gently) GPTGeminiEdit
Gently retouch this portrait so it looks natural.

- Even out the skin tone; the texture and pores stay visible.
- Remove temporary blemishes (spots, redness). Keep moles and laugh lines.
- Brighten the eyes just slightly.
- Smooth stray hairs at the edges.
Don't change the face, features, expression, background, light or colours.

🔀 Variants & series

4 variants of the same scene GeminiGPTGrok
Create 4 variants of the same scene so I can compare them. Each as a separate image.

Scene: [description of the scene].
The same in all of them: characters, clothes, style and palette.
Only the shot and the light change:
1. From the front, daylight
2. In profile, golden hour
3. From above, overcast
4. Close-up, artificial light
A series for social media GPTGemini
Create a series of 5 images for [Instagram] on the topic [how to start running].

Shared style: [palette, font, atmosphere].
Aspect ratio: [4:5].
Write only the text in quotation marks, nothing else:
1. Intro: a strong visual and at the top the caption “[Start today]” in a large bold sans-serif font.
2. Problem: a scene showing [the tiredness after the first kilometre], no text.
3. Solution: the same character with a positive change, no text.
4. How to: [3 steps as simple icons] and the caption “[3 steps]”.
5. Call to action: at the bottom centre, the caption “[Follow for more tips]”.

🎯 Logos, icons, app icons

A minimalist logo (symbol + name) GPTGemini
Design a minimalist logo for the brand “[Name]” (letter by letter: [N-A-M-E]) in the field of [field].

Brand values: [3 adjectives].
Style: simple geometric shapes and solid surfaces, readable even at the size of a browser tab icon.
Version A: the symbol only, a 1:1 square.
Version B: the symbol with the name “[Name]” next to it in [a modern sans-serif font].
Colours: two at most, [primary and secondary].
Show both versions on a white and on a black background. No gradients, shadows or 3D effects.
App icon (iOS / Android) GPTGeminiGrok
Create an icon for the mobile app [name], which [what the app does].

Motif: one bold symbol in the middle, [an abstract wave as a symbol of music], no text.
Background: a two-colour gradient, [dark blue to purple].
Shape: a 1:1 square with rounded corners.
Contrast: the motif must be recognisable even as a small icon on the phone's home screen.
Mood: [playful / professional / minimalist].

✒️ Vectors and logos: VectorCraft

Why a vector: a logo from ChatGPT, Gemini or Grok is an image made of pixels. For a website, a T-shirt or a billboard you need a vector (SVG) that can be enlarged without losing quality. VectorCraft is a vector graphics program in the style of Illustrator, free and open source. It opens and saves SVG and PDF, and with Image Trace it turns an image into vector curves. An AI agent (like Claude Code) can control it too via an MCP server; the project README describes how to connect it. It's for Windows, macOS and Linux, and as a web version you host yourself. Careful: it's an alpha. It was started on 30 September 2026, version 0.16.0 is from 23 October 2026, and the authors themselves say it's about two thirds of the way to an illustrator's real work.
Let an agent draw a logo in VectorCraft Claude CodeVectorCraft
In VectorCraft via MCP, draw me a logo for the brand “[Name]” in the field of [field].

Style: [simple geometric shapes, solid surfaces, 2 colours at most], readable even as a small icon.
Make 3 variants, each on its own canvas.
Save each as SVG plus a 512 × 512 PNG.
After each variant, show me a preview and wait for which one we'll refine further.
A logo from an image to a vector (Image Trace) Claude CodeVectorCraft
Open the image [logo.png] in VectorCraft and convert it into a vector with Image Trace.
Choose the preset for [a logo with solid colours / a black-and-white drawing].

Then unify the colours to [2 colours: #1a1a2e and #e94560], delete small smudges and smooth the curves.
Don't change the shape or proportions of the logo.
Save the result as SVG and show me a preview next to the original image.

You can do it by hand without an agent too: open the image, run Image Trace, pick a preset and save as SVG.

🧪 More apps from ArtCraft

VectorCraft is one of 13 apps from the ArtCraft team, which also makes an open-source studio for controlled AI images and video. All of them still describe themselves as work in progress or early alpha. Good for trying out, not yet for real work.

  • PhotoCraft — image editing. LightCraft — a photo library and RAW development.
  • FilmCraft — video editing. EffectCraft — motion graphics and effects.
  • FigCraft — interface design. DesignCraft — layout and print.
  • SoundCraft — recording and mixing sound. CADCraft — technical drawing.
  • WordCraft, DeckCraft, GridCraft, PdfCraft — texts, presentations, spreadsheets and PDF.

🧠 Advanced: references, exact text, the same character, data

A product from a photo in a new scene (several references) GPTGeminiGrokEdit
I'm sending two images:
Image 1: my product, [a bottle of olive oil].
Image 2: the setting and style, [a rustic kitchen in late-morning light].

Put the product from image 1 into the setting from image 2 as an advertising photo for [Instagram].
From the product, keep exactly the shape, colours, label and the text on it.
From the scene, take the light, colours and atmosphere. The light falls on the product too and casts a natural shadow.
Aspect ratio: [4:5]. Don't add any other text or logos.
A poster with exact text GPTGeminiCreating
Create a poster for [a concert in a club].

The text exactly like this and nothing else:
- at the top, a big headline “[NIGHT RIDE]”, a bold sans-serif font, white
- below it, smaller: “[12 Dec 2026 · 8 pm]”
- at the bottom: “[club name, city]”
Background: [a city at night in the rain, neon reflections in puddles], dark, so the text stands out.
Aspect ratio: [2:3].
The same character across a series GPTGeminiGrokCreating
This is my character: [attach an image, or describe her: a red-haired girl with round glasses, a green jacket, a yellow backpack].

Create [3] images with her for [a children's book], each separately:
1. [feeding ducks at a pond]
2. [reading under a tree]
3. [riding a bike downhill]

The same in all of them: the face, hairstyle, glasses, clothes, proportions and the style [a soft cartoon illustration].
An infographic from your numbers GPTGeminiCreating
Create a simple infographic for [LinkedIn] from these numbers:

[January 120 orders, February 150, March 210]

Type: [a bar chart] with the title “[Orders in Q1]”.
Every bar has a label with the exact number and the height of the bar matches the number.
Style: [clean corporate, dark blue and one accent colour], lots of free space.
Aspect ratio: [1:1]. Don't add any other numbers or text.
With advanced prompts: number the references and say what each one is for (the subject, the style, the background). When the character starts changing across a series, send the first image again and repeat what should stay the same. Always check lettering, numbers and labels after generating. Gemini does text in images best when you first come up with the texts together and only then ask for the image.

🎬 Bring an image to life in Grok

Grok Imagine makes a short video from an image or from a description, up to 15 seconds according to the xAI documentation. According to articles about Grok, the video has sound too, and the cheapest plan makes shorter clips in a lower resolution. Write like a director: what moves, what stays, how the camera moves, what the light is like and what can be heard. Don't animate photos of real people without their consent. More video tools are on the AI video page.

A video from an image Grok
Bring this image to life as a video [6 s] long.

What moves: [hair and a dress in the wind, clouds slowly drifting].
What stays: the face, clothes and composition exactly as in the image.
Camera: [a slow push-in towards the character].
Light and mood: as in the image, [a warm early evening].
Sound: [wind and distant seagulls, no music].
A short clip from a description (vertical) Grok
Create a video [8 s] long, vertical 9:16, for [Reels].

Shot: [a close-up of a barista's hands pouring milk into an espresso and drawing a heart].
Setting: [a small café in the morning, a blurred street outside the window].
Camera: [static, a slow pull-back at the end].
Light: [soft light from the window, warm tones].
Sound: [hissing steam, the clink of a cup, the quiet hum of the café, no music and no speech].
Style: real video, natural movement, no text.

⚠ What not to write in a prompt (common mistakes)

These are things AI can't deliver, or that confuse it and make the output worse. Take them out of your prompt and the image is instantly a class better.

🚫 General traps — they apply to ChatGPT, Gemini and Grok
  • Lettering without quotation marks — put the text that should be in the image in quotation marks and say where it goes and in what font. Spell out an unusual name letter by letter too. Check the spelling after generating. Better to add a long paragraph of text to the image in an editor (Figma, Canva).
  • “Nice/beautiful/cool” — says nothing. Write why it should be nice (minimalist, soft light, pastel palette…).
  • Negations like “no ugly hands” — the AI hears the word “hands” and adds them. Write what you want: an empty street instead of no cars, a shot from the waist up or hands in pockets.
  • 10 different styles at once — “watercolour + photorealistic + comic + 3D”. The AI can't manage it and makes kitsch. One style, at most 2 compatible ones (watercolour + line art, photography + film grain).
  • Brands, celebrities, specific people — “Tom Cruise”, “the Nike logo”, “Harry Potter”. The AI either refuses or makes a weird imitation. Instead, write a type (“a tall man of 40, short dark hair”), or check the licence first.
  • Exact sizes in pixels — “1920×1080 px”. Image AIs in chat take aspect ratios (1:1, 16:9, 4:5, 9:16), not resolutions. You set the resolution when exporting.
  • Numbers and charts without checking — new models can do infographics too (the prompt above), but check every number and label after generating.
  • Contradicting instructions — “a night scene on a sunny day”, “realistic but cartoon”. Pick one.
💬 Typical for ChatGPT
  • Sexual / explicitly violent scenes — it refuses. There's no point trying to get around it.
  • Someone else's face from a photo — don't put another person's face into an image without their consent.
  • A novel instead of instructions — 1–3 clear sentences are usually enough. For a complex image, split the instructions into parts (scene, character, details, limits) and then edit one change at a time.
🔷 Typical for Gemini
  • References without roles — Gemini takes up to 14 images at once. For each one, say what it's for: “image 1 is the shape, image 2 is the material”.
  • Political figures, elections, propaganda subjects — it refuses more strictly than GPT.
  • Outright porn / explicit erotica — Gemini reacts harshly to this, even to suggestive poses. If you want an art nude, phrase it neutrally (“anatomical study”, “classical nude painting style”).
⚫ Typical for Grok (Imagine)
  • Real people — don't upload or edit photos of people without their consent. Authorities in several countries are investigating Grok over deepfakes.
  • A paid plan — since March 2026 Imagine is only in SuperGrok. You can't create images in the free version.
  • Looser settings — it lets more through than ChatGPT and Gemini. Check its output for business use.
✓ Try this instead
  • Make 3–4 attempts with a slightly different prompt and pick the best.
  • Iterate one change at a time — the first version is a rough draft. In the next message, change one thing (“warmer light, make the building on the left smaller”) and repeat what should stay the same.
  • References — upload an image in the style you want and write “a similar style, but…”.
  • Specific vocabulary — “golden hour” instead of “nice light”, “bokeh” instead of “blurred background”, “a dark blue tweed jacket” instead of “a jacket”.

🧪 The universal recipe for a good prompt

A template that always works:
Create [type] for [purpose]: [who/what and what they're doing] in [setting], [shot + aspect ratio], [light], in the style of [style], palette [colours]. No [text, logos, watermarks].

7 things to always go through:

  1. Purpose — where it's going (website hero, logo, Insta, online shop)
  2. Subject and action — what/who and what they're doing (specifically, not “a nice scene”)
  3. Setting and shot — where, from what angle, how close
  4. Light — daylight, golden hour, neon, studio
  5. Style and palette — photography, watercolour, 3D, flat, and named colours
  6. Aspect ratio — 1:1 / 4:5 / 16:9 / 9:16
  7. What shouldn't be there — write positively what should be there. Explicitly exclude only text, logos and watermarks.
💡 Want more? Midjourney, Flux and other tools are planned for the next version. Tell us via pumpydumpy2visual.com what would be useful and we'll add it.