You need an image for a post about the home office tax deduction. The stock sites offer the same smiling people with calculators that are already in a hundred other feeds. There’s no designer, and your Canva templates are getting stale. So you open an image generator, type “image for a post about taxes” and get a glossy robot holding gold coins. Now you know exactly why people scroll past posts like that.
AI image generators in 2026 can do a lot: hold a style, edit your own photos, render short English text. What they can’t do is decide what to show, or whether a generated image belongs in the post at all.
Below: where AI images help and where they hurt, which tool fits which job, a prompt formula with ready examples, the tells that make images look fake and how to fix them, disclosure rules, and a monthly workflow.
Where AI images fit and where they don’t
A simple rule: generate when the image illustrates an idea; don’t when the image makes a promise.
| Good fit | Bad fit |
|---|---|
| Backgrounds and textures behind text or quotes | Your actual product “as is” |
| Metaphors for abstract topics: finance, law, psychology | Dishes from your menu |
| Carousel and article covers | ”Customers” or “team members” who don’t exist |
| Seasonal and holiday visuals | Before/after shots, interiors, finished work |
| Concepts and moodboards before a shoot | Screenshots, reviews, documents |
The right column isn’t about taste. If someone orders the dessert from the picture and gets a different one, that’s a broken promise and a bad review. In ads, it’s also a misleading-advertising risk. And a generated “happy customer” with an invented face will eventually be spotted. In the US, the FTC rule on fake reviews explicitly covers AI-generated testimonials.
The gray zone is your real product in a generated scene: your photo of a jar of honey placed on a wooden table in autumn light. That’s fine as long as the product itself doesn’t change: same colors, label and proportions. Zoom in on the label, because models love to “rewrite” small print.
Tools: which one for which job
| Tool | Strength | Trade-off |
|---|---|---|
| ChatGPT | Follows long instructions, edits in conversation, short text in images | Recognizable house style unless you set your own |
| Gemini | Photo editing: swap a background, keep a product or character consistent across shots | Sometimes “improves” what you asked it not to touch |
| Midjourney | Aesthetics, stylization, textures | Paid plans only, from $10/month |
| Ideogram | Typography, posters, lettering | Many prefer other tools for photorealism |
| Adobe Firefly | Trained on licensed content, safer for commercial use | Output is more cautious and “clean” |
| FLUX | Open models you can run locally or via API | Requires technical setup |
| Canva | Generation built into the templates you already use | Less control than dedicated tools |
For most social media work, one general-purpose tool (ChatGPT or Gemini) plus Canva or Figma for text on top is enough. Midjourney makes sense when visuals are your main differentiator. Read the commercial terms of whatever you pick: Midjourney, for instance, requires companies with over $1M in annual revenue to be on the Pro or Mega plan.
On text: short English headlines now come out clean most of the time, but other alphabets, long phrases and small print still break. The reliable approach is to generate without text and add it in your editor, in your brand font.
The prompt formula: seven parts
“Image for a post about taxes” gets you a robot with coins because the model has nothing to work with except the word “taxes”. Use this formula:
subject + scene + composition + light + style + palette + format, plus what must not appear.
Three examples to adapt:
1. Background for a quote card:
Soft paper texture, light shadows of monstera leaves in the top left corner, lots of empty space in the center for text, soft morning light, palette: beige #F3E9DC and dark green #2F4A3A, minimalist, no people, no text, 4:5 aspect ratio.
2. Metaphor for a home office deduction post:
Cutaway view of a house like a dollhouse, one room is a tidy home office with a neat stack of receipts on the desk, isometric, flat vector illustration, palette: blue #2D5BFF, light gray, orange accent, clean background, no text, 4:5.
3. Your product in a new scene (edit mode, upload your photo):
Keep the jar and the label exactly as they are. Place the jar on a wooden table by a window, warm autumn evening light from the right, blurred apples and a knitted blanket in the background, natural shadow under the jar, photorealistic, shot-on-a-phone look, 9:16.
To keep your feed from becoming a collage of random styles, save 2–3 prompts that worked as templates and change only the subject. Keep colors and fonts in one document, following the rules in our guide to visuals without a designer.
The tells that make AI images look fake
Followers spot generated images faster than you’d think. Check every image against this list:
- Plastic skin: poreless faces, glossy highlights.
- Hands and small objects. Fingers fail less often than a couple of years ago, but rings, mug handles, glasses and zippers still bend and merge.
- Pseudo-text in the background: signs, book spines, laptop screens with “almost letters”.
- Light from nowhere: glowing outlines, shadows in different directions, everything equally sharp.
- Oversaturated color and the familiar teal-and-orange grade.
- Sterile symmetry without a single random detail.
- Background faces melting into masks.
How to fix it:
- Add to the prompt “shot on a phone”, “natural light”, “subtle film grain” to kill the gloss.
- Switch from photorealism to illustration when the topic allows: flat graphics get fewer raised eyebrows.
- Crop or regenerate problem areas. Almost every tool supports inpainting.
- Start from your own photo and change only the background.
- Review on your phone at feed size, not on a big monitor.
Labels and rights
- Instagram, Facebook, Threads. Meta adds an “AI info” label when it detects technical signals of generation in the file. Photorealistic video and audio made with AI must be disclosed by you. For how platforms treat AI content more broadly, see our piece on AI content and search.
- YouTube and TikTok require disclosure of realistic synthetic content. Illustrations and thumbnails usually don’t fall under it.
- Copyright. In the US, images generated without meaningful human creative input aren’t protected by copyright, and other countries are still working it out. In practice, others can reuse your generated image and you’ll have a hard time stopping them.
- Other people’s styles, faces and brands. Don’t generate “in the style of [living artist]”, celebrities or other companies’ logos for commercial posts.
A workflow: 10 images in 40 minutes
- Once a month, go through your content plan and list the posts that need an illustration rather than a photo.
- Drop the topics into your saved prompt templates and generate 3–4 options each.
- Filter with the checklist above; crop or regenerate what fails.
- Add text in your own template in Canva or Figma.
- File them in folders named after the posts so scheduling is quick.
Captions give AI away just as easily; our guide to AI Instagram posts that don’t sound like AI covers that side.
When AI images won’t help: if your product is visual (restaurant, salon, clothing, renovation), people want real photos. If your audience is premium, generated visuals read as cutting corners. And if one image takes more than 30 minutes of fixes, taking a photo is faster.
The image is half the post; the caption is the other half. In Repost21, the AI assistant Luna writes captions from your brand knowledge base, and the finished post with your image goes out on schedule to Instagram, Facebook, Threads, X or Telegram. Start free.