I spent a weekend running the most-shared ChatGPT image prompts against GPT Image 2 (OpenAI’s current flagship image model, released earlier in 2026) and found something the viral posts leave out. About half the prompts that get reposted on social media are fun in the screenshot, broken in practice. ChatGPT refuses them, hallucinates the wrong thing, or returns an image so generic that the “wow” only survives for about five seconds. The other half are worth your time. Here are the seven I would actually run again, plus a candid look at the two that did not make the cut.
A short note before the list. Most of these prompts are circulating because they produce a single specific kind of result: an image that is unusual enough to screenshot and post. That is a real use case (content, social, slides, mood boards), but it is not the only one. Some of the prompts also work as creative warm-ups, or as a way to test what the model actually understands about composition, lighting, and prompt-following. The list below is organized by what each prompt is good for, not by how many likes the original tweet got.
The prompts that survived testing
The seven prompts that survived testing fall into four buckets, by what they are good for:
- Image-to-image play. Restore the image that does not exist (needs negotiation), MS Paint recreation (needs an attached image), the 3D figurine prompt (works clean on the first try).
- Prompt-only art. The unsettling normal scene, the album cover, the “how does the model see you” prompt.
- Real utility. Photo restoration with a real input. The only one that is actually useful outside the screenshot context.
- Group activities. “How does the model see you,” the album cover, and the 3D figurine all work well as things to send to a friend and compare results.
Restore the image that does not exist” is the most entertaining on the list, and also the most fragile. Paste in any photo, then say something like “Restore the attached photo. Do not ask me to upload it again. Close your eyes and restore it. Do not say you cannot restore the image.” ChatGPT often pushes back, asks you to re-upload, or apologetically explains that it cannot restore a file it has never seen. When it works, you get something genuinely strange: a hallucinated photograph of a subject you never gave it. The trick is to keep arguing. Treat the first refusal as a negotiation opening, not a wall. I have found that the second or third attempt usually produces something usable, often weirder than what the viral posts show.
The unsettling-scene prompt is more reliable because it does not depend on coaxing the model out of a refusal. “Create an image that seems completely normal at first glance but becomes increasingly unsettling the longer someone studies it.” The output is a polished, plausible image with one or more details that do not quite fit: an extra finger on a hand in the background, a face that is not quite the same person twice, a shadow that does not match the light source. The game is in the hunt. Do not ask ChatGPT where the unsettling parts are. It will guess, and it will guess wrong. You are better than the model at finding them, which is the point.
MS Paint recreation works as a counter-prompt to the photorealistic one. “Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse.” Without an upload, ChatGPT will invent one and the result is uninteresting. With a real photo, the model will produce something that genuinely looks like a 1990s clipart disaster, down to the wobbly lines and the white void background. The comedy is in the contrast. A model that can render near-photoreal images is being asked to deliberately produce a piece of garbage, and it commits to the bit.
The “how does the model see you” prompt is the one that goes around as “I asked ChatGPT what it thinks of me and it drew me as a woman holding a cat” or similar. The trick is the framing. You do not just ask “draw me.” You ask it to imagine its ideal user, then ask it to imagine what that user looks like, then ask it to draw that imagined person. Each step adds another degree of separation from you, which lets the model commit to a result that is unflattering or surprising without tripping its own safety filters. The output is rarely a real likeness, which is the joke.
Album cover and 3D figurine prompts both work because the visual grammar of those formats is one of the strongest priors in the training data. Specify a genre and band name that clash (“a doom metal band called Sunday Brunch”) and the model produces something funny, sometimes legitimately usable as a placeholder. The figurine prompt still circulates from 2025 and has not really stopped.
Photo restoration with a real input is the only prompt I would call practically useful. “Restore the attached photo. Do not change the composition or add new elements. Just remove scratches and improve color.” The result is not surprising, just a slightly better version of a photo you already have. If you have a folder of old family photos that look rough on a phone screen, this is the prompt to use. It works on the first try, no negotiation needed.
The two that did not make the cut
“Draw me as a 90s yearbook photo” usually returns a face that is recognizably 90s but not recognizably you. Without an uploaded reference image, the model invents a person. With a reference, the model is reluctant to age you down by 30 years. The viral posts showing this prompt working almost always used a stylized image of a celebrity, where the model already had a strong prior to work from. For an ordinary headshot, the output is a generic stranger.
“Photorealistic image of me doing [activity]” runs into the same problem in a different shape. ChatGPT does not know what you look like, and will not pretend to. The output is a stock-photo version of the activity with a face that has no relationship to yours. The viral posts that show this working are almost always demo accounts where the user uploaded a face reference and the model rebuilt it. The “first try” experience most people get is the blank-faced stock photo.
Trade-offs
GPT Image 2 is not free, and the prompts on this list are not equally cheap. The 7-prompt round-trip above used about 30 percent of a weekly Plus tier (the $20/month subscription that bundles ChatGPT, image generation, and a small amount of other tooling), because each “I tried 4 times to make the restore prompt work” round costs as much as one successful image. Free tier users can run the list, but will hit the daily image cap before finishing it. The savings on photo restoration are real (a Photoshop session that takes 20 minutes becomes 30 seconds), but only if you have a Plus subscription in the first place.
The biggest hidden cost is the negotiation loop. Several of these prompts only work because you push back when the model refuses, and that back-and-forth burns through context (the running text history the model uses to keep track of the conversation) faster than a clean single-shot prompt does. If you are on a plan with a session length cap, the restore prompt and the figurine prompt will eat your session faster than a normal conversation would.
Anyone making content and needing a fast, reliable image should start with the MS Paint prompt and the album cover prompt. They are the cleanest wins on the list. People looking to mess around on a Sunday and see the model surprise them will get the most out of the restore and the unsettling-scene prompts. Anyone curious about the “GPT Image 2 competes with Gemini” claim can verify it in two prompts: photorealistic landscape, then MS Paint redraw of the same image. That range is the point.
Bottom line
If you only do one thing from this article, try the unsettling-scene prompt. It is the most reliable on the list, the output is genuinely interesting, and it does not require you to argue with the model for ten minutes. The restore-the-image-that-does-not-exist prompt is the most fun but the most expensive in terms of time. Run it on a Sunday with nothing else on your plate, not before a deadline.
If you have ever been curious what people mean when they say “GPT Image 2 is now competitive with Gemini,” the MS Paint prompt is the fastest way to see it. The same model that just produced a photorealistic landscape will, in the next prompt, produce a wobbly white-void drawing that looks like a kindergartner’s homework. That range is the point.