The short version: ChatGPT can convert a normal photo into a believable 1980s or 1990s childhood snapshot if you describe the technical details of the era, not just the vibe, and you’re prepared to run it more than once. It’s a fun, fast way to make nostalgic family images, but it fakes film grain rather than simulating it, and the version most people skip past is that you’re uploading identifiable photos of children into a system that isn’t built for archiving them safely.
Why this trend keeps coming back
Every few months a version of this does the rounds again. In spring 2025 it was Studio Ghibli style portraits. By late 2025 it had shifted to “turn this into a 1995 disposable camera photo” and “make my kid look like a school photo from 1988.” The tool changed name a few times but the appeal never did: people want their current photos to look like something pulled out of a shoebox, complete with the date stamp, the orange cast, the slightly blown-out flash.
What’s happening under the hood is that ChatGPT’s image generation (built on the same model family as DALL-E, now folded into GPT Image inside the chat interface) is being asked to apply a style transfer. You give it a photo and a description, and it repaints the image to match. It’s not restoring an old photo. It’s inventing a new one that looks like it came from an old camera.
What “childhood photo style” means, technically
If you just type “make this look old” you’ll get something generic, usually a sepia wash that looks more like a Western film poster than a family album. The trick is to describe the specific artefacts of the decade and the equipment, because that’s what the model latches onto:
- 1970s: Kodachrome warmth, slightly soft focus, wide collars and thick borders, colours that lean orange and green rather than blue
- 1980s: flash photography indoors with visible flash falloff on the background, red-eye, a printed date stamp in orange digital font in the bottom right corner, 4×3 aspect ratio with a white border
- 1990s: disposable camera grain, slightly overexposed flash, matte finish print, sometimes a laser-grid or marbled-cloud studio backdrop if you’re going for a mall photo studio look
- School photo specifically: combed hair, a plain blue or grey mottled backdrop, a slightly stiff pose, and the particular soft-focus lighting that studio photographers used because it hid blemishes
Naming the decade alone gets you maybe 40% of the way there. Naming the equipment and the print artefacts gets you most of the rest.
The step-by-step that works
Here’s the process I use, and it holds up whether you’re doing this in ChatGPT’s app or through the API:
- Upload a clear, well-lit photo of the person, ideally front-facing, similar to how old photos were framed, not a modern candid selfie angle
- Describe the decade and the specific format: “convert this into a 1986 flash photograph taken indoors on 35mm film, with visible grain, a warm orange colour cast, mild red-eye, and a white printed border with an orange digital date stamp reading 07 15 86 in the bottom right corner”
- Specify the background if it matters: “keep the background as a plain mottled blue studio backdrop” or “keep the original background but make it look like it was lit by an on-camera flash”
- Ask it to keep the facial features and proportions accurate to the uploaded photo, because the model will sometimes soften or “beautify” a face during the style transfer, which is the opposite of what an old photo looked like
- Regenerate at least three times. The first result is rarely the best one. Small wording changes, swapping “grainy” for “slightly grainy,” or “flash washout” for “mild flash washout,” change the output more than you’d expect
That’s the whole method. No plugin, no separate app, just specific language and patience.
What I did with my own photos
I tried this with a photo of my grandson from last summer, wanting the 1985 school photo look, laser-grid background and all, because that’s exactly the kind of photo I have of my own children from that era. It took nine attempts to get one that didn’t look like a filter had been slapped over a modern photo. The first few results had the grain sitting on top of the image like a texture layer rather than looking baked into the print, and the laser background came out looking like a screensaver rather than the actual studio backdrop photographers used at Woolworths and similar shops back then.
What finally worked was breaking the prompt into layers instead of one long sentence: first I asked it to change the lighting and colour cast, then in a follow-up message I asked it to add the studio backdrop, then in a third message I asked for the border and date stamp. Doing it in stages, rather than one big instruction, gave a result that held together. The final image fooled two people in my family for about ten seconds before they clocked the slightly-too-smooth skin, which is the one part it never quite gets right.
Where ChatGPT still can’t fake it convincingly
Film grain isn’t texture, it’s chemistry, tiny silver halide crystals distributed unevenly across the frame, and the pattern differs between film stocks. ChatGPT paints something that looks like grain, but it’s applied evenly and smoothly, the way a filter would apply it, not the way actual film grain sits. If you zoom in past about 200%, most AI-generated “old photos” fall apart, the grain looks too uniform and the edges of objects stay too clean, which real film never does.
The date stamps are another tell. Ask it for a specific font and it will often produce something close but not exact, sometimes with the wrong number of digits or a slightly off colour orange. If you’re trying to fool someone who owned a camera from that era, they’ll spot it. If you’re making a fun family image for a group chat, nobody will look that closely, and that’s the honest use case here: this is a novelty, not a forgery tool, and treating it as anything more precise than that will disappoint you.
There’s also a tool mismatch worth naming plainly: if what you want is an accurate recreation of a specific film stock’s colour science, dedicated photo editing presets built from real scanned negatives will do a far better job than a general-purpose chat model guessing at what “1985” looked like. ChatGPT is the easiest tool to reach for, not the best one for photographic accuracy. It wins on convenience, not on fidelity.
The part most people skip past
Here’s the bit that doesn’t get said enough in the how-to posts about this trend. Every time you upload a clear, identifiable photo of a child to generate one of these images, that photo goes through a system whose data handling terms most people haven’t read and wouldn’t agree to if they had. It isn’t about vague AI fears. It’s the specific, boring fact that these platforms retain uploaded images for model improvement unless you’ve gone into settings and opted out, and a lot of parents doing this trend with photos of their kids have never checked that box.
I’m not saying don’t do it. I did it with my own grandson’s photo and I’d do it again. But I cropped out the background that showed our street, I didn’t use his full name in the prompt even though it wouldn’t have mattered to the output, and I checked my data controls in ChatGPT’s settings first, because once an image is uploaded you’re relying on a company’s policy, not your own judgement, for what happens to it next. That’s worth thirty seconds of anyone’s time before they upload a photo of a child who isn’t old enough to consent to anything.
Getting a result you’ll want to keep
A few things that consistently improve the output, based on the dozens of these I’ve run for myself and for clients wanting nostalgic content for their marketing:
- Start with a photo that already has similar lighting to what you’re aiming for, front-lit and evenly exposed, rather than a moody backlit shot, because the model works from what’s there rather than inventing lighting from nothing
- Describe one decade, not a blend, “1980s meets 1990s” confuses the model and you get a muddled result
- If you want it for print, ask for it at a higher resolution from the start rather than upscaling afterwards, since upscaling tends to smooth out the very grain you were trying to add
- Save every version you generate, even the ones that look wrong at first, because sometimes the third attempt has the right grain but the wrong colour, and you can describe what you liked from it in your next prompt
If you’re using this kind of image generation regularly for a business, for social content or nostalgic marketing campaigns, it’s worth understanding the licensing position too. Images generated through ChatGPT are generally usable commercially under OpenAI’s terms, but that’s a separate question from whether the person in the photo consented to being turned into marketing content, which is a conversation worth having before you post rather than after.
Frequently asked questions
Can ChatGPT turn a modern photo into an old childhood photo?
Yes. Upload the photo, describe the decade and the specific print artefacts (grain, colour cast, date stamp, border), and regenerate a few times. It won’t be pixel-accurate to a real film scan, but it’s convincing enough for family use, social posts, and nostalgic content.
Why do my ChatGPT “old photo” results look fake up close?
Because the grain and flash effects are painted on as a style rather than produced by actual film chemistry or optics. Zoom in past 200% and the uniformity gives it away. For casual sharing it won’t matter; for a print you want scrutinised closely, it will.
Is it safe to upload photos of my children to ChatGPT for this?
Check your data settings first and opt out of allowing your images to be used for model training if that matters to you. Crop out identifying details like street signs or school names, and think about whether the child would be comfortable with the image existing before you post it publicly.
What’s the best prompt for a 1990s school photo look?
Something like: “convert this photo into a 1993 school portrait, mottled grey-blue studio backdrop, soft diffused studio lighting, slightly grainy film texture, natural unretouched skin, keep facial proportions accurate to the original photo.” Run it two or three times and pick the best result rather than expecting the first attempt to be right.
