Why your AI portrait looks fake — and the fix
Your AI portraits read as AI in half a second. It's almost never the face — it's the cues you didn't add. Here's the 5-line template that fools the eye.
Scroll your feed right now. You can spot the AI portraits in half a second — and it's almost never the face that gives them away. The giveaway is almost always the same: the cues a real photograph picks up by accident, and the AI didn't think to add.
Highlights
The "AI look" comes from missing imperfections, not bad faces — halation, fabric draping, blister-pack barcodes, paparazzi flash hotspots.
Five cues separate convincing AI portraits from obviously synthetic ones — and they work the same across the figurine trend, the Met Gala "Fashion is Art" red-carpet shots, Y2K throwbacks, and scrapbook collages.
One four-slot prompt template — subject, style block, detail cues, output format — handles every viral AI photo trend on your feed.
Open Instagram, TikTok, or Pinterest this week and the same handful of AI photo trends keep cycling through: tiny boxed action figures of real people, Met Gala "Fashion is Art" red-carpet self-portraits, faded 90s disposable-camera memories, scrapbook collage pages. They feel like dozens of different apps. They are mostly one prompt template with a few words swapped — and the gap between the ones people screenshot and the ones people scroll past is almost never the AI model. It's the cues.
Most "AI-fake" reads collapse on one missing detail. The figurine without a barcode. The red-carpet gown without visible fabric weave. The Y2K throwback without flash overexposure. The chibi sticker without a die-cut white border. Name the right three or four cues and the same model that flopped yesterday ships a screenshot-worthy post today.
Sources: eWeek's 2026 trend rundown, Raxxo's dev.to prompt library, Rahul Ahir's Met Gala 2026 walkthrough.
The following images were generated using Nano Banana 2:
How to make any AI portrait stop reading as AI
Lock the face before anything else. Identity drift is the first reason a portrait reads as "AI of someone vaguely like me." Open every prompt with the identity instruction before you describe a single sequin or sneaker: "Keep the face identical to the reference. Preserve original facial landmarks and bone structure. DO NOT alter eye shape, nose, or jawline." Caps help on Nano Banana 2 and GPT-4o image. Identity-lock first, world-building second.
Name the materials, not the look. "A red dress" gives the model nothing — every default lands on the same shiny polyester. "A deep crimson sculptural gown with a corseted bodice, hand-stitched floral appliqué in ivory and gold, and sheer organza sleeves" gives it something real to build. Same for product: "blister-pack plastic on a printed cardboard backer" beats "toy packaging." Same for film: "35mm with visible grain and halation on highlights" beats "vintage." The materials are the cue.
Add the imperfections that prove it's not AI. A real photograph has accidents the AI won't generate unless you ask. Halation — the soft glow that bleeds off bright spots on real film — sells the analog look more than grain does. A faint price sticker on a blister pack sells the product look more than the figure does. A barcode, a torn paper edge, a die-cut white border, a small flash hotspot — every viral trend has one or two of these tells. Name them by their real-world part name, not their visual effect.
Specify the lighting style — not "good lighting." "Cinematic" means nothing. "Paparazzi flash from front-left with soft fill on the right and warm ambient glow from chandeliers above" is a brief the model can execute. "On-camera flash with cool indoor lighting and slight motion blur" tells the model you mean Y2K, not just retro. Name the source, the direction, the temperature, and the secondary fill. Keep it bright — dark AI portraits collapse at thumbnail size, which is the only size that matters in a feed.
Tell the model what the output is. Most beginners describe the subject and stop. The most underrated cue is the format: "product photograph for a retail listing," "candid 35mm film still," "festival-style one-sheet poster with title at the bottom and laurels," "die-cut vinyl sticker, sticker-sheet layout." Asking for the format unlocks layout cues — typography position, framing crop, aspect ratio — that pure subject prompts never get.
Worked example A: the Met Gala red-carpet portrait
The 2026 Met Gala theme — "Costume Art," dress code "Fashion Is Art" — turned out to be the perfect sandbox for AI fashion edits. The viral ones nail all five cues at once. Here's the prompt for the image below:
Prompt:
Keep my face, skin tone, and identity completely unchanged. Preserve original facial landmarks and bone structure exactly. Editorial fashion portrait of me standing on the Met Gala red carpet, wearing a sculptural couture jumpsuit with structured shoulders and dense crimson and ivory abstract embroidery in flowing wave shapes across the chest, hand-stitched fabric texture visible. Metropolitan Museum atrium behind, grand chandeliers, soft cool blue-white editorial light, paparazzi flash hotspot from front-left, soft fill on the right, visible micro-bokeh of camera flashbulbs in the background. Realistic skin texture and fabric draping, ultra-detailed, naturally photographed.
Identity lock first. Materials named (jumpsuit, structured shoulders, embroidery wave shapes, hand-stitched). Imperfections (paparazzi flash hotspot, camera-flash bokeh). Lighting specified (cool blue-white editorial + front-left flash + soft fill). Output format implied by "editorial fashion portrait." Five cues, one prompt.
Worked example B: the action figure, done right
The figurine trend is the most-shared AI photo style of 2026 — and it lives or dies on the packaging. Most miss because they describe the figure in detail and forget to describe the box.
Prompt:
Product photograph of a vinyl action figure inside clear blister-pack plastic on a printed cardboard backer. The figure is [SUBJECT]: [age, hair, outfit]. The cardboard backer header text reads [TITLE] in bold red and ivory Hasbro-style block letters. A small faded $24.99 price sticker in the lower-right corner, a real-looking barcode and UPC code on the back-corner of the cardboard. Subtle plastic reflections on the blister bubble. Clean light-grey retail shelf, bright clean front studio key light with a soft backlight catching the plastic edge.
Identity-lock the subject's face. Materials (blister pack, cardboard backer, vinyl). Imperfections (faded price sticker, barcode, plastic reflections — none of which the AI ships by default). Lighting (front key + edge backlight). Output format ("product photograph for retail listing"). Same five cues, different trend.
Chasing trends is the easy part — your figurine post lands the day the trend lands, and you ride the wave for a week. The compounding move underneath is identity consistency: an AI avatar of you that holds across the figurine, the red carpet, the Y2K throwback, and whatever next month's trend turns out to be. Lock the face once and you stop re-engineering the identity step on every post.
The bottom line
The viral AI photo of the week is never a new app. It's a known look plus three or four detail words that fool your eye into reading it as real. Halation. Blister pack. White sticker border. Paparazzi flash. Collect those cues and the figurine, the gown, and the disposable-camera throwback all stop reading as AI. The faster move is to build the identity layer once — an avatar that holds across every trend you want to chase. Build your AI avatar in 90 seconds.
FAQs
Q: Which model handles AI portrait identity-lock best right now?
A: For stylized trend photos (figurine, chibi, scrapbook), Nano Banana 2 holds likeness reliably. For high-fashion editorial like the Met Gala look, GPT-4o image handles fabric vocabulary and editorial lighting cues a little better. FLUX.2 is the open-weights option. Generate four versions and keep the one where the face is unmistakably the person.
Q: How do I stop the model from "fixing" my face?
A: Put the identity instruction in caps and put it first. "Keep the face identical to the reference image. DO NOT alter eye shape, nose, or jawline." If a generation drifts, follow up: "In the last image you altered my facial features — regenerate using only the reference photo, with no changes to face, bone structure, or skin tone." Iterate, don't re-prompt from scratch.
Q: Do I need a separate prompt for every trend?
A: No. One template with four slots — subject, style block, detail cues, output format. The style block is the trend ("collectible vinyl figurine," "Met Gala red carpet," "Y2K camcorder"). The detail cues are the 3–4 words that sell the illusion. Same skeleton, swap two phrases, ship the trend.
Q: Why do my AI portraits look dark and murky at thumbnail size?
A: Because the model defaulted to flat ambient light. Specify the key light direction and brightness ("bright front-left key, soft fill, well-exposed") and negative-prompt the darkness ("dark, murky, underexposed, crushed blacks"). On a bright feed, a dark thumbnail dies.
Q: How small can the image be before the trend stops reading?
A: If a stranger can't tell what the post is from a 200-pixel-wide thumbnail, the image is too busy. One readable subject, one clear cue. Crop tighter than feels comfortable — the feed scrolls fast.