Halloween AI Photo Prompts: 8 Viral Looks Tested on 3 Models

Every October the same thing happens: a handful of Halloween AI photo prompts go around TikTok and Instagram, everyone pastes them into Gemini or ChatGPT with a selfie, and half the results come back looking like a stranger in your clothes. This year's favorites are the Y2K flash photo, the VHS found-footage still, the masked figure standing behind you, and the usual vampires and witches.

The lists of prompts are everywhere. What nobody shows is what happens when you actually run them. So we took eight of this year's popular looks, wrote each one as a prompt you can copy, and ran all eight on three image models with the same selfie: two models from Google's Nano Banana family (the image models behind Gemini) and OpenAI's GPT Image 2.5 (the model behind ChatGPT Images 2.5).

Below are the eight prompts, the side-by-side results, which model keeps your face best, and how to turn a whole group of friends into Halloween photos in one go.

How we tested#

• One selfie, made for the test. The woman in every picture is fictional: we generated her selfie with an AI model rather than use a real person's photo. It is a plain, casual phone shot, the kind people actually upload.
• Eight looks, eight prompts. Each prompt describes the scene, the costume, the lighting and the camera style, and ends with the same identity line: keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle.
• Three models, as available in October 2026: Nano Banana Pro and Nano Banana 2.1 (Google), and GPT Image 2.5 (OpenAI, the Flare variant).
• Same settings everywhere: the selfie as the reference image, 2K resolution, 3:4 portrait, one attempt per prompt and no cherry-picking. All 24 images rendered on the first try.

One run per prompt is a small sample, so read the results as tendencies rather than rules. The patterns below did repeat across several of the eight looks.

Results at a glance#

The short version:

• To look like yourself: Nano Banana Pro, then Nano Banana 2.1, then GPT Image 2.5.
• For the most convincing costume and scene: Nano Banana 2.1.
• For the most "real film" look: GPT Image 2.5, if you don't mind a few extra years on your face.

The 8 Halloween AI photo prompts#

Each prompt below is exactly what the models received. Replace the costume details freely; keep the last two sentences.

1. Y2K flash photo#

Y2K flash photo: turn the person from the reference photo into a late-1990s disposable-camera Halloween snapshot. They sit on a messy bed in a dark teenage bedroom at night, wearing a simple homemade costume with cat ears, a string of orange lights and a carved pumpkin on the nightstand. Harsh direct on-camera flash, strong shadow on the wall behind, film grain, slightly overexposed skin. Keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle. No text, no logos.

The look of the season, and the one that showed the biggest differences. Nano Banana Pro kept her gray t-shirt from the selfie and just added cat ears. Nano Banana 2.1 went all in, with a black outfit and drawn-on whiskers, but the words "teenage bedroom" made her look several years younger. It also covered the walls with posters full of made-up lettering, despite "no text". GPT Image 2.5 nailed the harsh flash and the grain.

2. VHS found footage#

VHS found footage: turn the person from the reference photo into a still frame from a grainy 1990s camcorder home video on Halloween night. They stand in a dim hallway holding a flashlight, looking back over their shoulder, while a tall shadowy figure stands half-hidden in a doorway at the far end. VHS scan lines, color bleeding, soft focus, low-light noise, eerie greenish tint. Keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle. No text, no timestamps, no logos.

GPT Image 2.5 made the most convincing camcorder frame: scan lines, color bleeding, the green cast. But she is smiling, which kills the horror; the prompt never said "scared", and only GPT Image 2.5 needed telling. Both Nano Banana models gave her a startled look and a clearer figure in the doorway. Nano Banana Pro kept the gray t-shirt again.

3. Masked figure behind you#

Masked figure behind: recreate the reference photo as a phone selfie taken in a dark living room at night, the person smiling at the camera and unaware, while a figure in a plain white featureless mask and a black hooded robe stands silently behind their shoulder, partly lit by the phone screen. Realistic smartphone night photo, slight motion blur, cinematic horror mood. Keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle. No text, no logos.

All three got this one. The two Nano Banana models copied the phone from the original selfie into the frame; GPT Image 2.5 framed it as a front-camera shot. We described the mask as "plain white and featureless" instead of naming a movie character. Several prompt guides warn that named characters get refused more often, and a description gives you the same chill.

4. Gothic vampire#

Gothic vampire: turn the person from the reference photo into an elegant gothic vampire portrait at a candlelit gathering. Dark velvet and lace outfit, high lace collar, pale skin, deep red lips, subtle fangs, dark wood-paneled room with candelabras behind, warm moody candlelight, shallow depth of field, editorial photo. Keep the person exactly as they are: same face, same features, same face shape, same age, same hairstyle color. No text, no logos.

Every model added the fangs and the red lips. None of them made her skin pale, even though the prompt asked for it. Our guess is that "same skin tone" in the identity line won the argument. If you want the full undead look, try describing it as makeup, e.g. "very pale, almost white skin as part of the costume makeup" (we did not test that variation). Nano Banana Pro also brightened her hair to a vivid copper.

5. Classic witch#

Classic witch: turn the person from the reference photo into a classic Halloween witch portrait. Tall black wide-brim pointed hat, black lace dress with sheer sleeves, a small black cat on their shoulder, standing in a dim cottage room full of candles, dried herbs and old books, soft warm light, shallow depth of field, editorial photo. Keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle. No text, no logos.

The easiest look to get right: all three held the face well, and the only real difference is the cat. GPT Image 2.5 and Nano Banana 2.1 put it on or beside her shoulder as asked; Nano Banana Pro put it in her arms.

6. 1950s vintage party#

1950s vintage: turn the person from the reference photo into an authentic 1950s black-and-white Halloween party photograph. They wear a homemade vintage costume with a paper mask pushed up onto their forehead, standing in a 1950s living room decorated with paper bats and a jack-o'-lantern. Period-correct film photo, soft grain, slight vignette. Keep the person exactly as they are: same face, same features, same age, same hairstyle. No text, no logos.

Nano Banana 2.1 produced the most believable period photo, down to the wallpaper and other guests. Nano Banana Pro added a witch hat nobody asked for, curled her hair and gave her a big toothy grin. GPT Image 2.5 dressed her as a clown and, again, made her look older.

7. Skeleton face paint#

Skeleton face paint: turn the person from the reference photo into a close-up portrait with detailed black-and-white skeleton face paint covering the face, painted teeth over the lips, dark eye sockets, wearing a black hoodie, against a dark background with soft rim light and a few floating embers. Professional studio photo. Keep the person exactly as they are underneath the paint: same face shape, same eyes, same features, same age, same hairstyle. No text, no logos.

The best-looking results of the test from all three models. Face paint is also the most forgiving look: with the face covered, small changes to the features don't show. If you are making these for a group with one model, this is the safest look to include.

8. Pumpkin patch#

Pumpkin patch: turn the person from the reference photo into a warm autumn photo in a pumpkin patch at golden hour. They wear a cozy knit sweater and hold a large orange pumpkin, hay bales and rows of pumpkins behind, soft backlight, natural candid smile, professional lifestyle photo. Keep the person exactly as they are: same face, same features, same skin tone, same age, same hairstyle. No text, no logos.

The wholesome option, and close to a tie. Simple, well-lit scenes like this are where every model keeps the face best.

Faces up close#

Thumbnails hide most of the drift. Up close:

• Nano Banana Pro reads as the same woman in nearly every look: same eyes, same jaw, same freckles. It sometimes changes her expression (the big 1950s grin) but not her face.
• Nano Banana 2.1 is almost as close, except where the prompt pushed on age. In the Y2K photo she looks like a teenager, because the prompt put her in a "teenage bedroom".
• GPT Image 2.5 keeps the likeness but adds texture: deeper smile lines, more pronounced skin, a few more years. It was consistent across looks, so it is a style of the model, not bad luck.

Tips for writing your own Halloween prompts#

The first tip is what we saw in the test; the rest are fixes for the problems above. We did not re-run them, so try one image before a batch.

• End every prompt with an identity line. "Same face, same features, same skin tone, same age, same hairstyle" did most of the work in this test.
• Watch for age words in the scene. "Teenage bedroom", "school dance" or "grandma's attic" can leak into the person. Write "a messy bedroom with posters" instead.
• Ask for the outfit change in so many words. Nano Banana Pro kept the selfie's t-shirt twice. "Wearing a black cat costume instead of their own clothes" leaves less room.
• Describe, don't name. "A figure in a plain white mask and black hooded robe" works; famous horror characters are more likely to be refused.
• Say the emotion. If the scene is scary, write "looking frightened", or GPT Image 2.5 may give you a smile.
• "No text" is a request, not a guarantee. Posters and signs in the background can still come back with gibberish. Check before you post.

Make the whole group's Halloween photos at once#

In a chat app, every friend and every look is a separate round of upload, paste, wait and save: five friends and four looks is twenty rounds. With a bulk image generator it is one batch. Here are both of our fictional test subjects in four looks, made in a single run:

In BulkImagen:

1. Start a New batch and paste the prompts you want into the prompt box, one prompt per line.
2. Upload everyone's selfies as reference images and choose Per image under How to use the references. Every selfie is paired with every prompt, each person rendered separately.
3. Pick the model. Going by this test, Nano Banana 2.1 is the best default for groups: it follows the costume, keeps faces and costs 6 credits per image at 2K. Choose GPT Image 2.5 if you want the strongest film look.
4. Check the total and start. The form shows the image count and credits before you spend anything. Our group above was 2 selfies × 4 prompts = 8 images × 6 = 48 credits. Download everything as one ZIP.

Try one person first with your two favorite looks, then run the whole group. If faces drift on someone's photo, try a sharper, front-facing selfie of them. For more on keeping the same person across many images, see consistent character AI images; we ran the same kind of side-by-side test on old photo restoration.

FAQ#

Which AI model is best for Halloween photos of yourself? In our test, Nano Banana Pro kept the face most faithfully, Nano Banana 2.1 followed costumes best while keeping the face close, and GPT Image 2.5 had the strongest film looks but made the subject look older. With one run per prompt, test your own selfie before running a batch.

Can I use these prompts in Gemini or ChatGPT? Yes. They are plain language and work with any model that accepts a reference photo. Gemini's image generation runs on the Nano Banana family; ChatGPT's runs on GPT Image.

How do I stop the AI from changing my face? End the prompt with an identity line (same face, features, skin tone, age and hairstyle), avoid age words in the scene, and use a clear, front-facing selfie. The model matters too: the Nano Banana models changed faces least.

Why did the vampire prompt not make my skin pale? None of the three models did it in our test, most likely because the identity line asks to keep your skin tone. Describing the pale skin as costume makeup is worth a try.

Can I make Halloween photos for a group of friends at once? Yes. Upload all the selfies as reference images, use Per image mode, and every person gets every look in one batch.