Halloween AI Photo Prompts

What a Halloween prompt for Gemini needs

Halloween edits have a problem the other holidays don't: the look you want is dark. Fog, dusk, a lantern held low, and darkness is exactly where an image model loses detail on a face. «Make me a scary witch» makes it worse. It gives the model permission to change everything, your features included, and «scary» pulls toward things it won't draw. A costume edit that comes back looking like you settles five things.

1. The photo, and who's in it. «Use the attached photo» for one person, «the two attached photos, one of each person» for a couple, «the attached photo of my dog» for a pet.
2. The hold list. Face, skin tone, hair and body, «exactly as they are». For face paint, say «add only makeup», so the paint sits on your skin and your features stay underneath.
3. The costume, piece by piece. A wide-brimmed hat, a long cloak over simple clothes, worn leather boots. Pieces give you control over the result; a one-word costume leaves the model to fill in the rest.
4. One light you could hold. A lantern, a candle, a carved pumpkin, and where it sits. Then «face clearly lit». The spooky part comes from what's in shadow around you, and your face shouldn't be one of those things.
5. A tone and a no list. «Playful», «moody» or «eerie» sets the dose. Close with no gore, no weapons, no text, no watermark, no extra people. Costumes like pirate or hunter suggest props you probably don't want in your hand.

Put together, a request you can type from memory:

Use the attached photo. Keep my face, skin tone, hair and body
exactly as they are. Dress me as [costume, piece by piece] in
[one setting]. The only light is [a lantern / a candle / a pumpkin]
[where it sits], with my face clearly lit. [Playful / moody],
aspect ratio 4:5. No gore, no weapons, no text, no watermark,
no extra people.

Nine prompts on this page are that request worked out for one costume or setting. Three build a picture from nothing and need no upload. For restyles that have nothing to do with October, the Gemini photo prompts use the same hold-list method.

Which Halloween prompt for which look

Pick the look, then check what to upload. Costume edits want a clear, front-lit photo; the dark comes later.

A costume request, before and after

A scarecrow isn't one of the prompts on this page, which makes it a good test of the method. We wrote both versions below to show the difference.

turn me into a creepy scarecrow for halloween, scary

The horror version of a scarecrow has a stitched sack where the face should be, so «creepy» and «scary» together point the model at replacing your face, the one thing you wanted kept. Nothing protects it anyway. There's no setting and no light, so you'll get the model's average of every scarecrow picture it has seen.

Use the attached photo. Keep my face, skin tone, hair and body
exactly as they are. Dress me as a scarecrow: a battered straw
hat, a patched plaid shirt, rope tied at the waist, straw sticking
out of the cuffs and collar. Add a few painted stitch lines at the
corners of my mouth as makeup only, my features unchanged. I'm
standing in a dry cornfield at dusk, a lantern hanging on a post
beside me lighting my face warm from the side, cool blue sky
behind, crows on a fence far away. Playful and a little eerie,
aspect ratio 4:5. Photorealistic. No gore, no weapons, no text,
no watermark, no extra people.

The costume is now all clothes and makeup, so there's nothing left for the model to redraw on your face. The eerie part lives in the scene: dusk, dry corn, crows in the distance. And the lantern hangs beside you, which keeps one side of your face in warm light and lets the other fall into shadow. The stitch-line makeup borrows from the Skeleton Face Paint Edit, and the lantern from the Witch Costume Photo Edit.

Follow-up lines that fix a Halloween edit

Close? Reply in the same chat. The model still has your photo and your request, so a one-line correction is usually quicker than starting over with a new prompt.

Where Halloween edits go wrong

Most failures here trace back to the dark, and a few to the costume itself.

• Low light costs likeness. The less light on a face, the less of it the model has to hold on to. That's why every edit prompt on this page keeps one warm source on the face, however dark the rest of the frame gets.
• Underlighting changes your shape. Light from below is the classic spooky look, and it flips the shadows people know your face by. Jack O Lantern Glow pairs it with a gentle expression. If the result doesn't read as you, raise the light, as in the second follow-up.
• Two photos, two kinds of light. Couple Costume Portrait takes one photo of each person, and if one was shot in daylight and the other under a kitchen lamp, one face often comes out better than the other. Choose two front-facing photos taken in similar light.
• A single word can stop the job. When a costume prompt keeps coming back empty or refused, look for a word that reads as violent out of context, like a blade on a pirate or «bloody» in a costume name, and take it out. The pirate and couple prompts already say no weapons.
• Someone else's face. A costume edit of a friend is still their face on the internet. Edit yourself, your pet, or adults who've said yes.

Running these prompts in Gemini and ChatGPT

All twelve run as typed in either app. These details come from Google's and OpenAI's own help pages and are current as of October 2026.

• Gemini without a plan. Google lists image generation with Nano Banana 2 for every Gemini plan, including no plan. Limits are compute-based and refresh every five hours up to a weekly cap, and Google says media generation uses more of that budget than chat, so a long run of retries adds up.
• Age. Google requires you to be 18 or older to edit images in Gemini, which covers every costume edit on this page.
• The couple prompt in Gemini. Nano Banana 2 accepts multiple reference images. With the Gemini model set to Flash-Lite, it uses Nano Banana 2 Lite, which Google says isn't designed for multiple image references, so switch to Flash or Pro before you upload two photos.
• ChatGPT. OpenAI lists ChatGPT Images on all tiers, free included, and you edit by uploading a photo and describing the change. To fix one spot, such as a hat, the selection tool marks just that area; OpenAI notes edits can spread past the selection, so check your face again afterwards.

Plan prices are in our Gemini and ChatGPT reviews. For the image model itself, see the Nano Banana prompts, and when the pumpkins come down, the Christmas photo prompts use the same method for cards.