Every October the same thing happens. Someone turns a selfie into a vampire, posts it, and their friends reply "who is that?" An AI Halloween photo is only fun if it still looks like you, and the prompt lists going around don't tell you which edits keep your face and which ones quietly swap you for a stranger.
This guide sorts the popular edits by how much of you they keep. Costume swaps and scene changes are safe. Full creature transformations cost you some resemblance, and it's worth knowing that before you spend credits on twenty zombie attempts.
What an AI Halloween photo edit is, and why some look like a stranger
You upload a real photo to an image editing model and describe the change in words. The model redraws the picture with your change applied. It isn't a filter laid on top. Every pixel gets regenerated, including your face, which is why results can drift. You can try it on a single page with edit a photo with an AI prompt, or open Image to Image for every editing model side by side.
The rule of thumb: the more of your face the prompt asks to change, the less of you survives. "Put a witch hat on me" leaves your features alone. "Turn me into a rotting zombie" asks the model to rebuild your skin, eyes, and mouth, and it fills the gaps with a generic zombie face that shares your haircut.
The Nano Banana editors are the obvious place to start. Google's own Halloween post from October 2025 pitched Nano Banana for ghostly portraits, spooky invites, and scary movie posters, and Google says Nano Banana Pro can keep the resemblance of up to five people in one image. In our Image to Image tool you get Nano Banana Edit, Nano Banana 2 Edit, and Nano Banana Pro Edit, plus Seedream 5.0 Pro Edit, Flux Kontext Pro, and GPT Image 2.5 Flare Edit if one of them isn't landing.
Pick your starting photo: light, angle, one face
Most failed edits were doomed by the input. A good source photo does more for resemblance than any prompt trick.
- Even, soft light on the face. Hard shadows get read as features and come back as scars or bruises.
- Face toward the camera, or a slight three-quarter turn. Profiles give the model half a face to work from.
- Your face big in the frame. A full-body shot from across the room leaves too few pixels to hold on to.
- One person, unless you want a group edit. Crowds in the background tend to get costumes too.
- No sunglasses. The model will invent eyes, and they won't be yours.
- Pets follow the same rules. A cat looking at the camera in daylight edits far better than a blurry shot of it under the couch.
Costume swaps that keep your face
This is the safest AI Halloween costume photo, and the one people like most. Change the clothes and props, and tell the model outright what to leave alone. That instruction does real work. Models change less when you name what to keep.
Try: "Dress this person as a 1920s witch: black pointed hat, high-collar black dress, a broom in one hand. Keep the face, hair color, expression, and pose exactly the same. Same lighting. Photorealistic."
Swap the costume words and the rest of the prompt carries over: pirate captain, astronaut, the Bride of Frankenstein's hair, a knight's armor. For a pet, keep it playful: "Put this black cat in a tiny wizard hat and cape. Keep the cat's face, eyes, and fur exactly as they are."
Where it breaks: hands and props. A broom can fuse into fingers, and a held sword sometimes grows a second hilt. If the face is right and the hand is wrong, don't regenerate from scratch. Take the result back in and ask for one fix, like "fix the left hand holding the broom, change nothing else."
Full transformations: ghost, vampire, zombie, and the identity cost
These are the edits everyone wants, and the ones where you pay in resemblance. Here's roughly how they rank, from gentlest to harshest.
| Look | What changes | How much of you survives |
|---|---|---|
| Vampire | Pale skin, red eyes, fangs, cape | Most of it. Bone structure mostly stays |
| Victorian ghost | Translucent skin, period clothes, cold light | Good, if you ask to keep pose and framing |
| Skeleton face paint | Paint over your real features | Good. It's makeup, not new anatomy |
| Zombie | Skin, eyes, teeth, wounds | Patchy. Often a stranger with your hair |
| Werewolf or creature | The whole head | Little. Expect a family resemblance at best |
Google's own example prompt for the ghost look is a good template because it pins the pose down: "Transform this person into a semi-skeletal spectral translucent Victorian ghost. Victorian clothing. Maintain existing pose and framing. ultra-realistic photograph."
For zombies and creatures, go in steps. First make the face gray and sunken with clouded eyes. Save it. Then add the wounds and torn clothes in a second edit. Each small edit keeps more of you than one big one. We cover the same idea in more depth in keeping a character consistent across images.
Swap the scene, not the person
If you care most about looking like yourself, leave your face and clothes alone and change the world around you. A foggy cemetery, a candlelit haunted hallway, a pumpkin patch at dusk with green light coming from behind.
Try: "Place this person in a foggy graveyard at night, lit by a single lantern from the left and green moonlight from behind. Keep the person exactly as they are, same face, clothes, and pose. Match the lighting on the person to the new scene."
That last sentence matters. Without it, you get a studio-lit face pasted into a dark scene, which looks like a sticker. The cost of asking for relit skin is a little drift, since the model is now allowed to touch your face. It's usually worth it.
Vintage horror posters with readable titles
A 1950s B-movie poster with you screaming on it makes a great party invite. The hard part used to be the lettering. Google says Nano Banana Pro is its best model for legible text inside an image, so pick Nano Banana Pro Edit for this one.
Try: "Turn this photo into a 1950s horror movie poster. The person looks terrified, lit from below. Big hand-painted title at the top that reads "THE THING IN APARTMENT 4B". Small text at the bottom: "Oct 31, 8pm". Worn paper texture, faded red and yellow inks."
- Put the exact words in quotes.
- Keep text short. A title and one line lands far more often than a paragraph of credits.
- Check every letter before you send it. If one word is off, edit just that word instead of rerolling the whole poster.
Make it move: a 5 second spooky clip from your edit
Once you have a still you like, open it in Image to Video with the "Use as input" button on the result. Wan 2.7 animates from your edit, and 5 seconds is its default length. That's plenty for a looping story or a profile video.
Describe motion, not the picture, since the picture already exists: "Slow push in. Fog drifts across the floor. The candle flickers. She turns her head slowly toward the camera and smiles." Keep it to one or two actions. The more the person moves, the more their face can shift between frames, so a slow head turn keeps you recognizable better than a lunge at the lens. There are more tips in how to animate a still image.
AI Halloween photo FAQ
Can I make an AI Halloween photo of my pet?
Yes, and pets are some of the most forgiving subjects. Use a sharp photo where the face fills the frame, and say what to keep: "keep the cat's face, eyes, and fur pattern". Costumes and props work best. Full creature transformations drift on animals just like on people.
Will the AI change my face?
It can. Every pixel is redrawn, so the more of your face the prompt touches, the more it drifts. Costume and scene swaps that say "keep the face exactly the same" hold up well. Zombies and monsters hold up better when you build them in two or three small edits instead of one big one.
Which model should I use?
Start with Nano Banana Pro Edit for faces and for posters with text. Seedream 5.0 Pro Edit and Flux Kontext Pro are worth a second try on the same photo when the first result looks off. Running the same prompt on two models costs two generations, so try it on your best photo only.
How much does it cost?
Each edit and each clip uses credits from your plan, and the studio shows the cost before you press Generate. Video costs more than a still, so get the still right first. Failed generations are not charged.
The short version
- Start with a clear, evenly lit, front-facing photo of one person or pet.
- Costume swaps and scene swaps keep your face. Always say what to keep.
- Ghost and vampire looks keep most of you. Zombies and creatures drift, so build them in small steps.
- Use Nano Banana Pro Edit when the image needs readable text.
- Animate the finished still with slow, simple motion.
Upload a photo in Image to Image and start with a costume swap. It's the edit most likely to still look like you.
