AI Halloween Costume Generator

For Halloween 2025, Jordan built a costume builder: snap a photo with your webcam, pick a costume or type your own, and AI dresses you up and turns the result into a short "Happy Halloween" video. It runs on desktop and phone. He built it in Replit with Nano Banana for the image and Veo 3.1 Fast for the video, both through Replicate, and showed it in Episode 18 (October 24, 2025). He didn't give a build time; Sam said it had been out for 48 hours.

Why we built it

The idea came from Paige Bailey, DevEx engineering lead at Google DeepMind, who posted a webcam demo using Nano Banana and Veo 3.1. Jordan promised to credit her and built our own version as a fun Halloween tool for the team.

The stack

• Replit: the app, with API keys stored there.
• Nano Banana: turns the photo into a costumed image.
• Google Veo 3.1 Fast: animates that image into a Halloween greeting.
• Replicate: serves both models.

So just to kind of run through them real quick, I built this in Replit. The APIs that I'm using are nano banana from Google, as well as Veo 3.1 fast.

— Jordan Metzner, Episode 18

How we built it, step by step

1. Build the app in Replit. Jordan said it's "not a very complicated application."
2. Capture a webcam photo. Works on computer and phone.
3. Pick a costume. Preset options (pirate, superhero, ninja) or type your own.
4. Send photo and costume to Nano Banana via Replicate. A result usually comes back in about 20 to 40 seconds.
5. Use that image as the source for a Veo 3.1 Fast video with a short Halloween prompt.
6. Add downloads, a gallery and technical details.

How it turned out

Live, Jordan became a ninja, then an 1800s gunslinger. The images kept his pose, office and hand position almost exactly, and removed his background, but the face wasn't always quite his: "I don't know if those are exactly my eyes." The videos worked, though the characters' speech was mostly mumbling rather than real words. Jordan compared it to how early AI images garbled text, something that has mostly been fixed since.

The team used it a lot. People made everything from vampires to wizards to custom costumes, and Sam shared this at 08:19:

A teammate this morning told me, based on one of the videos, she it actually inspired what she was going to be for Halloween.

— Sam Nadler, Episode 18

Sam also saw a business angle: we already use Veo for B-roll, and turning a plain webcam shot into publishable video is much faster with these tools.

What we'd do differently

• Other stacks work too. Jordan went with Google's models because Paige's demo did, but said Sora 2 could do the same job.
• Don't expect real dialogue yet. Keep any speech short and treat it as a gag.

The next week we went further with Mirror, Mirror on the Wall, the Guess My Costume app and the Trick-or-Treat route planner. See also ChatGPT vs Nano Banana.

FAQ

How do you make an AI Halloween costume generator from a photo?

Jordan's app captures a webcam photo, sends it with the chosen costume (pirate, superhero, ninja or your own) to the Nano Banana API, and gets back a costumed image. That image then becomes the source image for a short Veo 3.1 Fast 'Happy Halloween' video.

How do you use Nano Banana and Veo 3.1 together in an app?

Jordan called both through Replicate, with his API keys stored in Replit. Nano Banana makes the costumed still, and Veo 3.1 Fast animates it from a short Halloween prompt.

How long does Nano Banana take to generate an image?

In Jordan's app, the costumed image usually came back in about 20 to 40 seconds.

Do AI Halloween videos get the person's likeness right?

Partly. The images kept body pose and hand position well but weren't always a perfect likeness, and the characters' speech in the videos was often mumbled non-words.