By skyforbes 
What was originally going to be maybe 50-100 images, in the end was over 16,500 images. I then used Codex to create a website. Gemini helped to create prompts based on movie villains, survivors, halloween costumes I found online, etc. I took a single image for each employee from our directory then used batch processing inference to cut our cost in half.
I also added a feature where you can use our prompts to create an image of your own.
I'm using custom system instructions to keep the style consistent to match the theme of our party. We then process these on gemini-2.5-flash-image (Nano Banana), store the asset in a GCP bucket, and use Imgix to apply a watermark, text, and resize the image in your browser. Give it a shot!
Finally, I created a video of a manually selected image for each person and brought it to life with Veo 3.0. The images from NB were often too graphic for Veo. Blood or weapons depicted in an image would not pass the prompt validation or would be rejected after generation time had passed for a similar reason.
I think we should all share prompts and processes more, but there are over 300 prompts in my project. So, here are just a few samples:
a photorealistic image of this person styled as Jigsaw, pale face lit by flickering orange light from a television screen, faint spiral cheek designs visible as smeared makeup, wearing a black suit with crimson bow tie, background of red-lit workshop walls and rusted chains, cinematic psychological horror portrait.
a photorealistic image of this person clutching a makeshift weapon, torn denim jacket soaked in rain, orange city lights behind, expression full of relief and disbelief, tight portrait framing.
a photorealistic image of this person clutching a flashlight in a collapsing basement, orange beam illuminating their horrified expression, cobwebs and smoke swirling behind, high-contrast horror framing.