Trying to make AI video feel less “AI-generated”: notes from my short film project
I recently finished a Korean AI short film inspired by the idea of AI hallucination.
The concept was not just “AI makes mistakes,” but the idea that humans also misremember, misread, and sometimes edit reality inside their own minds.
Most of the video was made with Seedance 2.0.
Music was created with Suno.
I used ChatGPT mainly for prompt development, but the story itself was written by me.
The hardest part was trying to make the footage feel grounded and realistic instead of obviously AI-generated. I wanted to avoid relying too much on flashy AI visuals. Instead, I focused on small details: wet streets, phone screens, reflections, Korean city textures, subway interiors, and the uneasy feeling of “did that really happen?”
A few things I learned:
- Realistic lighting matters more than big spectacle.
- Reflections are useful, but they can easily become too “AI shiny.”
- Keeping visual continuity between shots is still one of the hardest parts.
- Simple camera movement often works better than complicated movement.
- A strange idea becomes more believable when the environment feels ordinary.
For anyone else making AI short films, how do you deal with keeping scenes realistic across multiple shots?