An interactive horror story set in a hospital at night, made with AI-generated footage.
You wake in a hospital at night. On-screen choices move you through the story as you try to escape.
H3 Max Turbo’s speed made me want to try generating every scene as the player progressed through the story. Charlie Clark’s Diiverge, a point-and-click adventure that grows with players’ choices, was a reference for that idea.
A hospital at night feels frightening to me. The setting provides familiar spaces and characters: wards, operating rooms, patients, nurses and doctors. I wanted the world to feel real enough that exploring it would be unsettling.
I found it easier to browse visual styles in Midjourney than to search for movie or TV references. I used the same hospital prompt for each style test.
I liked the fisheye lens because it felt immersive.
For character concepts and scene stills, I used GPT Image 2. The images used as the videos’ final frames were edited with Nano Banana Lite Edit through fal. FLUX.2 Klein 9B Edit was used to refine the Surgeon scene.
I introduced the Surgeon as a silhouette because wondering who was in the doorway felt scarier to me than seeing him clearly.
I wanted choices to change what happened later, but the story still needed an ending. In the cabinet prototype, players could choose between a notebook, a syringe and a map. With a story generated dynamically, the LLM would have to carry that choice into later scenes.
Keeping an improvised story coherent was difficult, so I chose a limited set of written branches with the consequences decided in advance.
Generating every scene live would also mean accepting the takes as they arrived. A bad clip could break the immersion. Making most of the footage in advance let me choose the takes and limit production costs. Only the name-reveal scene is generated live.
I still wanted to showcase H3 Max Turbo’s speed with a scene generated for each player. I chose the name reveal because seeing your own name inside the hospital makes the story feel as though it is addressing you directly. The name is generated into the video itself.
The generated scene still had to connect to the footage on either side. I supplied H3 Max Turbo via fal with a first frame matching the preceding ward view and a last frame with the Nurse in position for the next scene. The frames and the rest of the prompt stay fixed; only the player’s name changes.
I wanted the name scene ready before the player reached the reveal. Generation starts as soon as they submit their name. The instructions, the dark ward and the Surgeon’s appearance give it time to run.
Larger video files took longer to download, so I chose 480P to help the scene finish downloading during the opening. If it still isn’t ready when the Surgeon’s scene ends, the game loops an idle ward clip until the name scene is ready.
1.3K playthroughs 16.5K views on X