Dreamcore
AI image-to-video pipeline / World building
A pastel, half-flooded world where ordinary infrastructure has been left to the weather: a bus in the shallows, a door in a waterfall, a level crossing in an empty field. I built it with image and video models running on my own GPUs.
Pipeline
Stills come from Krea 2 in ComfyUI, steered by a bank of my own reference images. Each picked still is animated with MiniMax H3 image-to-video, with the camera locked. Both models run locally. The larger megastructure scenes are rebuilt in Blender Cycles.
Scale
Between September and November 2026 I generated about 1,700 stills and 362 clips, and one overnight run alone made 444 stills and 54 clips. Each animation is a 5.9 second clip.
World building
One rule keeps it one place: a pastel palette, a locked camera and one impossible thing per image. Rounds on terrain, water, weather, bridges and giant interiors reuse it, so any two shots sit together.
