Alien Craps is an R&D short that answers one question: can AI video hit an exact physical action? Not a mood. An action with rules everyone knows. A dice throw is the perfect test, because everyone can see when dice fall wrong.
The challenge
Video models make a great-looking casino. They cannot count. Ask for two dice and you get dice that melt, merge and change faces in the middle of a roll.
The approach
I built the motion first and let the model dress it. I simulated the throw with real rigid-body physics in MuJoCo, rendered a clean 3D guide in Blender (dice paths, blocking and camera), and then generated the shot in Seedance from two video references: one for the cast and the action, one for the dice and the camera. The look and the cast came from Midjourney and were locked in reference sheets, so nothing drifted between shots.
The use case
Any shot where an object must behave: a product that falls and lands logo-up, a pour, a car that takes a precise line, a ball that bounces into frame. Physics first, then AI. The method is now a reusable 3D previz workflow I apply to client work.
AI tool stack:
MIDJOURNEY ∙ GPT IMAGE 2.5 ∙ MUJOCO ∙ BLENDER ∙ SEEDANCE 2.5 ∙ TOPAZ ∙ ELEVENLABS
2 video references steering one shot
13 second final shot, finished in 4K
1:51 breakdown film
