Each sample is a triple: (previous frame, action, next frame). A procedural renderer makes them, so no real giraffe had to pose.
Step 0
Loss —
Neural confidence: Sees nothing (untrained)
Conditional DDPM · 40 timesteps · tiny U-Net. About 1500 steps gets you a giraffe-shaped belief.
Press an action. The network will imagine what happens next.
Streak 0 · Best 0