world models · real time · on-device

ADS Dream
Simulator

live capture

A world model running on iPad Pro at 30 FPS capped.

30 fps

on iPad Pro, capped

model card · ADS Dream Simulator

research preview

architecture: Flow-matching diffusion transformer (DiT) with adaLN-Zero conditioning, pixel-space.
output: 64 × 64 RGB at 30 Hz, conditioned on the previous frame.
conditioning: Player input as a 4-bit action mask (WASD) plus the prior frame — the model predicts the next frame given what you press.
training data: ~3M frames of bot self-play captured at 30 Hz in our own Godot simulators — desert, ocean, grass and street worlds.
training: PyTorch, distributed data-parallel on 8 × NVIDIA A100 on AWS (p4d.24xlarge), float32 with TF32 matmuls, torch.compile.

the reel

4 models, 1 ipad

Go outside. Touch grass. Play world models.

The world model interprets this ocean wave as a hill for it to drive on.

Desert patrol — a police cruiser rolling through a neural desert, one continuous rollout.

A gloopy little game, dreamed entirely by a local world model.