04 / lab

Diffusion, by hand.

A toy reverse-diffusion sampler, the same idea behind the hero and behind the driving planners I study: start from pure noise, then denoise toward a target. Pick a target, drag the steps, and sample.

target

Fewer steps resolve faster but rougher; more steps converge cleaner. That speed-versus-fidelity tradeoff is exactly what DDIM buys you over DDPM, which I implemented from scratch in av-policy-lab.

No model runs here; it is a particle toy that mirrors the schedule. The real samplers live in the repo. See the work →

// drive · perception + avoidance

Don't hit the cars.

Steer with ← / → (or A / D, or tap a side). The amber boxes are detections; keep the ego clear of them. A toy nod to the closed-loop avoidance behind av-policy-lab.