Denoise

The model is dreaming. Name what it is hallucinating before it finishes.

How Denoise Works

This is a diffusion model running its reverse process, live, in front of you. Every round starts as a tensor of near-pure noise and progressively denoises — sharpening into a recognizable pixel-art subject over a few seconds. Your job: identify the latent before it fully resolves.

  1. A sprite begins as a shimmering field of static
  2. It sharpens a little more every frame as the sampler does its work
  3. Tap (or press 1–4) for the label you think it is
  4. Guess earlier — with more noise still present — for more points

Scoring & Lives

Points are time-weighted: catching a subject at 90% noise pays roughly ten times more than waiting for the clean image. Consecutive correct IDs build a combo multiplier up to x4. You get 3 lives — a wrong guess re-injects noise and costs one, and letting the sample fully resolve without answering also forfeits a life. The latents wait for no one.

Slop Fact: Real diffusion models start from Gaussian noise and iteratively predict-and-subtract the noise over dozens of steps. You're doing the same eval, except your "denoiser" is a primate that has seen a cat before. Guessing at 90% static isn't perception — it's prior. Embrace the overfit.

Back to the Slop