Everyone becomes the denoiser. The facilitator says the words, noise sits on every canvas, and the class draws in timed steps — biggest shapes first, details last, a veil of noise thinning between rounds on a fixed schedule. Each drawing exports as a PNG and as a step-by-step trajectory GIF.
Live preview · launch for the interactive version
The brush shrinks every step, forcing a coarse-to-fine drawing trajectory. That echoes a common denoising pattern without claiming every model follows this exact order.
A noise-remaining meter drops by the same fraction at every commit. Between steps, a veil of noise falls — thick early, thin late, gone after the last step.
Undo works only inside the current activity step so the trajectory stays visible. A sampler moves forward through timesteps, while later updates can still revise structure in its current representation.
Pick step count and noise, type the words, press Copy setup link, paste it in chat. Everyone opens the same conditioning.
About a minute per step. Step 1 is big shapes only — "find them in the noise." Commit together and watch the veil fall.
Each participant exports the finished image and the animated trajectory. Nothing uploads; sharing is a consent decision.
Tile the GIFs on the Evidence Wall, then open the Diffusion Step-Through Viewer. The human drawing and machine denoising make different kinds of iterative change visible.