Tiny Diffusion Lab

01 / Dataset

0 / 50 imgs

Feed me pixels

drop, click, or paste · 5–50 images recommended

resolution

color

schedule

t = 30

We add noise gradually. The network learns to undo it. Slide to watch your first image dissolve into television static.

timestep√ᾱ = 0.611

t14

t28

t42

t57

t71

t85

t99

no model

epochs200

batch size16

learning rate2.0e-3

model size

epoch

0/200

step

loss

—

s/epoch

—

dream previews · every 10 epochs

Nothing yet. Early dreams look like soup; this is normal.

idle

images

sampling steps50

temperature1.00

seed

Train a model first. Generating without one is just… noise.

The hall is empty and echoing. Be the first blob.