🧠 Neural Diffusion Playground

Watch a real neural network denoise your images — step by step

⬆

Drop an image here or click to upload

PNG, JPG, WEBP — resized to 64×64 for neural processing

Noise Steps 50

Number of discrete timesteps in the diffusion process. More steps = finer gradation of noise.

Noise Schedule

Controls how noise variance increases over time. Cosine preserves more signal in early steps.

Learning Rate 0.003

How quickly the network updates its weights during training. Too high = unstable, too low = slow.

Architecture

Conv
3×3
→8

→

Conv
3×3
→16

→

Conv
3×3
→16

⤴ skip

Conv
3×3
→3

Tiny U-Net: 4 conv layers with skip connection from layer 1 to layer 4. Predicts noise ε(x_t, t).

Original Image

64×64 RGB

Current State

t = 0

Network Prediction

ε̂(x_t, t)

Training Loss

Idle

Epoch 0/0

Conv Filter Weights

Real-time visualization of the first convolutional layer's learned filters.

Diffusion Timeline

No GPUs were harmed in the making of this diffusion model 🍓

Remix on Berrry