🧠 Browser GAN Lab — 64×64

A real neural network. No APIs. No shortcuts. Pure math. 64×64 pixels of pure creativity.

Training Data

📁

Drop images here or click

Resized to 64×64 RGB

Upload similar images (faces, textures, patterns) for best results. Network trains on downscaled representations and upsamples to 64×64.

Generator

z[64]→Dense 256→8×8×16→↑16→↑32→↑64→64×64×3

Discriminator

64×64×3→Pool→16×16→Dense 128→Dense 64→σ[1]

G: 280675D: 106753Total: 387428

Output: 64×64×3 RGB (12288 values per image)

Learning Rate0.0010

Batch Size2

Latent Dimension64

Internal Spatial Res8×8

Channel Width16

Turbo Mode

📐 Output: 64×64 (fixed)

Step 0 / ∞

Awaiting training data...

Generated 64×64 images will appear here during training

Timeline appears during training...

Train a model first, then drag to explore its latent space and generate 64×64 images

📊 Performance & Memory Notes

• Estimated model params: 387428

• Each 64×64 RGB image: 48 KB as Float32

• 0 training images ≈ 0.0 MB

• For best results: use batch size 1-2, internal res 8×8, channels 8-16

• Turbo mode runs 3 steps per frame (may lag UI)

• If your tab freezes: reduce channels/batch size or close other tabs