Train a real generative model in your browser.

Five small but real architectures (DCGAN, StyleGAN v1, BigGAN, DDPM and VQGAN) run on your GPU with TensorFlow.js. Use a classic dataset or your own images. Your coach watches every rep and calls out mode collapse, lazy codebooks and runaway discriminators.

01 data02 model03 train04 generate05 benchmark

Choose your training data

Every image is resized to a small square so training stays fast on a laptop GPU.

Or load any Hugging Face image dataset

Settings

Smaller images train much faster. 32 px is the sweet spot.

Resolution

Samples to load

Color

Horizontal-flip augmentation (skip for digits)

Nothing loaded yet.

Dataset preview

Load a dataset to see a sample grid here.

status: empty

Pick your model

Each one is a scaled-down build of the real architecture, with the parts that make it what it is.

What's inside

Build the model to see its layer summary and parameter count.

Hyperparameters

Training plan

Load checkpoint

Step0

Speed–

Elapsed0:00

ETA–

Epochs0

Params–

Loss curves

Live samples

Fixed seed, so you can watch the same samples improve.

Coach Ada watches your run and calls it as she sees it

Generator

No model yet.

How many

Seed

Train (or load) a model, then hit Generate.

Benchmark arena

Each selected model gets a fresh start on the loaded dataset with the same budget. We measure speed, sampling time, and sample quality with kernel MMD against real images (lower is better) and diversity (closer to the real data is better).

Steps per model

Batch size

Load a dataset first.

Results

Model	Samples	Params	ms / step	Sample 64	Final loss	MMD ↓	Diversity
No runs yet.

Global leaderboard

#	Model	Dataset	Res	Steps	MMD ↓	ms/step	Params	Athlete	Device
Loading…