Can your model fool the room?
Hand in your drawings. Make your calls. Meet the artists behind the impostors.
Students: add and verify your @columbia.edu email in GitHub email settings before signing in. It can stay private and need not be your primary address.
Get your notebook
Download a notebook and its matching data-and-weights ZIP, open Google Colab, choose File → Upload notebook, then Runtime → Change runtime type → GPU. Upload the matching ZIP using Colab’s Files sidebar (folder icon). Wait for the upload to finish, then run the cells in order and change one setting.
For a first experiment, try Stroke GRU or Flow matching. VAE and GAN are included for exploring model limitations.
- VAE notebook — 60,000 steps · experimental; weak new samples. Data + weights ZIP
- Stroke GRU notebook — 18,000 steps. Data + weights ZIP
- Diffusion notebook — 60,000 total steps. Data + weights ZIP
- GAN notebook — 30,000 steps · experimental. Data + weights ZIP
- Flow matching notebook — 40,000 total steps. Data + weights ZIP
- VQ + token prior notebook — 40,000 total steps. Data + weights ZIP
Keep the asset ZIP compressed; the setup cell checks that it matches your notebook. By default, each experiment fine-tunes the supplied pretrained weights for 100 additional steps; it does not train from scratch. Each notebook shows learning curves with raw losses and a 10-step moving average, a human/baseline/experiment comparison grid, and samples from three fixed seeds. These plots are also saved in my-experiment.zip. At the end, run the notebook’s hand-in cell and upload classroom-submission.json below. Training data: Quick, Draw! · Google Creative Lab · CC BY 4.0.
Run the class round
Start voting once everyone has submitted. Reveal closes voting and makes names and results visible to the class. A new round archives this one.
Your final 16 drawings
Run the final Hand in your drawings cell in your notebook. Upload classroom-submission.json here. No weights or code needed.
Already using an older notebook? Copy that export cell into a new cell at the end and run it after your experiment.
Sign in to submit. Your GitHub handle and entry name appear in the class gallery after the reveal.
Human or machine?
Everyone sees anonymous drawings mixed with held-out human examples. All use the same black-and-white 32×32 display. You won’t be assigned your own drawings.
The instructor opens voting after submissions close.
Make your own call before discussing a drawing. Answers stay hidden until the final reveal.