| --- |
| license: cc-by-4.0 |
| tags: |
| - diffusion |
| - ddpm |
| - ddim |
| - onnx |
| - quickdraw |
| - image-generation |
| library_name: onnx |
| pipeline_tag: unconditional-image-generation |
| datasets: |
| - quickdraw |
| --- |
| |
| # You Are the Model - Quick, Draw! diffusion |
|
|
| A tiny class-conditional diffusion model that draws 32x32 doodles, small enough |
| to run a full 25-step DDIM sample in the browser. |
|
|
| It powers **You Are the Model**, a daily drawing puzzle where the player takes |
| over the denoising loop: at each round you pick which noise prediction to apply, |
| your choice is fed back into the real sampler, and a bad pick genuinely derails |
| the drawing. |
|
|
| ## Model |
|
|
| | | | |
| |---|---| |
| | Architecture | class-conditional U-Net, epsilon prediction | |
| | Parameters | 1,929,313 (7.7 MB fp32) | |
| | Resolution | 32x32 grayscale, values in [-1, 1] | |
| | Classes | 24 (+1 null class for classifier-free guidance) | |
| | Noise schedule | cosine, 1000 train timesteps | |
| | Sampler | DDIM, 25 steps, guidance 3.0 | |
| | Trained | 60,000 steps, batch 256, AdamW, EMA 0.9995, bf16 | |
| | Hardware | NVIDIA GeForce RTX 3090, 1.23 h, about $0.167 | |
|
|
| ### Classes |
|
|
| `fish`, `flower`, `house`, `sailboat`, `smiley face`, `star`, `umbrella`, `envelope`, `t-shirt`, `ice cream`, `apple`, `banana`, `ladder`, `lightning`, `mountain`, `pants`, `wine glass`, `eyeglasses`, `snowman`, `car`, `hand`, `crown`, `donut`, `camera` |
|
|
| ## Files |
|
|
| | File | Purpose | |
| |---|---| |
| | `yatm_v2.onnx` | the browser asset - opset 17, dynamic batch | |
| | `yatm_v2.meta.json` | class list, sampler settings, parity numbers | |
| | `yatm_v2.abar.json` | cosine alpha-bar table, so the JS sampler matches exactly | |
| | `ema.pt` | EMA weights (PyTorch) | |
| | `samples.png` | fixed-seed sample sheet, one row per class | |
| | `eval.json` | classifier gate and guidance sweep | |
| | `metrics.jsonl` | training and validation curves | |
| | `config.resolved.json` | the exact resolved config and environment | |
|
|
| ## Inputs and outputs |
|
|
| ``` |
| x: float32 [batch, 1, 32, 32] noisy latent |
| t: int64 [batch] timestep, 0..999 |
| y: int64 [batch] class index, or 24 for unconditional |
| -> eps: float32 [batch, 1, 32, 32] |
| ``` |
|
|
| Classifier-free guidance: run the batch twice (class and null) and combine with |
| `eps = eps_null + w * (eps_class - eps_null)`. |
|
|
| ## Evaluation |
|
|
| A held-out CNN classifier (95.9% accurate on real test drawings) labels **100.0%** of fixed-seed samples as the class they were conditioned on, at the best guidance scale of 3.0. Weakest class: `fish` at 100.0%. |
|
|
| Python and ONNX Runtime agree to `8.3e-02` maximum absolute difference over |
| a full 25-step sample, so the browser and the training code produce the |
| same picture from the same seed. |
|
|
| ## Dataset and attribution |
|
|
| Trained on the [Google Quick, Draw!](https://github.com/googlecreativelab/quickdraw-dataset) |
| dataset (`numpy_bitmap` 28x28 drawings, padded to 32x32), which is released under |
| [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) by Google Creative Lab. |
| 12,000 drawings per class were downloaded, filtered for empty and |
| over-inked drawings and exact duplicates, then split deterministically. |
|
|
| This model is released under the same CC BY 4.0 licence. |
|
|
