ARotting/glyph-forge-cvae-model

GlyphForge CVAE

GlyphForge is a compact conditional variational autoencoder that generates 8x8 handwritten digits. A requested digit label conditions the decoder while an eight-dimensional Gaussian latent captures style.

The benchmark measures:

• held-out reconstruction MSE;
• KL divergence to the unit-Gaussian prior;
• generated-class fidelity through the frozen Tiny Vision classifier;
• within-class sample diversity.

Reproduce

uv run python projects/tiny-vision-foundry/prepare_data.py
uv run python projects/glyph-forge-cvae/train.py

The saved sample grid and Gradio app are model outputs, not hand-authored examples.

Verified results

• Parameters: 9,504
• Latent dimensions: 8
• Held-out reconstruction MSE: 0.02394
• Mean held-out KL divergence: 0.3978
• Conditional samples evaluated: 1,000
• Frozen-judge class fidelity: 99.90%

Nine digit classes achieved 100% judge fidelity; digit 1 achieved 99%. Every class had nonzero mean pixel variance across its 100 samples, ranging from 0.0066 to 0.0168. The judge is the 2,198-parameter Tiny Vision labels-only student with 98.52% accuracy on real held-out images, so this metric measures recognizability to that specific classifier rather than human perceptual quality.