What is it
This is one deliberately overfit function. It maps a time-group and a spatial patch coordinate directly to four monochrome output patches. The browser reconstructs the sequence from weights and compact lookup codes rather than decoding a conventional video stream.
2,986,276 Bcomplete gzip artifact
75.7939%mean boundary F1 / 9,203 evaluated frames
2,899,994trained scalar parameters
9,581generated frames at 480 x 360
Strictly feed-forward: the function receives no previously generated image. Seeking to any frame is direct and deterministic.
Actual forward path
(t, x, y) → 96 Fourier/bit features → 512-wide projection → 10 residual MLP blocks (512 → 128 → 512) → rank-176 direct head + rank-128 PQ lookup → four 34 x 34 halo patches → 480 x 360 binary frame
- 192 spatial patches per four-frame group.
- Product quantization: five 8-bit codebooks across 128 lookup dimensions.
- Overlapping two-pixel halos feather patch boundaries.
- The eight maps on the left are computed from these tensors on the GPU, not decorative animation.
Complete artifact
| gzip size | 2,986,276 B |
| strict 3 MB margin | 13,724 B |
| raw payload | 4,373,892 B |
| tensor records | 67 |
| lookup rows | 460,032 |
| index values | 2,300,160 |
| SHA-256 | 62531850...58dc1 |
The loader losslessly reverses temporal-context bucket packing, expands quantized tensors, and submits the same network to WebGPU or the mobile WebGL2 fallback.