| --- |
| license: other |
| tags: |
| - 3d |
| - neural-fields |
| - siren |
| - hypernetwork |
| - shape-generation |
| --- |
| |
| # Hypernetwork → Shape pipeline checkpoints |
|
|
| Trained models and processed data for image-to-3D experiments documented at |
| **[BOB-THE-BUILDER-in/Hypernetwork](https://github.com/BOB-THE-BUILDER-in/Hypernetwork)**. |
|
|
| ## Contents |
|
|
| - `tier_essential.tar.gz` (1.98 GB) — anchors, autoencoder, mappers, results |
| - `tier_data.tar.gz` (2.67 GB) — watertight meshes, SDF samples, image-SIRENs, shape-SIRENs |
| - `tier_hypernets.tar.gz` (6.69 GB) — 100 trained hypernets (one per training shape) |
|
|
| ## Source |
|
|
| Mesh data derived from [Objaverse 1.0](https://objaverse.allenai.org/). |
| Individual creators retain rights to original models. |
|
|
| ## Restore |
|
|
| See the GitHub repo's README for full restoration instructions. |
|
|
| ## Key results |
|
|
| - Direct weight prediction (264K-dim) at N=100 produces fragmented OOD predictions |
| - Latent autoencoder compression to 128 dims rescues OOD generalization |
| - Mapper from hypernet → 128-dim latent achieves MSE 1.4e-5 with 9M× scrambled-cond ratio |
