metadata
license: apache-2.0
language:
- en
size_categories:
- n<1K
tags:
- 3d
- sdf
- siren
- hypernetwork
- implicit-neural-representations
pretty_name: Hypernet Prior Topic 03 Archive
Hypernet — Prior Topic 03 Archive
Complete archive of the prior topic 03 research thread: per-shape SIREN decoders, per-layer hypernetwork architectures, mapper experiments, and ancillary checkpoints. This work predates the current image-to-3D pipeline documented in hypernet-image-to-3d and the main dataset.
100 shapes, naming convention obj_NN for NN in [0..99].
Contents
| Path | Size | Description |
|---|---|---|
watertight/ |
~5.6 GB | 100 watertight .obj meshes (filename obj_NN.obj) |
sdf_samples/ |
~591 MB | 100 .npz files with point/SDF pairs |
shape_sirens/ |
~505 MB | 100 trained per-shape SIREN decoders + reconstructions in meshes_w30/ |
hypernets/ |
~6.7 GB | 100 trained per-shape hypernetworks (full weight space) |
image_sirens/ |
~629 MB | 2400 image-SIREN files (24 views × 100 shapes) |
checkpoints/ |
~220 MB | VAE, direct-mapper, and h90 variant checkpoints |
diagnostics/ |
~1 MB | Training diagnostic plots and metric dumps |
hypernet_meshes/ |
~188 MB | Sample reconstructions from trained hypernets |
Note on naming
The corresponding folder in the source repo was named sdf_samples_old10;
it has been renamed to sdf_samples here for clarity. The _old10 suffix
was a local distinguisher when this codebase ran alongside other variants.
Citation
Jain, A. (2026). Hypernet research archive — prior topic 03 (per-shape hypernetworks
and SIREN decoders). HuggingFace: bobthebuilderinternational/hypernet-prior-topic03.
License
Apache 2.0. Underlying meshes are derived from publicly available sources under their own licenses.