bobthebuilderinternational/hypernet-image-to-3d-data

Hypernet Image-to-3D Dataset

Companion dataset to the hypernet and image-to-3d-deepsdf repositories.

Contents

Path	Size	Description
manifest.json	<1 MB	obj_idx ↔ uid ↔ LVIS category mapping for 1000 shapes
views.json	<1 KB	64 Fibonacci-sphere camera poses
watertight/	~21 GB	985 watertight .obj meshes (filename = uid hash)
sdf_samples/	~3 GB	976 .npz files, each with 200K (point, sdf) pairs
multiview/	~1.3 GB	976 .tar files; each tarball contains 64 PNG renders for one shape
dino_features.pt	184 MB	DINOv2 ViT-B/14 CLS features for all (shape, view) pairs, shape (976, 64, 768)
checkpoints/phase1_pershape/	~21 MB	199 per-shape SIREN decoders + anchor (used by phase 3-5)
checkpoints/phase1_relu_pershape/	~422 MB	976 per-shape ReLU+PE decoders + anchor (used by phase 6-10)
checkpoints/phase3_hypernet/	~125 MB	Canonical trained hypernets (Topic 03 prior work)
checkpoints/phase11_deepsdf/	~7.7 MB	★ Working: shared DeepSDF decoder + 976 per-shape latents
checkpoints/phase12_image_to_latent/	~61 MB	★ Working: image-conditioned latent DiT

Total: ~26 GB

Why multiview is tarred

The 62,464 individual PNG renders (976 shapes × 64 views) are bundled into one tarball per shape (multiview/obj_NNNN.tar) to keep the file count manageable for HuggingFace's per-commit limits. To use them as loose PNGs:

# Download the multiview folder (~1.3 GB)
hf download bobthebuilderinternational/hypernet-image-to-3d-data \
    --repo-type dataset --local-dir data/ \
    --include "multiview/*"

# Extract all tarballs into a single multiview/ directory of loose PNGs
mkdir -p data/multiview_extracted
for tar_file in data/multiview/*.tar; do
    tar xf "${tar_file}" -C data/multiview_extracted/
done

Each tarball follows the original filename convention obj_NNNN_view_KK.png for K = 0..63.

Source

Shapes are drawn from a 1000-shape curated subset of Objaverse, filtered to the LVIS category vocabulary. After watertight conversion (Houdini VDB pipeline) and SDF sampling, 976 shapes had clean output across both stages and form the working set.

Working pipeline

input image(s)
    → DINOv2 ViT-B/14 (CLS features per view)
    → DiT (checkpoints/phase12_image_to_latent/dit.pt)
    → 64-dim shape latent z
    → DeepSDF shared decoder (checkpoints/phase11_deepsdf/shared_decoder.pt)
    → SDF
    → marching cubes
    → output mesh (.obj)

How to use

Download checkpoints only (~70 MB, sufficient for inference)

hf download bobthebuilderinternational/hypernet-image-to-3d-data \
    --repo-type dataset --local-dir data/ \
    --include "checkpoints/phase11_deepsdf/*" \
              "checkpoints/phase12_image_to_latent/*" \
              "manifest.json" "views.json"

Download full dataset (~26 GB)

hf download bobthebuilderinternational/hypernet-image-to-3d-data \
    --repo-type dataset --local-dir data/

Single-file load (Python)

from huggingface_hub import hf_hub_download
ckpt_path = hf_hub_download(
    repo_id="bobthebuilderinternational/hypernet-image-to-3d-data",
    filename="checkpoints/phase11_deepsdf/shared_decoder.pt",
    repo_type="dataset",
)

What's NOT included

The mode-collapsed DiTs from phases 6-9 (the failed weight-space diffusion attempts) and the broken weight autoencoder from phase 10 are intentionally excluded — they only produce broken outputs. The diagnostic results from those experiments are documented in docs/thesis.docx of the hypernet repo.

Citation

Jain, A. (2026). Image-to-3D via DeepSDF Shared Decoder + Latent Diffusion at Small Data Scale.
GitHub:      BOB-THE-BUILDER-in/hypernet
HuggingFace: bobthebuilderinternational/hypernet-image-to-3d-data

License

Apache 2.0. The underlying Objaverse meshes are released under their own licenses; see Objaverse license terms.