cyjcyj91/Hypersim-Distractor

Hypersim Synthetic Distractor Pairs

Clean/distractor image pairs for multi-view feature restoration, generated by compositing selected Objaverse objects and their estimated shadows into Hypersim indoor views. This is a derived dataset, not an official Hypersim or Objaverse release.

Release contents

Packaged pairs: 20,965 (train 20,885, eval 80).

Archives: 165; total 42.22 GB (decimal). Inventory updated: 2026-09-14T20:58:19.985376+09:00.

Collection	Pairs
scenes_002	3,283
scenes_003	3,445
scenes_004	3,141
scenes_005	3,017
scenes_006	1,256
scenes_v2	6,823

summary.json is the current cumulative inventory. manifest.jsonl has one record per packaged pair, including its archive, scene/camera/frame, split, asset UIDs, generation version, and SHA256 of each member. archive_index.json lists archive sizes and checksums. No image data is downloaded automatically from this card.

The first snapshot was frozen on 2026-09-14 at approximately 20:38:30 Asia/Seoul. Synthesis continues independently. Each snapshot lists the exact included frame IDs, not a maximum frame number: earlier missing frames can complete in a later batch. The legacy scenes collection is excluded to avoid duplicate versions of _001 data.

Files and layout

TAR archives are grouped by collection and scene, then timestamped. TAR is uncompressed because PNG is already compressed; the PAX format supports large files and long paths. These are ordinary TAR archives, not WebDataset shards. Download and extract them for the existing GARD filesystem loader.

archives/scenes_006__ai_001_006__<batch>.tar
  hypersim_pairs/scenes_006/ai_001_006/cam_00/
    clean_multiview.json
    frame.0000/
      clean.png
      distractor.png
      mask_object.png
      mask_shadow.png
      metadata.json
      complete.json

• clean.png: tone-mapped clean Hypersim color; it may differ in exposure from the official preview JPEG.
• distractor.png: composited RGB with inserted objects and estimated shadow effects, using the clean image's fixed tone mapping.
• mask_object.png: raster mask of visible inserted objects.
• mask_shadow.png: shadow mask, using the stored relative luminance-change threshold (typically 25%), excluding object coverage. It is not a ground-truth mask of all possible appearance changes.
• metadata.json: seed, asset IDs, placement/scale/orientation, camera and lighting settings, coverage metrics, and generation policy where present. Original local path strings are provenance, not required download paths.
• complete.json: synthesis completion marker.
• clean_multiview.json: clean-geometry visibility candidates for the camera trajectory. It can reference frames absent from a release; the GARD manifest builder filters candidates to downloaded pairs.

The training subset omits raw Hypersim HDF5, GLB meshes, render logs, instance IDs and shadow-strength maps. The included metadata is useful for tracing generation, but exact regeneration additionally requires the corresponding source data/assets and generation code/version. A few original review.jpg files are included under previews/ as examples.

Generation

Blender/Cycles renders inserted Objaverse assets against a proxy surface reconstructed from Hypersim ground-truth geometry. Illumination is approximated from Hypersim diffuse illumination with ambient/directional components; it does not reproduce the original V-Ray scene lights. The composite is formed in linear HDR and tone-mapped with shared exposure for the pair.

The collections contain multiple generation versions; use each pair's metadata as authoritative. Recent _002–_006 settings use 1–3 objects, uniform proposals over the prepared indoor asset pool, a 1.5 scale multiplier for small assets, and both in-scene support surfaces and an assumed floor extension outside the camera view. Partial border occlusions are included. Acceptance typically requires 5–30% visible object coverage with additional major-scene occlusion checks. Proposal sampling is uniform; accepted asset frequencies need not be uniform. Earlier scenes_v2 results can use different asset and placement policies.

Objects are placed per frame; this release does not guarantee the same distractor identity/pose across neighboring views. Clean background geometry supplies the multi-view relationship. Synthetic shadows, lighting and unseen supporting geometry are approximations.

Train/evaluation split

All packaged ai_001_001 views are evaluation-only; all other scene IDs are training. This is the GARD experiment split, not the official Hypersim split. Keep the same complete held-out image set, covisibility cache and seed for stable evaluation groups. Training and evaluation are separated by scene ID, but other scenes in the same Hypersim volume remain eligible for training.

The GARD recipe samples an anchor and overlapping neighboring views from a camera trajectory (directional clean GT covisibility >= 0.25), with input distractor probability 0.7 independently per view and clean targets. Training uses 1–4 views; evaluation uses fixed 4-view groups. This recipe is generated after downloading, rather than baked into the TAR layout.

Download and use

Once this directory is uploaded to a Hugging Face dataset repository, substitute its actual ID below:

pip install -U huggingface_hub
hf download YOUR_ACCOUNT/YOUR_DATASET --repo-type dataset --local-dir /datasets/hypersim_release
cd /datasets/hypersim_release
sha256sum -c checksums.sha256
python tools/extract_release.py --destination /datasets

Then, from GARD-DFrecon:

export HYPERSIM_PAIRS_ROOT=/datasets/hypersim_pairs
python scripts/check_hypersim_pairs.py
CUDA_VISIBLE_DEVICES=0,1 bash run_scripts/train/train_GARD_hypersim_pairs.sh

For training, include scenes_v2 with its evaluation scene as well as any additional collections. No trained checkpoint is included.

Incremental packaging and verification

See PACKAGING.md. Archives and per-archive receipts are append-only. New runs create timestamped delta TARs containing only newly completed pair directories. All archive versions are needed for the cumulative release. SHA256 manifests, exact frame inventories and frozen snapshots record the packaging boundary. Packaging never changes the synthesis output.

Sources, attribution and licenses

The Hypersim dataset is released under CC BY-SA 3.0. Its authors are Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb and Joshua M. Susskind (ICCV 2021). This release modifies its appearance through tone mapping, inserted objects and estimated shadows.

Objaverse 1.0 has database-level ODC-By terms and individual Creative Commons asset licenses. Individual asset attribution is in attribution/objaverse_assets.json; see LICENSE_NOTES.md for the observed license inventory. No blanket permissive license is assigned here to override underlying asset terms. Refer to the per-asset licenses and the Hypersim terms when distributing or reusing derived images.

Generation code: hypersim-synthesis. Training code: GARD-DFrecon. Current public generation defaults are not a guarantee of pixel-identical reproduction of every historical collection.

Please cite Hypersim (Roberts et al., ICCV 2021) and Objaverse: A Universe of Annotated 3D Objects (Deitke et al., CVPR 2023), in addition to identifying the exact release snapshot used.