AdamWang0224/rendered_images

PhysXNet / PhysX-Mobility: renders, part assets and annotations

Multi-view renders, per-part meshes, preprocessed geometry, part masks and text annotations for the PhysXNet (PhysX-Anything) and PhysX-Mobility (PartNet-Mobility) objects, plus articulation-state augmented ("posed") copies of these objects.

Licensing. Everything here is derived from upstream datasets that keep their own terms: PhysXNet / PhysX-Anything, and PartNet-Mobility (SAPIEN), which restricts redistribution. Use this data only under those upstream licenses. Access is gated and intended for collaborators.

0. Conventions

• Data root. Every archive extracts relative to one data root, written $DATA. The layout under $DATA is the same as on the original cluster (where $DATA=/group/articulated-physx-gen/xiaoye). Section 1.2 says where each archive goes.
• Object ids. PhysXNet ids are integers (10017). Mobility ids are PartNet-Mobility ids (102588). Posed objects are <base_id>-s<k> (10017-s1): the same object re-articulated to a sampled joint state. A posed object has its own directory next to the original in every tree, so the posed archives extract into the same folders as the original ones.
• Renders are 1024×1024 RGBA PNGs (alpha = object mask), 25 random Hammersley views per object (000.png … 024.png), with cameras in transforms.json ({aabb, scale, offset, frames[25]{file_path, camera_angle_x, transform_matrix}}).

1. Original objects (PhysXNet 32,047 · Mobility 2,024)

File	Size	Contents	Objects / files
renders_cond.tar.zst	372.9 GB	PhysXNet 25-view renders	32,047 objects, 833,222 files (511.6 GB unpacked)
renders.tar.zst	22.9 GB	Mobility 25-view renders	2,024 objects, 52,624 files (32.0 GB)
physxnet_part_obj_with_tex.tar.zst	26.8 GB	PhysXNet semantic-part meshes with textures	32,047 objects, 287,958 files (94.1 GB)
partseg_new.tar	6.2 GB	Mobility raw part data: objs/ (part meshes + *.convex.stl), images/, part_obj_with_tex/	2,024 objects, 399,360 files (5.9 GB)

1.1 Layout inside each archive

renders_cond.tar.zst                 renders_cond/<id>/{000..024.png, transforms.json}
renders.tar.zst                      renders/<id>/{000..024.png, transforms.json}
physxnet_part_obj_with_tex.tar.zst   <id>/part_obj_with_tex/{*.obj, *.mtl, textures}     (no root directory)
partseg_new.tar                      partseg/<id>/{objs, images, part_obj_with_tex, part_renders_shared_views*}

1.2 Where to extract

mkdir -p $DATA/PhysX-Anything/preprocessed_data/renders_output $DATA/physx_mobility/PhysX_mobility $DATA/physxnet/version_1/partseg
tar -I zstd -xf renders_cond.tar.zst  -C $DATA/PhysX-Anything/preprocessed_data/renders_output   # -> renders_output/renders_cond/<id>/
tar -I zstd -xf renders.tar.zst       -C $DATA/physx_mobility/PhysX_mobility                      # -> PhysX_mobility/renders/<id>/
tar -xf partseg_new.tar               -C $DATA/physx_mobility/PhysX_mobility                      # -> PhysX_mobility/partseg/<id>/
tar -I zstd -xf physxnet_part_obj_with_tex.tar.zst -C $DATA/physxnet/version_1/partseg            # -> partseg/<id>/part_obj_with_tex/

2. Additions of 2026-10 (sharded, uploaded 2026-10-23)

70 tar.zst shards, 6,465,564 files, 1,458.5 GB unpacked / 736.0 GB on the Hub. Every shard extracts relative to $DATA:

tar -I zstd -xf <shard> -C $DATA

The one exception is G0_meta/repo_eval_splits_and_records_000.tar.zst, which extracts relative to the project code repository root. Shards of one unit never split an object directory, so you can extract any subset. MANIFEST/ has each shard's file list (lists/<shard>.list.gz), SHA256SUMS and inventory.json (files, bytes, sha256 per shard). Check downloads with sha256sum -c MANIFEST/SHA256SUMS.

Folder / unit	Shards	Files	Unpacked	On Hub	Contents → extracted path (under $DATA unless noted)
G0_meta/annotations_*	1	471,646	1.5 GB	0.2 GB	Text annotations PhysX-Anything/txt_rep_meshframe_v3_grid64/, txt_rep_mobility_all_meshframe_v3_grid64/; mesh tokens PhysX-Anything/3DVQVAE_mesh_token/{physxnet,mobility}/; raw annotations physxnet/version_1/{finaljson,bbox,bbox_shapellm_style}/, physx_mobility/PhysX_mobility/{finaljson,bbox,bbox_shapellm_style,urdf}/; posed-object metadata PhysX-Anything/pose_states/posed_meta/; camera files …/renders_output/cond_*rank*.json. Includes posed objects
G0_meta/part_boxes_2d_*	1	1,547,600	5.7 GB	0.5 GB	Per-view 2D part boxes …/renders_output/2D_seg_bbox/per_object/, physx_mobility/PhysX_mobility/2D_seg_bbox/per_object/ (originals + posed)
G0_meta/repo_eval_splits_and_records_*	1	35,741	1.0 GB	0.1 GB	Code-repo root: evaluation id splits, single-view GT part boxes, frame-correction tables, posed-object generation records
G1_jsonl/jsonl_v3_filtered_*	1	11	10.9 GB	0.1 GB	Conversation JSONL, original objects (785,275 rows / 31,411 objects) + val/test splits
G1_jsonl/jsonl_posed_filtered_*	1	6	9.9 GB	0.1 GB	Conversation JSONL, posed objects (695,825 rows / 27,833 objects)
G1_jsonl/jsonl_merged_pose4_*	1	10	20.8 GB	0.2 GB	Both of the above concatenated (1,481,100 rows) + val/test and posed test splits
G2_images/physxnet_renders_posed_*	16	598,338	363.2 GB	267.6 GB	…/renders_output/renders_cond/<id>-s<k>/, 23,013 posed PhysXNet objects
G2_images/mobility_renders_posed_*	4	135,675	80.4 GB	59.1 GB	physx_mobility/PhysX_mobility/renders/<id>-s<k>/, 5,025 posed Mobility objects
G2_images/physxnet_renders_fixed_view_*	1	64,312	21.1 GB	16.0 GB	…/renders_output/renders_cond_fixed_view/, fixed-view evaluation renders (32,156 objects incl. 109 posed)
G3_gt_geometry/physxnet_preprocessed_partseg_*	16	1,563,711	375.1 GB	87.5 GB	PhysX-Anything/preprocessed_data/tmp/partseg/<id>/, voxels 16/32/64 + merged meshes, 55,169 objects (originals + posed)
G3_gt_geometry/mobility_preprocessed_partseg_*	1	116,486	18.2 GB	3.9 GB	PhysX-Anything/preprocessed_data/tmp_mobility/partseg/<id>/, 7,049 objects
G3_gt_geometry/physxnet_raw_partseg_*	4	537,172	90.4 GB	22.7 GB	physxnet/version_1/partseg/<id>/{objs,imgs,…}, originals + posed; part_obj_with_tex/ is excluded (it is in section 1)
G3_gt_geometry/mobility_raw_partseg_posed_*	1	440,240	5.7 GB	0.9 GB	physx_mobility/PhysX_mobility/partseg/<id>-s<k>/{objs,images}, posed only (originals are in partseg_new.tar)
G4_part_masks/physxnet_part_masks_*	18	833,222	411.2 GB	250.9 GB	…/renders_output/rendered_partseg_replay_from_transforms/<id>/, 25-view part masks of the 32,047 original objects
G4_part_masks/mobility_part_masks_*	2	52,624	25.7 GB	14.8 GB	physx_mobility/PhysX_mobility/rendered_partseg_replay_from_transforms/<id>/, 2,024 original objects
G4_part_masks/part_masks_fixed_view_*	1	68,770	17.8 GB	11.3 GB	…/renders_output/rendered_partseg_fixed_view/ and the Mobility counterpart

…/renders_output = PhysX-Anything/preprocessed_data/renders_output. For a complete tree, extract section 1 first, then every shard of section 2 into the same $DATA. Posed objects and originals live in different directories, so nothing overwrites anything.

3. Needed for training but not uploaded, and how to reproduce it

Sections 1–2 hold everything that is expensive or impossible to regenerate. Training also needs a few derived files. They were left out because they are large, take only hours to rebuild, or depend on the software environment. The scripts named below live in the project code repository (shared with collaborators, commit 3e3c319). Their slurm launchers carry cluster-specific #SBATCH headers and default paths, so set ROOT_DIR / IMAGE_ROOT / OUTPUT_ROOT for your machine.

Item	Needed for	Size	Rebuild cost	Section
JSONL image paths pointed at your $DATA	training + eval	—	seconds	3.1
DINOv3 patch features for every render	training (and eval on fixed views)	2.57 TB	≈20 GPU·h (H200)	3.2
Tokenized dataset cache (LLaMA-Factory Arrow)	training	22–42 GB each	≈1 h on 32 CPU cores	3.3
GT depth maps	only depth-supervised variants	19 GB	CPU hours	3.4
25-view part masks of posed objects	only re-deriving posed 2D boxes	—	≈56 GPU·h	3.5
Annotations / JSONL / posed objects / part meshes from scratch	only if the uploaded ones are lost	—	CPU/GPU days	3.6

Already on the Hub and needed as-is: renders (sections 1–2), JSONL (G1_jsonl/), frame-correction tables and evaluation splits (G0_meta/repo_eval_splits_and_records_000, which extracts into the code repository), and the GT geometry and masks used by evaluation (G3_gt_geometry/, G4_part_masks/, section 1).

3.1 Point the JSONL at your data root

Every absolute path in the JSONL files (the images field) starts with /group/articulated-physx-gen/xiaoye. Either create that path as a symlink to $DATA, or rewrite it:

for d in $DATA/PhysX-Anything/runs/physx_attr_preprocess/*/llamafactory_data; do
  sed -i "s#/group/articulated-physx-gen/xiaoye#$DATA#g" $d/*.jsonl
done

Do this before building a cache (3.3). For a cache built with the old paths, scripts/rewrite_tokenized_image_paths.py --input-cache … --output-cache … --image-root-override … --mobility-image-root-override … rewrites it in place of a rebuild.

3.2 DINOv3 features (required for training)

One .pt file per rendered view, at the mirrored path <image_root>_dinofeat/<id>/<view>.pt (1.62 MB, bf16, shape (4, 196, 1024): ViT-L/16 hidden states of layers 4/11/17/23 at 224 px). Training finds a view's feature by appending _dinofeat to its image root, so write the features next to the renders:

Image root (under $DATA)	Feature root	Views	Size
PhysX-Anything/preprocessed_data/renders_output/renders_cond	…/renders_cond_dinofeat	1.43 M (originals + posed)	2.23 TB
physx_mobility/PhysX_mobility/renders	…/renders_dinofeat	0.18 M	0.29 TB
PhysX-Anything/preprocessed_data/renders_output/renders_cond_fixed_view	…/renders_cond_fixed_view_dinofeat	64 k (evaluation)	0.05 TB

# DINOv3 code: https://github.com/facebookresearch/dinov3 ; weights dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth (request access from Meta)
python data_preprocess/physxnet/dino_features/export_dino_features.py \
  --image-root physxnet_25view=$DATA/PhysX-Anything/preprocessed_data/renders_output/renders_cond \
  --output-root $DATA/PhysX-Anything/preprocessed_data/renders_output/renders_cond_dinofeat --mirror-image-root \
  --backend dinov3_local --dinov3-repo <dinov3 checkout> --dinov3-vit-type vitl \
  --dinov3-vitl-weights <dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth> \
  --layers 4,11,17,23 --image-size 224 --pool-patch-grid 0 --dtype bf16 --batch-size 64 \
  --skip-valid-existing --skip-existing --num-shards 5 --shard-index <i>

The same command with --image-root mobility=…/renders and physxnet_fixed=…/renders_cond_fixed_view covers the other two roots. Sharded slurm launchers: data_preprocess/physxnet/slurm_export_dino_features_{renders_cond,mobility_renders,renders_cond_fixed_view}_sharded_a100.sh (one array task per shard). Throughput is ≈83k views/h on one H200. The run is resumable (--skip-valid-existing re-does files that are missing or contain NaNs).

3.3 Tokenized cache (required for training)

LLaMA-Factory reads a pre-tokenized Arrow cache built from the JSONL and the base model's tokenizer. The cache depends on package versions, so pin transformers 5.2.0 / datasets 4.0.0 / accelerate 1.11.0 / peft 0.18.1 / trl 0.24.0 (pip install trl pulls datasets 5.x; reinstall 4.0.0 afterwards). With those pins, caches built on different machines are byte-identical. Build it CPU-only:

# DATA_TAG picks the JSONL: v3allclass25v (originals) or v3allclass25v_pose4 (originals + posed).
# ARM / ABL pick the tokenization variant; the current default is ARM=2 ABL=nodepth+nosil (cache "B").
ARM=2 ABL=nodepth+nosil DATA_TAG=v3allclass25v_pose4 \
  bash code_checks/mesh_query_latent_ablation/scripts/run_prepare_cache_cpu.sh   # or sbatch it

Before running it, three things must match your machine:

• ROOT_DIR at the top of the script must point to your checkout of the code repository.
• The micromamba env name is hard-coded as shapellm-omni-blackwell. Create or rename yours, or edit the check.
• The data symlink: runs/ in the code repository must link to $DATA/PhysX-Anything/runs, so the JSONL is found at runs/physx_attr_preprocess/<dataset>/llamafactory_data/.

The script sources code_checks/mesh_query_latent_ablation/scripts/ablation_env.sh for the data directory, cutoff (5120) and grid constants (BBOX_QUANTIZATION_BINS=64 SIMREADY_VOXEL_DEFINE=64 ARTICULATION_TEXT_CONVENTION=meshframe_v3_grid64). It writes <cache>.building, runs a gate check, then renames the cache into place. Measured: 785k rows ≈ 50 min, 1.48 M rows ≈ 65 min on 32 cores. Use one thread per worker (OMP_NUM_THREADS=1), because more threads per worker deadlock in libgomp.

3.4 GT depth maps (optional)

scripts/render_gt_depth_targets.py --datasets physxnet,physxnet_fixed_view,mobility --npz-root <out> --shard i/N ray-casts the GT mesh with open3d from each view's transforms.json. It writes <out>/<tree>/<id>/<view>.npz (504×504 float16 z-depth, hit mask, 18×18 token depth). It runs on CPU and self-checks the hit mask against the PNG alpha (IoU ≥ 0.97 expected). Only depth-supervised variants use it.

3.5 Part masks of posed objects (optional)

Masks for posed objects were deleted after their 2D part boxes (uploaded in G0_meta/part_boxes_2d_*) were exported. To recreate them, replay each object's cameras in Blender: data_preprocess/physxnet/4_render_cond_partseg.py --camera_mode replay_from_transforms (Mobility: data_preprocess/mobility/render_cond_mobility_partseg.py). Then export boxes with code_checks/partseg_sam_validation/scripts/export_gt_part_bboxes.py. Set BLENDER_THREADS=<n> and OMP/OIDN_NUM_THREADS, because Blender and OIDN size their thread pools by the machine's core count, not the job's CPU limit. Throughput is ≈500 objects/h per GPU.

3.6 Rebuilding upstream stages (only if needed)

Data	Scripts
meshframe_v3 annotations → JSONL	data_preprocess/physxnet/meshframe_v3/build_meshframe_v3_dataset.py (posed objects: --append), code_checks/mesh_query_latent_ablation/scripts/build_v3_filtered25v.sh (filter: ≤ 50 parts, turn-1 ≤ 4000 chars), code_checks/data_and_vqvae/articulation_state_aug/scripts/full_jsonl_sof1_cpu.sh (merge with posed)
Posed objects (re-sampling joint states)	data_preprocess/physxnet/pose_states/{pose_object,materialize_posed_object,make_posed_manifests,verify_posed_object,sweep_gate}.py, code_checks/data_and_vqvae/articulation_state_aug/scripts/full_*; gate rules in instruction_md/agent_prompts/20260929_articulation_state_augmentation_plan.md §3.4
Renders	data_preprocess/physxnet/4_render_cond.py, 4_render_cond_fixed_view.py, data_preprocess/mobility/render_cond_mobility.py (Blender; same BLENDER_THREADS caveat)
Voxels / merged meshes (tmp*/partseg)	data_preprocess/physxnet/1voxel*.py, 4_generate_merged_mesh.py
part_obj_with_tex	data_preprocess/per_part_render_preprocess/export_semantic_part_meshes.py, package_part_obj_with_tex.sh
Raw datasets	official PhysX-Anything (PhysXNet) and PartNet-Mobility releases

4. Verification (2026-10-23)

Each archive in section 1 was streamed and indexed member by member (path, size, md5) and compared with the cluster copy. All files matched by path and size, with nothing found only on the Hub. Random md5 samples matched in every archive: 1,000/1,000 for renders_cond.tar.zst, 400/400 for each of the others.

For section 2, each shard was checked twice. Before upload, its member count had to equal its file list. After upload, its size and sha256 on the Hub had to equal the local copy (70/70 passed, with no "file changed while reading" warnings). A round trip on one JSONL shard (download, extract, md5) reproduced the source file byte for byte.