Ani4D dataset (TBZ + Objaverse + Objaverse-XL, curated 2026-09-16)

The curated articulated-character dataset behind Ani4D (PoseAE codec + VoxelFM), plus the 50-asset Ani4D-vs-ActionMesh benchmark under benchmark/. 11,206 kept assets: train 11,008 (TBZ 58 / Objaverse 2,137 / Objaverse-XL 8,813), val 148 (8 / 60 / 80), test = benchmark 50 (10 / 20 / 20); the split is poseae_split.json at the root and bench50.txt lists the benchmark. Every kept asset has a complete front-view render that is 1:1 with its pose sequence, is normalised to the frame-0 bounding box (centred, longest side 1), and passes the curation rules below. Statistics: dataset_stats_2026-09-16.md.

Layout, one folder per asset (records and ready-to-train cache merged):

inventory_<dataset>.jsonl          per-asset scan record; bucket=excluded rows carry the curation reason
clean/<dataset>/<name>/
    rig.json, motion.json          bake records (skeleton + per-frame motion, source frame range)
    frame_0000_anchor.glb          rest-pose mesh (ground-truth topology); frame_0000_textured.glb, video*.mp4 where produced
    manifest.npz                   rest mesh + skeleton + skinning weights
    theta_motion.npz               per-frame joint parameters (theta: rot 3 | stretch 1 | trans 3 per joint), frame-0 units
    frames.npy                     per-frame ground-truth vertices [T, V, 3], frame-0 units
    meta.json, RENORM_FRAME0.json, REALIGN_ROOT.json, TRIMMED.json   provenance markers
    slat_frame0.npz                AniGen frame-0 structured latent (z_s, z_skin, z_skl on a 64^3 grid)
    pose_vox_*.npz                 pose-voxel pools of the codec
    frontview.tar                  the conditioning render: levelled front view (az -90, el 0), 1024^2,
                                   one PNG per pose frame (frame_%04d.png), uncompressed tar
    frontview_rgba.tar         RGBA twin of the render (the ActionMesh baseline's input) -- present on 2,354 assets, 149 GB
    video_feats_crop_g.npz     DINOv2-giant features of the subject crop -- present on 2,322 assets, 1.10 TB.
                               DERIVED: recomputable from frontview/ with scripts/tools/recache_video_crop.py
                               (--res 518 --model dinov2_vitg14_reg). They are ~half the download.

The two render directories ship as uncompressed tars, one per asset (tar xf frontview.tar): unpacked they are 1.69 M of the dataset's 1.83 M files -- 92 % of the file count for 36 % of the bytes -- and a HuggingFace repo is capped at 1,000,000 files. Packed, the dataset is ~159 k files at the same size. Everything else is stored as ordinary files.

Curation (applied in this order; every removal is recorded in the inventories): texture present and shaded; T >= 32; no idle clips (fastest joint p95 < 1 deg/frame and root travel < 0.5 bone lengths); no implausible motion (> 90 deg/frame, jerk spikes, root jumps); theta->GT reconstruction <= 5 % of the bbox diagonal; grey/magenta placeholder renders; single-joint rigs; vertex count <= 300k; blank renders (subject < 0.1 % of the frame); render frame count == T; plus eight manual review rounds. Skeletons: one connected tree, no phantom leaves, root realigned onto the first skinned joint (UniMate's filters). The benchmark's Objaverse/XL skeletons are unseen rig instances (joint set, hierarchy and rest bone lengths never seen in training); its TBZ clips hold out two whole animals (goat, seagull).

Sources & licensing -- derived from Truebones Zoo (TBZ), Objaverse and Objaverse-XL. The underlying assets keep their own licences; verify them before redistributing derived data.