Ani4D dataset (TBZ + Objaverse + Objaverse-XL, curated through 2026-11-09; 16 cameras per clip)

The curated articulated-character dataset behind Ani4D (PoseAE codec + VoxelFM), plus AniBench, the 50-clip held-out benchmark (TBZ 8 / Objaverse 29 / Objaverse-XL 13), under benchmark/<source>/. 10,921 kept assets: train 10,725 (TBZ 57 / Objaverse 2,861 / Objaverse-XL 7,807), val 133 (10 / 55 / 68), test 63 (8 / 42 / 13) = AniBench 50 + the clips held out with them (sharing an AniBench clip's source or rig); the split is poseae_split.json at the root, bench50.txt lists AniBench and bench_windows.tsv the clip whose evaluated window does not start at frame 0. Every kept asset is rendered from 16 cameras, every frame 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_2026-11-09.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, MERGE_ZERO_BONES.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_rgba.tar             THE conditioning render, camera 0: the levelled front view (az -90, el 0), 1024^2,
                                   one straight-alpha RGBA PNG per pose frame (frame_%04d.png), uncompressed tar -- 10,921 assets, 731 GiB
    multiview/view_01.mp4 .. view_15.mp4
                                   cameras 1..15 (ActionMesh's 16-camera layout, camera 0 being the front view): camera k at
                                   azimuth -90 + 22.5 k, elevation (20, 35, 5, 50)[k mod 4] degrees above the horizon, in the
                                   asset's view frame; the front view's scene and framing, 1024^2, every pose frame. Each MP4
                                   holds two LOSSLESS H.264 streams: stream 0 = the straight RGB (libx264rgb, -qp 0), stream 1 =
                                   the alpha (gray, -qp 0), a keyframe every 16 frames -- 10,921 assets, 1.50 TiB
    multiview/cameras.json         the clip's cameras: layout, resolution, lens / sensor / field of view, framing (union-bbox
                                   centre and diagonal, fill 0.8, headroom 1.15), view frame, source frame window, per view
                                   azimuth / elevation / location / matrix_world (Blender world, Z up) / file / frames
    frontview.tar                  the older OPAQUE render of camera 0 (the same scene over a grey world), kept because the
                                   published models were trained on it; it equals the transparent render composited as below
                                   (not made for the Objaverse refilter clips added on 2026-11-04: transparent renders only)

Reading the views: ffmpeg -i view_01.mp4 -map 0:v:0 -f rawvideo -pix_fmt rgb24 - gives the RGB frames and -map 0:v:1 -pix_fmt gray the alpha, byte for byte the rendered PNGs (each MP4 was decoded and compared before upload). The opaque render is the transparent one composited onto Blender's grey world: ground = 171 + d, where d is Blender's fixed 8-bit output dither (grey_render_dither_1024.png at the root: uint8 d + 1, the same in every frame, clip and channel); opaque pixels are the RGB itself; a semi-transparent pixel is srgb(lin(rgb) a + lin(171) (1 - a)) + d (1 - a), blended in linear light as Blender does. Measured on 50 clips, DINOv2-giant tokens of that composite sit 0.014 relative L2 from the opaque render's (one frame of motion: 0.22).

The two PNG render directories ship as uncompressed tars, one per asset (tar xf frontview_rgba.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. The 15 extra views are one MP4 per camera (16 files per clip with cameras.json). 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; rest joints on the mesh; eight manual review rounds; a texture-patch audit (lost textures, see-through bodies; 2026-09-30); and a review of EVERY clip (2026-10-01): the maintainer's drop marks plus two Gemini passes -- a content review and a shading review (unlit / looks 2D) -- applied wherever the maintainer had not marked the clip Keep; and a consistency pass (2026-10-04): four drop marks made after the 10-01 application, a two-joint rig with a zero-length bone and a ten-vertex skinning test rig. Then (2026-11-02 to 11-04) an Objaverse-1.0 refilter: assets the original ingest had excluded, re-baked under loosened ingest rules, checked with the same curation rules and reviewed clip by clip; the kept clips joined the split (3 more left it because their pose was subsampled from a longer source than their renders show). A consistency pass (2026-11-04) made the inventories agree with the split: every clip outside it is excluded with its reason, including the refilter clips the review did not keep. On 2026-11-09 one Objaverse-XL clip left: the refilter had ingested the same file as an AniBench clip. Some renders carry explicit, per-asset repairs (a 3ds Max layered face decal, six TBZ animals' wing / hair card textures, a camera view frame for an asset whose source is not Z-up); the code repo's scripts/texture_repairs.json and scripts/render_views.json list them. Skeletons: one connected tree, no phantom leaves, root realigned onto the first skinned joint (UniMate's filters), and no redundant zero-length bones (2026-10-09): a joint at its parent's rest position is merged with it wherever the skinned mesh and every other joint stay put (760 clips, 2,533 joints; on the merge's own test, at most 1e-4 of the frame-0 diagonal for any vertex coordinate; MERGE_ZERO_BONES.json lists the merges, and those clips' motion.json is written for the merged skeleton, while rig.json keeps the source's); pairs that cannot be merged without moving the mesh or another joint stay (mostly two skinned joints turning about one point). Every benchmark skeleton is unseen in training: the TBZ part is two whole held-out animals (goat, seagull), and no training asset shares an Objaverse/XL benchmark asset's rig instance (joint set, hierarchy and rest bone lengths).

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