hz6666/3d-wm-atomic-v2

3d-wm-atomic-v2 — Synthetic CAD construction videos

Per-op animated frame sequences for CadQuery construction programs, with per-frame atomic op labels in continuous raw mm. Designed as training data for video → action (IDM) and image → next frame (video-gen) models.

Format version: `v2-continuous-mm` (2026-05-29 onward).

One clip = one (case, view) pair

data_*/{bNNNN}/train/shard-NNNNNN.tar
└── {case_uid}_v{NN}/
    ├── 0000.png .. NNNN.png   # 256x256 rendered frames
    ├── meta.json              # all metadata + per-frame op_spans
    ├── ops.json               # full atomic op sequence
    └── frame_codes.json       # per-frame partial CadQuery code

A case is one CAD program rendered from 5 sibling views (iso, front, side, top, random) sharing the same anchor — physical mm/pixel identical across the 5 views, only the camera angle changes.

Atomic op vocabulary (continuous mm, no quantization)

Op IDs are STABLE. Sparse IDs (0, 1, 2, 5, 6, 7) are reserved.

Categorical alphabets

• —plane_idx → ["XY", "YZ", "ZX"]
• —extrude_kind → ["fresh", "union", "cut"]
• —edge_idx → ["|Z", "|X", "|Y"]

Example op tuple (from a real clip)

[8, 1, -30.0, 0.0, 0.0, 16.0, -20.0]   # SKETCH_BEGIN plane=YZ origin=(-30,0,0) start=(16,-20)
[15, 30.0]                              # SKETCH_CIRCLE r=30.0
[12]                                    # SKETCH_CLOSE
[13, 32.0, 0]                           # SKETCH_EXTRUDE h=32.0 kind=fresh
[14]                                    # END_SKETCH

meta.json key fields

op_spans per-frame entry

Each frame has exactly one span. For animated extrude frames, action_op contains SKETCH_EXTRUDE with h = full_h × op_progress so each frame carries a distinct numeric target.

{
  "frame_idx": 5,
  "op_idx": 4,
  "op_type": "SKETCH_EXTRUDE",
  "op_progress": 0.864,
  "kind": "extrude",
  "keyframe": false,
  "action_op": [13, 27.65, 0],
  "action_op_human": "SKETCH_EXTRUDE(h=27.65, kind=fresh)"
}

op_progress is the fraction of the full extrude already drawn at this frame. Mostly in (0, 1.0]. With probability ~30% the second-to-last extrude frame overshoots to (1.0, 1.3] — explicitly trained-on negative-feedback signal for inverse-dynamics models.

Translation canonicalization

The renderer shifts each case so the final-mesh bbox is centered at world (0, 0, 0) before generating PNGs and actionop. Two cases that differ only by world translation produce identical PNGs + identical ops.json + identical actionop tuples, with the difference recorded ONLY in canonicalize_offset_mm. Recover original world coords with:

original_world_xyz = canonical_xyz + meta["canonicalize_offset_mm"]

No rotation / scale canonicalization is applied — those would be lossy or brittle.

Coordinate frames

• —Continuous params (length / coord) are in canonical-frame mm (canonicalization above).
• —world_bbox_* fields are also in canonical frame.
• —cam_elev_azim_deg is spherical: elev measured from XY plane, azim from +X axis CCW around +Z. front_dir = (cos·elev·cos·azim, cos·elev·sin·azim, sin·elev). Camera position = bbox_center + camera_distance × front_dir, looking back at bbox_center.

Per-clip vs dataset-wide normalization

Default: per-clip auto-fit (each clip fills ~62% of the frame). To recover physical mm from a PNG pixel:

mm_per_px = meta["approx_mm_per_pixel"]
mm_offset_from_image_center = pixel_offset × mm_per_px

If meta["reference_bbox_diag"] is non-null, the renderer forced a constant world_diag_mm across the dataset → constant mm/pixel.

Sibling views share scale

For each base_uid, all sibling clips {base_uid}_v00..vNN share exactly the same world_bbox_*, world_diag_mm, and approx_mm_per_pixel. Only cam_elev_azim_deg differs. Use this contract to train view-invariant features.

Recommended downstream normalization recipe

# IDM training (predict action_op given video frames):
half_extent = max(
    max(abs(c) for c in meta["world_bbox_min"]),
    max(abs(c) for c in meta["world_bbox_max"]),
)
def normalize_coord(mm): return mm / half_extent          # -> [-1, 1]
def normalize_length(mm): return mm / (half_extent * 2)   # -> [0, 1]

Apply the inverse at sampling / inference time.

For video frames: standard (x / 255 - mean) / std with ImageNet stats or per-dataset stats.

Format transition note

This dataset was migrated to v2-continuous-mm on 2026-05-29. Earlier flat prefixes have been removed. The data_hula_b01..b06 prefixes are being progressively re-rendered and overwritten to v2; until that chain finishes, a minority of bNNNN shards may still carry the older bin-quantized integer encoding (length [0,60] mm @ 128 bins, coord [-30,30] mm @ 128 bins). Always branch on meta["format_version"]: trust only shards whose meta/ops report "v2-continuous-mm".

Source

Cases are parsed from `filapro/cad-recode` training splits and re-rendered with per-op animation.