3D-WM Atomic v3: Synthetic CAD Construction Videos with Atomic Action Operations

Synthetic CAD construction videos are paired with per-frame atomic action operations for training video-to-action and image-to-video world models. Version 3 extends the v2 dataset with batches b07 through b14, maintaining an identical format. The dataset was created by author hz6666 and was last updated on June 4, 2026.

Use Cases

• Train video-to-action (inverse-dynamics) models based on the per-frame atomic action operations.
• Train image-to-video world models based on the synthetic CAD construction video sequences.
• Benchmark model performance on translation-canonicalized data where the final mesh bounding box is centered at the world origin.
• Experiment with raw float-millimeter operation parameters without quantization as described.

Strengths

• Dataset is an extension, adding batches b07-b14 to the previous v2 version (batches b01-b06).
• Data format is specified as 'v2-continuous-mm', indicating raw float-millimeter parameters without quantization.
• Clips are translation-canonicalized, a specific preprocessing step mentioned in the description.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

hz6666 on Hugging Face

Collection Method

Synthetically generated.

Freshness

Last updated 2026-06-04 18:06:35; freshness should be verified.