Davd

Procedurally generated 4D geometry for multiview reconstruction, camera inference, surface/interior/air trajectories, normals and feature supervision. Generation target: 6,000 accepted physical scenes. This is an incremental release. Read index.json for the actual verified scene count and available shards; the target is not a claim that generation has finished. Only indexed shards are ready. The release contains original procedural assets, losslessly compressed authoritative motion state, camera/appearance recipes, and 768-dimensional procedural feature programs. It does not persist every dense image, projected feature tensor or query trajectory. Replay renders RGB, metric depth/range…

Gated internal mirror. Each dataset lives under a neutral code (data-NNN/). Clip shards follow the p4d-1 / p4d-1.2 WebDataset format (one tar member per frame and modality,.idx.json member offsets for range reads), with a per-dataset data-NNN/meta.tar holding dataset.json (splits, licenses/provenance, conventions, depthsemantics, quality, multiview), index.parquet (one row per clip) and the shard indices. The exception is data-013, which holds per-(sequence, camera) source units read by its own streaming loader. multiplies the per-pixel depth regression weight (lossweights(status, confidence=...)); see dataset.json modalities.depthconfidence.rule / calibration. resolution as depth) on real…

Public browser assets for the C4D scene viewer, including the wider-family gallery. The full training exports are in smerkd/c4d-scenes. This repository contains browser previews and display-mapped supervision examples, not full training supervision. Browser videos are compressed previews; training RGB is stored losslessly in the scene archives. Gaussian displays use a bounded subset with experimental, uncertified covariance/opacity conventions. Preview assets are checksum-verified and gallery URLs refer to immutable revisions.