Source
Wang et al. (2023), LIT-385 — ARXIV-2312.14132 (DUSt3R). Wang et al. (2025), LIT-384 — ARXIV-2503.11651 (VGGT).
The dependency being inverted
Classical multi-view reconstruction has a fixed order. Calibration and pose come first, because triangulation is defined in terms of them — DUSt3R’s own words are that they are “mandatory to triangulate corresponding pixels in 3D space, which is the core of all best performing MVS algorithms”.
So the pipeline’s stages are not a decomposition somebody chose for convenience; each one exists because the next one cannot start without it. That is what makes this a reframing rather than a better solver: regress the 3D structure directly, and recover matches, relative and absolute camera from the output. Everything the old order required in advance is available afterwards.
The inversion is first made in LIT-385 (DUSt3R), which casts pairwise reconstruction as pointmap regression with no calibration or pose supplied, and shows pixel matches and relative and absolute camera being recovered from the regressed 3D output rather than demanded before it.
Two things fall out that the classical framing cannot express. Monocular and multi-view reconstruction become the same problem with different amounts of evidence, rather than different problems — a single view has no baseline to triangulate across. And the projective camera model stops being a constraint to satisfy and becomes a property the answer turns out to have.
Why two papers rather than one
DUSt3R shows the inversion works and still needs an explicit global alignment step once there are more than two images. VGGT removes it: one feed-forward pass, one to hundreds of views, producing camera parameters, point maps, depth maps and 3D point tracks together — and beating methods that post-process with geometry optimisation, in under a second.
That second step is LIT-384: it reports state of the art across camera parameter estimation, multi-view depth, dense point cloud reconstruction and 3D point tracking from the one network, and shows pretrained VGGT improving downstream non-rigid point tracking and feed-forward novel view synthesis when used as a backbone.
That last part is the strong claim. The iterative refinement that justified the classical pipeline is not, on these benchmarks, buying what it was thought to be buying.
Conditions
This is emerging, not settled. The classical pipeline is decades of tooling, well understood failure modes, and what most production reconstruction still runs on. Nothing in this record measures where the crossover is, and “a forward pass beat optimisation on these benchmarks” is not “optimisation is obsolete”.
The guarantees are different in kind. A triangulated point has a geometric justification and a residual you can inspect; a regressed pointmap has neither. Where a reconstruction has to be certified rather than merely used — survey, metrology, anything with a tolerance — that difference is the whole question, and neither paper addresses it.
Scale of evidence. Two groups. The second’s backbone result — pretrained VGGT improving downstream tracking and view synthesis — is the best available evidence that something geometric was learned rather than four output heads fitted, but it is still one team’s finding.
What this gives the record
The geometry half of 3D vision, which #154 found was entirely absent. The anthology already recommends how to represent a reconstructed scene (SOTA-205, from LIT-108, LIT-086, LIT-109, LIT-064) and said nothing about how the scene got reconstructed — so a reader arriving with a pile of images found four papers assuming the poses they came to ask about.
It also gives LIT-111 (OnePose) its lineage. That reading builds an SfM point cloud and matches against it with a generic network; until now this record held that machinery only as a step inside somebody else’s method, never as a subject with papers of its own.