Read from LIT-384 — ARXIV-2503.11651.
What it adds over the paper before it
LIT-385 established that reconstruction without calibration or pose is possible, and left two things: an explicit global alignment step whenever more than two images are involved, and a pairwise formulation that scales awkwardly in the number of views.
This removes both. A single feed-forward pass, “from one, a few, or hundreds of its views”, producing camera parameters, point maps, depth maps and 3D point tracks together.
The two claims worth separating
That the tasks were never separate. The paper’s framing is that 3D vision models “have typically been constrained to and specialized for single tasks”. Camera estimation, depth, dense reconstruction and point tracking are four literatures because the pipeline had four stages, not because they are four problems. One network doing all four at state of the art is the argument.
That optimisation was not buying what it seemed to. Under one second, and still “outperforming alternatives that require post-processing with visual geometry optimization techniques”. This is the stronger claim and the more surprising one: the iterative geometric refinement that justified the classical pipeline is beaten by a forward pass.
The backbone result is the quiet one
Pretrained VGGT improves downstream non-rigid point tracking and feed-forward novel view synthesis. That matters for two reasons.
It is evidence about what was learned — four output heads bolted onto a shared trunk would not transfer; something geometric did. And it is the point at which geometry estimation joins the set of things this record treats as components rather than subjects. The corpus is full of papers that take a pretrained encoder for granted; this is the same move one level out.
What this does to the record
Sources SOTA-236 jointly with LIT-385. Two groups, two years apart, and the practice rests on the pair rather than on either — DUSt3R showed the dependency could be inverted, this showed the inversion scales and needs no optimisation to finish the job.
Not filed: the specific architecture, and the “under one second” figure as a target. The first dates immediately and the second is hardware-bound. What the practice records is the choice available to somebody holding a pile of uncalibrated images.