Breaking

MVInverse infers materials and geometry from multi-view images in seconds, accepted to CVPR 2026

July 1, 2026 at 13:16 EDT

Research & Papers
Media Generation
Open Source

MVInverse, an inverse rendering method that predicts scene geometry and physically based materials from multiple RGB views in a single forward pass, has been released. Its arXiv paper and GitHub code appeared in December 2025, and the work has been accepted to CVPR 2026.

CVPR 2026 · Tsinghua × HKUST

MVInverse: Feed-Forward Multi-View Inverse Rendering in Seconds

Recovers geometry and physically based materials from multiple images or video in a single inference pass — no per-scene optimization. Code, models, and a runnable demo are openly available.

~10 fps

Material estimation speed

1 pass

Single feed-forward inference — no iterative refinement

Intrinsic maps output per view

TWO OLD APPROACHES vs MVInverse

Relative processing speed — taller column means slower

Slow

Multi-view optimization
(per-scene refinement)

Inconsistent

Single-view methods
(ignore inter-view links)

Seconds

MVInverse
feed-forward, consistent

PIPELINE

DINOv2

patch tokens + alternating attention

→

ResNeXt

multi-resolution features

→

DPT head

dense intrinsic maps

albedo roughness metallic diffuse shading surface normals

SEEN AS NOTEWORTHY

DINOv2 + ResNeXt features with video consistency training; resolves the previously problematic cross-view albedo drift. Claims state-of-the-art results on Mip-NeRF 360, Tanks and Temples, and DL3DV.

OPEN QUESTIONS

Production limitations and commercial-use availability are not spelled out — this is research code, leaving those points for practical verification.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.