NeumaTex¶

Development line: project:neumatex · thread neumatex-development
Last event: 2026-07-02 · 1 dated since 2026-07-02 · Researched: 2026-09-05 · confidence: medium

What it is¶

NeuMatEx — a research pipeline for graphics and 3D-asset teams that turns a known 3D model plus multi-view images into a relightable neural material rather than PBR texture maps.

• predicts base color, neural-material latents, and uncertainty with the Large Material Reconstruction Model (LMRM)
• refines the result with uncertainty-guided inverse path tracing
• models clearcoat, haze, dust, fuzz, and scattering beyond fixed PBR lobes

Development line¶

• 2026-07-02 — NeumaTex project page shared. LMRM initialization followed by uncertainty-guided inverse rendering, targeting relightable materials whose specular effects would otherwise be baked into PBR base color.

What changed¶

2026-07-02 — NeuMatEx was presented as a multi-view neural-material extraction method: LMRM initialization followed by uncertainty-guided inverse rendering, targeting relightable materials whose specular effects would otherwise be baked into PBR base color.

How to use this¶

As of 2026-07-02, practitioners should consult the NeumaTex project page before relying on or adopting the project; the sealed link alone does not establish specific capabilities or usage guidance.

1. Assess fit first: the documented method needs a supplied 3D model and a set of multi-view images, not one casual image. — https://nvlabs.github.io/neumatex/
2. For a research reimplementation, reproduce the two stages: predict base color, neural latents, and uncertainty with LMRM; then refine with uncertainty-guided Monte-Carlo inverse rendering. — https://arxiv.org/abs/2606.26715
3. Do not plan an ordinary install or inference run yet: the official project page currently links the paper, supplementary PDF, video, and BibTeX, but no source repository or model download. — https://nvlabs.github.io/neumatex/

Best practices¶

• Use calibrated multi-view observations and known geometry, poses, and lighting where possible; the paper notes inverse-rendering recovery is sensitive to ambiguous or poorly calibrated data. — https://arxiv.org/abs/2606.26715
• Keep the uncertainty-guided test-time optimization stage; the reported ablation shows it improves material decomposition over feed-forward initialization alone and reduces baked-in lighting. — https://arxiv.org/abs/2606.26715
• Check relighting results and inspect real-world, out-of-domain materials for residual specular artifacts, especially red tinting and artifacts in crevices. — https://arxiv.org/abs/2606.26715

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• No public source repository, checkpoint, license, hardware requirement, or reproducible runtime procedure was found on the official project page reviewed on 2026-09-05.
• The 2026-07-02 report said source code was not yet published; availability outside the sources checked here is unverified.
• The supplied display spelling is “NeumaTex”; the official project and arXiv title use “NeuMatEx.”
• event_findings: [{"event_date":"2026-07-02","finding":"The linked work’s official name is NeuMatEx, short for the paper Extracting Neural Materials from Multi-view Images; it is arXiv:2606.26715 by Kim Youwang, Jon Hasselgren, Peter Kocsis, Andrea Weidlich, Tae-Hyun Oh, and Jacob Munkberg. The method requires a 3D model plus multi-view images, rather than being a single-image texture generator.","source_date":"2026-06-25","source_url":"https://arxiv.org/abs/2606.26715"},{"event_date":"2026-07-02","finding":"A contemporaneous report adds the practical release boundary: source code had not been published, so the announcement did not make a runnable public package available.","source_date":"2026-07-02","source_url":"https://nnets.ru/news/nvidia-predstavila-neumatex-novyj-metod-3d-tekstur-po-foto"}]
• new_events: [{"date":"2026-06-25","finding":"The arXiv v1 preprint was posted, establishing the research method before its 2 July coverage: LMRM initialization plus uncertainty-guided inverse rendering for neural materials.","source_date":"2026-06-25","source_url":"https://arxiv.org/abs/2606.26715"}]

Sources¶

Agent brief¶

• Subject: project:neumatex, thread neumatex-development, 1 dated events 2026-07-02 → 2026-07-02.
• Practical note: As of 2026-07-02, practitioners should consult the NeumaTex project page before relying on or adopting the project; the sealed link alone does not establish specific capabilities or usage guidance.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.