跳到正文
DEV Community· Fathom News·· 10 天前AI 評分48

Nvidia Déjà View 重用單一層,媲美參數量大 10 倍的模型

Nvidia's Tiny 3D AI Matches Models 10 Times Its Size By Reusing One Layer

AI 導讀

Nvidia 研究人員打造 Déjà View,以 117 million 個參數將不同角度拍攝的照片轉換成 3D 場景,並在重用同一層的設計下,媲美或超越參數量為其 8 至 10 倍的模型。模型在五項 3D 重建基準測試中,平均相機位置準確度領先,所需運算能力亦比測試中的最大競爭模型少 1.9 至 2.3 倍。論文獲選為 NeurIPS 2026 口頭報告,Nvidia 已釋出程式碼。

正文

Cover image for Nvidia's Tiny 3D AI Matches Models 10 Times Its Size By Reusing One Layer

Fathom News

Nvidia researchers have built Déjà View, a 117 million-parameter AI model that turns photos taken from different angles into 3D scenes and matches or beats models up to 10 times its size by reusing a single layer over and over.

Key Takeaways

  • Déjà View turns multi-angle photos into 3D geometry and camera positions with just 117 million parameters
  • It matches or beats models with 8 to 10 times more parameters on five 3D reconstruction benchmarks
  • It also uses 1.9 to 2.3 times less computing power than the largest rivals it was tested against
  • The paper is an oral presentation at NeurIPS 2026, and Nvidia has released the code

Nvidia's AI research team shared the model on Wednesday. The work comes from 12 researchers at Nvidia, ETH Zürich, the University of Toronto and the University of Modena and Reggio Emilia.

How One Layer Does The Work Of Many

Most 3D reconstruction models have grown past 1 billion parameters by stacking many different layers. Déjà View instead runs the same transformer block again and again on features from each photo, sharpening its estimate of the scene and camera positions with every pass.

Users can choose how many passes to run from one trained model, trading speed for accuracy depending on how much computing power they have. Nvidia says the looped design even beat an otherwise identical version with separate weights for each step, trained on the same data with the same compute.

Also Read: Humanoid Robots Running Google’s Gemini Just Did Factory Work At Mercedes-Benz, And Got 96% Right

How It Stacks Up

On five benchmarks covering indoor rooms, outdoor scenes, single objects and driving footage, Déjà View led on average camera pose accuracy while being the smallest model tested. Its rivals included Depth Anything 3 at 1.2 billion parameters, VGGT at 1.26 billion and π³ at 959 million.

Read Next: Meta’s AI Code Checker Was Barely Better Than Chance, Then Its Engineers Tuned It


Originally published on Fathom News.

來源:DEV Community · dev.to