AIAny
Icon for item

GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

Compresses a geometry-foundation model’s multi-level features into a compact latent that decodes jointly to RGB, depth, cameras and point maps — enabling a conditional flow to generate 3D-consistent video and novel views with measurably improved coherence.

Introduction

Most visual generators evolve appearance-centric latents and treat geometry as an afterthought; GAE flips this assumption by putting 3D inductive bias directly into the generated state. The core insight is that a shared, geometry-native latent — one that decodes natively to both appearance and geometry — makes it much easier for standard generative models to produce cross-view consistent output.

Key Findings
  • A two-stage recipe: (1) train a codec that compresses a frozen geometry backbone’s multi-level features into one compact bottleneck decodable to RGB and geometry; (2) freeze the codec and train a DiT-style conditional flow in that latent space. This separates representation design from flow training.
  • Quantitative gains: replacing a conventional latent with GAE reduces FVD by 12.7% on RealEstate10K and 23.1% on DL3DV, and halves camera-trajectory error on RealEstate10K — showing improvement in both perceptual quality and independent 3D coherence metrics.
  • Latent structuring matters: GAE organizes the bottleneck at token and relational levels (token alignment to local features and pairwise similarity matching to preserve relational geometry), forcing a single compact state to represent both appearance and cross-view geometry.
Who it's for & tradeoffs

Great fit if you train conditional generative models for camera-controlled video, reference-conditioned novel-view synthesis, or any task where cross-view geometric consistency matters. It’s also useful when you can afford a perception backbone and a two-stage training pipeline. Look elsewhere if you need a zero-shot, single-stage generator without a frozen geometry backbone, or when minimal implementation complexity is the top priority.

Where it fits

GAE is a representational complement to recent flow- and diffusion-based generators: instead of adding geometry outputs to an appearance-first latent, it reparameterizes a geometry foundation model’s features into a generator-friendly latent. Practically, this makes perception and generation share the same interface, improving consistency without changing the downstream flow architecture.

Information

  • Websitearxiv.org
  • OrganizationsThe Hong Kong University of Science and Technology, ARC Lab, Tencent IEG, The University of Hong Kong, The University of Texas at Austin
  • AuthorsJiahao Lu, Minghao Yin, Wenbo Hu, Hengyu Liu, Wang Zhao, Sai-Kit Yeung, Ying Shan, Yuan Liu
  • Published date2026/09/21

More Items

An all-in-one multilingual scene text recognition approach that pairs a shared visual encoder with a script-aware Mixture-of-Experts (ScriptMoE) decoder to route each image to top-2 script experts. Introduces TextMuSS-10M, a synthetic dataset covering 10 scripts and 229 languages, and reports state-of-the-art accuracy and large end-to-end OCR F1 gains while remaining parameter-efficient.

Converts natural-language SVG instructions into instance-aware six-item rubrics and uses a vision–language judge to score rendered SVGs item-by-item as dense RL rewards. Produces stronger human alignment and lifts benchmark rubric scores substantially on MMSVG datasets.

Provides a unified data and evaluation suite for measuring gameplay capabilities across short, medium, and long temporal horizons. Includes an automated annotator, a 5,000-hour AAA gameplay dataset with temporally aligned frames, actions and multi-horizon instructions, and a reproducible offline + stepwise online benchmark.