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

Hugging Face Daily Papers Papers

Summary

The paper introduces GAE, a geometry-native autoencoder that creates a compact latent space for generating 3D-consistent scenes, enhancing visual quality and coherence over existing methods.

We present a compact geometry-native latent space as a shared foundation for perception and generation. Visual generators can produce photorealistic frames without preserving a consistent 3D scene. We argue that this is not only a modeling problem but also a representation problem: generators typically evolve appearance-centric latents, while perception models recover geometry in a semantically rich space that encodes cross-view structure. Rather than adding geometry as another output, we reparameterize a geometry foundation model's features into a compact latent space for generation. We realize this shift with the geometry-native autoencoder (GAE), whose latent is jointly decodable to appearance, depth, cameras, and point maps. With this state, a standard conditional flow supports diverse generation tasks. In controlled comparisons that hold the generator and training protocol fixed, replacing the latent with GAE improves both visual quality and independently measured 3D coherence: FVD falls by 12.7% and 23.1% on RealEstate10K and DL3DV, and camera-trajectory error is halved on RealEstate10K. Together, these results show that the latent space is central to geometry-consistent generation and can serve as a shared interface between perception and generation.
Original Article
View Cached Full Text

Cached at: 09/23/26, 03:30 AM

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

Source: https://huggingface.co/papers/2609.24981

Abstract

Wepresentacompactgeometry-nativelatentspaceasasharedfoundationforperceptionandgeneration.Visualgeneratorscanproducephotorealisticframeswithoutpreservingaconsistent3Dscene.Wearguethatthisisnotonlyamodelingproblembutalsoarepresentationproblem:generatorstypicallyevolveappearance-centriclatents,whileperceptionmodelsrecovergeometryinasemanticallyrichspacethatencodescross-viewstructure.Ratherthanaddinggeometryasanotheroutput,wereparameterizeageometryfoundationmodel’sfeaturesintoacompactlatentspaceforgeneration.Werealizethisshiftwiththegeometry-nativeautoencoder(GAE),whoselatentisjointlydecodabletoappearance,depth,cameras,andpointmaps.Withthisstate,astandardconditionalflowsupportsdiversegenerationtasks.Incontrolledcomparisonsthatholdthegeneratorandtrainingprotocolfixed,replacingthelatentwithGAEimprovesbothvisualqualityandindependentlymeasured3Dcoherence:FVDfallsby12.7%and23.1%onRealEstate10KandDL3DV,andcamera-trajectoryerrorishalvedonRealEstate10K.Together,theseresultsshowthatthelatentspaceiscentraltogeometry-consistentgenerationandcanserveasasharedinterfacebetweenperceptionandgeneration.

View arXiv pageView PDFProject pageGitHub120Add to collection

Get this paper in your agent:

hf papers read 2609\.24981

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.24981 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.24981 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.24981 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

@zhiwen_fan_: paper from dust3r’s team

X AI KOLs Timeline

A paper from the DUSt3R team proposes sparse auto-regressive modeling for 3D scene generation from multi-view images, using a voxel-aligned 3D latent space and an occupancy-aware masked autoregressive transformer.