Tencent ARC Releases GAE: Geometry-Native Latent Space for 3D-Consistent World Generation

Tencent ARC, with CUHK members, released GAE (Geometry-native Autoencoder) with open weights on Hugging Face. Its central claim: 3D-inconsistent generation stems from representation, not modeling — existing generators evolve appearance-centric latents, while perception models recover geometry only in semantic spaces encoding cross-view structure. GAE compresses features from geometric foundation models into a shared compact latent space so generation happens natively in that space.

Confirmed

GAE compresses 3072-channel geometric features into a compact 128-channel latent representation
Conditional flow models generate directly in this geometry-native latent space; the same latent decodes into both RGB appearance and 3D point-cloud geometry, yielding 3D-native generation rather than post-hoc 3D reconstruction from RGB
Moving geometric representation learning into the autoencoder lets DiT focus on distribution modeling; the resulting latents act as a flexible 'soft skeleton' for persistent worlds
Results: FVD reduced by 23.1%, camera error roughly halved
Weights (GAE-D64-1B) uploaded to Hugging Face; code and models open-sourced
GAE is positioned as a foundational building block for persistent world models, the first to embed a compact geometry-native autoencoder directly into world modeling

Why it matters

It reframes 3D consistency from a modeling patch to a representational redesign, offering a new latent-space paradigm for world models
One latent producing both appearance and geometry unifies video and 3D point-cloud generation, with potential value for 3D content creation and embodied-AI simulation
Open weights lower the barrier for the community to reproduce and build generative world models on top

2026-09-22 ~ 2026-09-24 · 7 related posts

Primary sources

Tencent ARC's GAE: geometry-native latent space cuts FVD by up to 23% in world generation — TencentARC · 2026-09-23 11:17 (UTC+08:00)
GAE: A Geometry-Native Autoencoder Cuts World Model FVD by 23.1% — yshan2u · 2026-09-22 17:26 (UTC+08:00)

[source] GAE: A Geometry-Native Autoencoder Cuts World Model FVD by 23.1% — yshan2u · 2026-09-22
GAE generates video and 3D geometry natively in a shared geometry latent space, code released — yshan2u · 2026-09-22
[source] Tencent ARC's GAE: geometry-native latent space cuts FVD by up to 23% in world generation — TencentARC · 2026-09-23
GAE Paper: Geometry-Native Latent Space Cuts FVD by 23% and Halves Camera Error — yshan2u · 2026-09-23
Tencent ARC releases GAE: geometry-native latents halve camera error, cut FVD up to 23% — CSProfKGD · 2026-09-24

2 near-duplicate retellings: yshan2u · AdinaYakup