Tencent ARC releases GAE: geometry-native latents halve camera error, cut FVD up to 23%
CSProfKGD · x · 2026-09-24
Tencent ARC released GAE (Geometry-Native Autoencoder), a geometry-native latent space for 3D-consistent world generation. The idea: move geometric representation learning into the autoencoder so the DiT can focus on modeling the distribution, yielding latents that act as a flexible "soft skeleton" for persistent worlds.
- Compresses 3,072-channel geometry features into a compact 128-channel latent
- A single generated state decodes into RGB, depth, camera trajectories, and point clouds
- Halves camera-trajectory error and cuts FVD by up to 23% on standard benchmarks
More from Multimodal
- Claude models, textures and rigs characters in Blender in pure code — majidmanzarpour · 2026-09-24
- Meta launches Muse Realtime Avatar: sub-second voice-video synced avatar — alexandr_wang · 2026-09-24
- Two-word prompt video cutouts: SAM 3.1 on Meta's Model API demoed — nikhilaravi · 2026-09-24
- DiffusionGemma served locally via SGLang recognizes terrible cat doodles — ChengleiSi · 2026-09-24
- GLM-5.3 crushes stealth model Space Bunny Alpha on Newton's cradle physics test — rohanpaul_ai · 2026-09-24
- $1,874 of Opus tokens generated an entire explorable 3D island running at 60 FPS — DeryaTR_ · 2026-09-24