Newest papers

Latent Bridge: From Video Latents to 3D Geometry via Feature Alignment

Shuai Zhang, Zhanpeng Ouyang, Zhengqing Chen, Ruohong Mei, Shaoyu Chen, Weiqiang Ren, Hui Sun, Qian Zhang, Wenyu Liu, Xinggang Wang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28521 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

La generación de vídeo y la reconstrucción de feed-forward 3D generalmente utilizan representaciones separadas, requiriendo decodificación RGB y re-codificación visual para recuperar geometría de latentes generados. Proponemos Latent Bridge, que conecta espacios latentes generativos de vídeo con espacios de características geométricas para predecir directamente la geometría 3D de latentes de vídeo. Desarrollamos puentes específicos de representación para Wan2.1 y V-RAE, permitiendo la predicción de feed-forward de cámaras, profundidad y nubes de punto 3D sin el camino RGB intermedio. Cada puente está entrenado en videos RGB a través de la destilación de características, mientras que el codificador de vídeo y el modelo de geometría permanecen congelados. El entrenamiento no requiere

Machine translation (Meta M2M100); the English original follows.

Abstract

Video generation and feed-forward 3D reconstruction typically use separate representations, requiring RGB decoding and visual re-encoding to recover geometry from generated latents. We propose Latent Bridge, which connects video generative latent spaces with geometric feature spaces to directly predict 3D geometry from video latents. We develop representation-specific bridges for Wan2.1 and V-RAE, enabling feed-forward prediction of cameras, depth, and 3D point clouds without the intermediate RGB pathway. Each bridge is trained on RGB videos through feature distillation, while the video encoder and geometry model remain frozen. Training requires neither ground-truth geometry nor execution of the downstream geometric aggregator or prediction heads. At inference, the same video latents support both RGB decoding and geometry prediction. The same trained bridge can be reused for image-to-video and text-to-video generation within the same video latent space, without bridge retraining. Experiments on Co3Dv2 and WildRGB-D demonstrate competitive geometric reconstruction quality, efficient end-to-end generation and reconstruction, and faster latent-to-geometry decoding, highlighting latent alignment as a practical and reusable interface between video generation and 3D reconstruction. Our code is available at: https://github.com/hustvl/Latent-Bridge.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Artificial intelligence, Feature (linguistics), RGB color model, Computer vision, Encoder

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange