VideoWeave: Unlocking Geometric Consistency in Video Generation via Joint Geometry-Video Modeling

A new post-training framework called VideoWeave aims to reduce geometric drift in AI-generated video by jointly modeling appearance and implicit geometry features in a shared denoising space, according to a paper submitted in 2026 [1]. Large-scale video diffusion models frequently fail to preserve 3D structure over time, producing implausible motion when the viewpoint changes [1]. Existing approaches often rely on explicit geometry reconstructions — depth maps, point clouds, or reconstructed 3D structures — to provide conditions or supervision signals, but this makes the generator sensitive to errors from upstream geometry pipelines [1]. VideoWeave instead adapts implicit geometry-model features into geometry latents and jointly models them with video latents, allowing geometry to shape the generative distribution during training without rigid reconstruction constraints [1]. The framework is designed as a post-training module that can be applied to existing video generation models [1]. To support the method, the authors built GeoVid-80K, a dataset of 80,000 videos with paired appearance and geometry representations [1]. Experiments on text-to-video and image-to-video generation showed that VideoWeave improves geometric coherence while preserving strong visual quality [1]. The work arrives as major AI laboratories continue to advance generative video models. Google DeepMind, for instance, has developed the text-to-video model Veo as part of its broader generative AI portfolio, alongside the text-to-image model Imagen and the Gemini family of large language models [3]. The VideoWeave paper was submitted to arXiv on 12 June 2026 [1].