Daily AI Paper

Today’s pick is a clever way to turn video generation models into geometry solvers. The problem is that estimating depth and surface normals from a single image usually needs lots of labeled data, and many recent generative methods either train separate models for each geometry task or fine-tune large diffusion backbones with expensive supervision. This paper, GeoNeXt, takes a different route: it repurposes a pretrained video generative model and frames geometry prediction as next-frame forecasting. That lets the model use the temporal structure it already understands as a strong prior, while jointly learning image-to-geometry and geometry-to-image relationships. The result is a more data-efficient system that can produce both depth and surface normals, even in zero-shot settings. What makes this interesting is the broader idea: instead of building a new model from scratch for every vision task, we can sometimes unlock powerful representations by changing how we ask the model to predict. That matters for robotics, AR, 3D reconstruction, and any application where geometry is needed but labels are scarce.