AnyStitch: Generative Large Parallax Image Stitching

Wide field-of-view panorama stitching is important for computational photography and immersive visual applications, but often faces some challenges like severe geometric disparities. The typical "warp-and-blend" stitching technique only works with limited parallax and fails on large-parallax scenes, due to its inherent restriction to 2D geometric transformations that cannot resolve underlying 3D structures. To break this limitation, we present AnyStitch, a generative framework that can merge multiple reference images with large parallax into a seamless panorama. Unlike previous methods, we formulate the stitching problem as a reference-based image generation process, which naturally avoids stitching artifacts. Specifically, we propose a Hierarchical Projected Rotary Positional Encoding (HPRoPE) based on a 3D feed-forward model to establish accurate geometric relationships between the inputs and the panoramic canvas while preserving fidelity. Given that the resolution of the generated image varies with the number of inputs, we design a spatial autoregressive generation strategy to generate visually coherent panoramas at flexible resolutions. Finally, to enable the supervised fine-tuning of the proposed model, we construct a paired dataset utilizing 3D Gaussian Splatting (3DGS) scenes. Experiments demonstrate that AnyStitch can seamlessly stitch a variable number of inputs (like 22 views in teaser) with large camera motion, support flexible control of target view and projection type, and handle simple dynamic scenes. Project page: https://cecret3350.github.io/AnyStitch/.