AnySplat: Feed-forward 3D Gaussian Splatting from Unconstrained Views

We introduce AnySplat, a feed forward network for novel view synthesis fromuncalibrated image collections. In contrast to traditional neural renderingpipelines that demand known camera poses and per scene optimization, or recentfeed forward methods that buckle under the computational weight of dense views,our model predicts everything in one shot. A single forward pass yields a setof 3D Gaussian primitives encoding both scene geometry and appearance, and thecorresponding camera intrinsics and extrinsics for each input image. Thisunified design scales effortlessly to casually captured, multi view datasetswithout any pose annotations. In extensive zero shot evaluations, AnySplatmatches the quality of pose aware baselines in both sparse and dense viewscenarios while surpassing existing pose free approaches. Moreover, it greatlyreduce rendering latency compared to optimization based neural fields, bringingreal time novel view synthesis within reach for unconstrained capturesettings.Project page: https://city-super.github.io/anysplat/