AllocSplat adapts reconstruction capacity with one model. (a,b) Reconstructions at 36.9K Gaussians with insets of the boxed region at 9.2K, beside all decoded Gaussian centers in 3D; (a) an input view, (b) a held-out view. (c) DL3DV-140, four posed inputs, no test-time optimization. AllocSplat and ZipSplat sweep capacity with fixed weights; the others sit at their evaluated capacities. At 36.9K Gaussians AllocSplat reaches 23.87 dB, 85.9% fewer Gaussians than TokenGS at 0.99 dB higher PSNR, and it leads ZipSplat by 0.29 to 1.65 dB at equal primitive counts.
Feed-forward 3D Gaussian Splatting reconstructs renderable scenes from a few posed images, but storing, streaming and rendering them require an output budget chosen by the application. At a smaller budget, each Gaussian must cover more of the scene, with its required extent depending on where the others lie in 3D. We introduce AllocSplat, which allocates capacity through spatial anchors and uses their local spacing to control Gaussian decoding. From the depth and features of a visual geometry model, a non-learned rule selects M anchors among back-projected candidates. Each anchor decodes K Gaussians whose offsets and scales are constrained by its distances to neighbouring selected anchors. Recomputing this local scale and the contextualized decode lets one set of weights serve a range of budgets, emitting exactly KM Gaussians per request. On DL3DV-140, the default budgets achieve higher PSNR and SSIM than the evaluated feed-forward baselines with fewer Gaussians at two, four and six input views. At four views, AllocSplat reaches 23.87 dB with 36.9K Gaussians, 0.99 dB above TokenGS with 85.9% fewer primitives, and 0.29–1.65 dB above ZipSplat at equal counts. Controlled comparisons show that the local scale improves on both fixed-footprint and global budget-scaling controls, and that bounding Gaussian offsets as well as sizes by it helps at every budget.
One checkpoint serves every budget: V = 2, 4 or 6 posed views, M = 256 to 1,944 anchors, K = 36 Gaussians per anchor.
AllocSplat. Sparse views and their known cameras pass through a visual geometry model (VGM), whose multi-layer features and depths are unprojected into a pool of 3D anchor candidates. A joint geometry–appearance farthest-point rule allocates exactly M of them, and the resulting density sets each anchor's local spatial scale ρi, so the same centers are decoded differently at different budgets. The Gaussian head emits K primitives per anchor, KM in total.