The Flow has not summarised this story yet — read it at arXiv Computer Vision.
F4Splat introduces a feed‑forward predictive densification strategy for 3D Gaussian splatting that allocates Gaussians based on a densification‑score guided by spatial complexity and multi‑view overlap. The method predicts per‑region scores to estimate required Gaussian density, enabling explicit control over the total Gaussian budget without retraining. This adaptive allocation reduces redundancy in simple regions and minimizes duplicate Gaussians across overlapping views, yielding compact yet high‑quality 3D representations and superior novel‑view synthesis performance with fewer Gaussians.
The paper introduces SPAR3S, a sparse voxel‑aligned 3D latent generative model that completes 3D scenes from sparse, unconstrained multi‑view images. By representing only occupied voxels in a compact latent space and training a masked autoregressive transformer with photometric supervision via differentiable 3D Gaussian Splatting, the method predicts missing latent tokens and spatial support, enabling efficient and spatially consistent generation of unseen regions. Experiments on synthetic indoor scenes and RealEstate10k demonstrate higher novel‑view quality and real‑world applicability compared to prior work.
VoxelTTO is a feed‑forward framework that reconstructs 3D Gaussian splatting scenes from multiple images by aggregating dense image features into a global voxel representation and decoding Gaussians from voxel features, thereby eliminating the pixel‑to‑Gaussian correspondence. It incorporates test‑time optimization with lightweight LoRA modules to adapt to known camera parameters while keeping the pretrained visual foundation model frozen. The method replaces standard rasterization with stochastic solid volume rendering, improving geometric fidelity, and demonstrates superior RGB‑D novel‑view synthesis and camera‑pose estimation on Replica, Tanks and Temples, and DTU datasets.
The paper introduces SPAR3S, a sparse voxel‑aligned 3D latent generative model that completes 3D scenes from sparse, unconstrained multi‑view images. It learns a compact voxel‑aligned latent space using photometric supervision via differentiable 3D Gaussian Splatting, and employs a masked autoregressive transformer to predict missing voxel occupancy and latent tokens. Experiments on synthetic indoor scenes and RealEstate10k show that SPAR3S achieves higher novel‑view quality than prior methods and generalizes to real‑world data.
CompactSplat introduces a feedforward scene reconstruction method that adaptively allocates 3D Gaussian primitives based on texture-guided spatial partitioning and hierarchical sampling. By concentrating dense Gaussians in complex regions and using expanded primitives for flat areas, it reduces redundancy and better captures planar geometry. The framework also allows dynamic sparsity control during inference, enabling trade-offs between rendering quality and memory usage without retraining.
The paper introduces WINGS, a reference‑free Gaussian splatting inpainting technique that operates directly in 3D. It uses a large, pre‑trained 3D generative prior and a structure completion network to reconstruct missing geometry and appearance without generating reference views. The method claims faster performance and reduced multi‑view inconsistency compared to 2D diffusion‑based approaches, and its effectiveness is validated through experiments and a user study.