🏆ReSplat achieves state-of-the-art performance on DL3DV and RealEstate10K, across various input views (2, 8, 16) and resolutions (256x256-540×960).
Check out our paper for detailed results: arxiv.org/abs/2510.08575

🧠To initialize the recurrent process, we design a compact reconstruction model that operates in a 16x subsampled space, producing 16x fewer Gaussians than previous pixel-aligned models. This substantially reduces computational overhead and allows for efficient Gaussian updates.

🎯Key idea: the Gaussian splatting rendering error provides a rich feedback signal, guiding the recurrent network to learn effective Gaussian updates. This feedback naturally adapts to unseen data, enabling robust generalization across datasets and resolutions.

⚡Feed-forward Gaussian splatting is fast but limited: it only makes a single forward pass. ReSplat introduces recurrent refinement, enabling the model to iteratively improve the 3D Gaussians. ReSplat converges fast, requiring only three iterations.

🚀Excited to share our recent work on test-time scaling for feed-forward Gaussian splatting:
we learn a recurrent model ReSplat that is able to iteratively improve the reconstruction quality in a feed-forward manner!
haofeixu.github.io/resplat/

Wanna scale your feed-forward Gaussian Splatting model to 4K resolution? Come check out our #CVPR2025 poster PanSplat today 10:30–12:30 (June 14) at ExHall D, Poster #74!