Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting
Yixing Lao, Xuyang Bai, Xiaoyang Wu, Nuoyuan Yan, Zixin Luo, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Shiwei Li, Hengshuang Zhao
TL;DR
LGTM is a novel feed-forward framework that enables high-resolution 4K view synthesis by predicting compact primitives with textures, overcoming previous scalability limitations of Gaussian Splatting methods.
Contribution
Introduces LGTM, a method that decouples geometric complexity from resolution, allowing efficient 4K synthesis without per-scene optimization or excessive primitives.
Findings
Enables 4K view synthesis with fewer primitives.
Decouples geometry from rendering resolution.
Achieves high-fidelity results without per-scene optimization.
Abstract
Existing feed-forward 3D Gaussian Splatting methods predict pixel-aligned primitives, leading to a quadratic growth in primitive count as resolution increases. This fundamentally limits their scalability, making high-resolution synthesis such as 4K intractable. We introduce LGTM (Less Gaussians, Texture More), a feed-forward framework that overcomes this resolution scaling barrier. By predicting compact Gaussian primitives coupled with per-primitive textures, LGTM decouples geometric complexity from rendering resolution. This approach enables high-fidelity 4K novel view synthesis without per-scene optimization, a capability previously out of reach for feed-forward methods, all while using significantly fewer Gaussian primitives. Project page: https://yxlao.github.io/lgtm/
Peer Reviews
Decision·ICLR 2026 Poster
* The idea of combining low-resolution Gaussian primitives with high-resolution texture maps to achieve high-resolution feed-forward Gaussian Splatting is well-motivated and sound. Such an insightful finding may significantly boost the feed-forward 3DGS community to explore higher-quality synthesis. * The introduced module is architecture-agnostic and is thoroughly evaluated across several state-of-the-art models and large-scale benchmarks. All experiments show consistent quantitative and quali
* Lack of multi-view results. It would be better to provide video results or multi-view images to better illustrate the impact of the texture modules. I am concerned that the texture module may potentially destroy multi-view consistency to some extent. * Lack of evaluation under dense multi-view settings. Most experiments are conducted with 1, 2, or 4 views. Since the multi-view model is based on VGGT, which natively supports dense input views, it would be better to include results with denser
1. A novel topic on feed-forward 3D-GS. 2. A fair well performance compared with baseline methods.
1. High-resolution rendering requires accurate geometric prediction. However, the proposed method seems more like a trick—it projects a high-resolution image onto relatively coarse geometry. While this may work well when the input views have small viewpoint differences, it would be helpful if the authors could include additional experiments to evaluate the novel-view synthesis quality under different camera pose settings. 2. The paper claims to "decouple" geometry and appearance 20, yet the arch
1. The paper is well-motivated. 2. The paper is well-written and generally easy to follow. 3. The experiments are throughout and the proposed method significantly boosts the performance of the baselines. 4. The results is visually good.
1. Necessity of integrating 4K texture to 3DGS: another solution to get 4K renderings can be a general feed-forward 3DGS followed by an image super-resolution model. A comparison should be made between it and the proposed method. 2. Evaluation: As the paper claims an immersive user experience, a non-reference perceptual metric such as Niqe [1] or Q-align [2] should be added as an evaluation metric. 3. Line space issue in L.474. [1] Zhang, L., Zhang, L., & Bovik, A. C. (2015). A feature-enriched
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Vision and Imaging · Video Coding and Compression Technologies · Computer Graphics and Visualization Techniques