acceptodds
Under review as a conference paper at ICLR 2027

LeanSplat: Low-Overhead Language Gaussian Splatting

Abstract

Modeling 3D language fields with Gaussian Splatting suffers from high storage and computational costs due to the large number of Gaussian primitives, high-dimensional language features, and complicated retrieval procedures. Primitives from a pretrained 3DGS scene are typically carried into the language branch, so even those that contribute little to semantic prediction require high-dimensional language features and participate in semantic rendering. Query pipelines that reconstruct high-dimensional pixel-level language features before text matching repeat this high-dimensional computation for each query. In this paper, we propose LeanSplat, a low-overhead framework for Language Gaussian Splatting. LeanSplat alternates semantic-opacity optimization with multi-view language-feature aggregation to progressively refine primitive-level semantic visibility. The resulting cross-view semantic rendering contributions guide pruning of low-contribution primitives, while semantic opacity weights the construction of shared prototypes that compress the high-dimensional language features of retained primitives. At query time, text scores are computed directly in prototype space, mapped to primitives via prototype indices, and rasterized without reconstructing the full high-dimensional language feature field. Experiments on LERF show that LeanSplat removes 83.25% of language primitives while achieving an average mIoU of 65.5%. With a prototype budget of \(K=256\), the average language storage is reduced to 2.94 MiB/scene, while query throughput exceeds 500 queries/s on an A100 GPU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian Splatting;3D Language Fields;Open-Vocabulary 3D Understanding;Efficient Representation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XXdKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents LeanSplat, an efficient Language Gaussian Splatting framework. It alternates semantic-opacity optimization with multi-view feature aggregation and prunes low-contribution Gaussians from the language branch. It further compresses language features using opacity-weighted K-means. At query time, text scores are computed on shared prototypes and directly rasterized into pixel-level score maps, avoiding high-dimensional feature reconstruction.

Critical strengths
  • The method combines semantic point pruning, shared feature prototypes, and direct score rendering to reduce language-branch storage and query costs.

  • On LERF, it removes 83.25% of language primitives and reduces language storage to 2.94 MiB per scene while achieving 65.5% mIoU. Query throughput exceeds 500 queries/s on an A100 GPU.

Critical weaknesses

1. Limited justification for optimizing only semantic opacity.

The method freezes Gaussian positions, scales, and rotations and optimizes only semantic opacity. While this preserves the pretrained RGB representation, it limits the ability to adapt Gaussian support to semantic boundaries. An ablation allowing language-specific geometric adjustments would clarify the accuracy–efficiency trade-off of this design. Why not optimize other Gaussian attributes?

2. The interpretation of semantic opacity as visibility requires stronger justification.

Why do the authors use semantic opacity for prototype clustering? The authors refer to this as “Visibility-guided prototype construction” but why use opacity rather than the alpha-blending weight, which reflects actual visibility?

3. Unclear ScanNet evaluation protocol after pruning.

Tables 2 and 3 evaluate pruned models only on retained points. Why does reducing the number of Gaussian primitives reduce the number of ScanNet points used for evaluation? Could the models before and after pruning be evaluated on the same full ScanNet point cloud? Please clarify how predictions are mapped to ScanNet annotation points and how the retained points are selected.

Questions for authors

Please refer to the weaknesses above.

Reviewer_aZVHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to reduce the storage and query cost of the language branch in language Gaussian splatting. On top of a frozen RGB 3DGS, it learns a separate semantic opacity, alternating its optimization with multi-view re-aggregation of CLIP-SAM features. The learned opacity is then used to prune language primitives and to weight a k-means codebook that replaces per-primitive features, and text queries are scored directly in codebook space before rasterization. On LERF and ScanNet the method reaches accuracy comparable to recent methods, with about 3 MiB of language storage per scene and roughly 500 queries/s on an A100.

Critical strengths

(1) Decoupling semantic opacity from RGB opacity is simple and reasonable. It gives a clear gain over using the RGB compositing weights and holds up better than RGB significance under aggressive pruning. I think this is the most useful takeaway of the paper.

(2) The resulting representation is compact and fast in practice, and the appendix is fairly transparent about how storage and latency are counted.

Critical weaknesses

(1) The accuracy improvements are small and not very convincing. LERF has only four scenes, there are no repeated runs, and the final mean (65.5) is about the same as CCL-LGS, ReLaGS and LaGa. Without the codebook the method is actually below them (63.69). The codebook with K = 256 adds almost 2 points, but the K sweep is not monotonic, and K and the pruning threshold seem to be chosen on the same test scenes. The iterative refinement itself adds less than 1 point over plain optimization. On ScanNet, the pruned model is evaluated only on the points it kept, which isn't comparable to the baselines. Also, Table 9 compares two models that already differ by about 3 points before any pruning.

(2) The efficiency claims are weaker than presented. Table 11 shows that without pruning the query is slower than the latency reported for LangSplatV2, so the speedup comes from pruning rather than from the decode-free query. The decode-free trick is essentially reordering a matrix product and could be applied to LangSplatV2 or any other codebook-based method. The comparison is also limited to LangSplatV2 with numbers copied from its paper, while methods (LightSplat, CF3) are cited but not compared. Since each component is close to existing practice, I would need stronger evidence on either accuracy or efficiency.

Questions for authors

(1) How were K and the pruning threshold selected? Could you report results over multiple seeds?

(2) What are the ScanNet results of the pruned model when evaluated on all points?

(3) Can you time LangSplatV2 in your own setup, with the same codebook-side text projection and with pruning applied? A comparison with LightSplat would also help.

(4) Is the rendered feature normalized before computing relevancy?

Reviewer_VjfjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LeanSplat, an iterative semantic visibility learning method to build a compact language representation. It identifies the overhead of the existing Language Gaussian Splatting method: a large primitive set optimized for RGB, costly storage of high-dimensional per-primitive features, and costly high-dimensional feature reconstruction at query time. To address this, LeanSplat introduces a separate semantic opacity, which is optimized alternately with training-free CLIP feature aggregation. The learned semantic compositing weights are then used to prune the primitives. Furthermore, the retained full-dim features are compressed into a small set of shared prototypes. Each retained primitive then stores only one prototype index per level plus its semantic opacity. The method is validated on LERF-OVS and ScanNet, reporting state-of-the-art open-vocabulary segmentation performance.

Critical strengths
  • Compact representation with near-SOTA performance. The average language storage for LERF is 2.94 MiB, with an overall pruning ratio of 83.25%. The average results on LERF and ScanNet is on par with or slightly above the SOTA.
Critical weaknesses
  1. The presented approaches only provide limited novelty. Semantic-specific opacity has already been explored by work e.g. 3DVL-GS or OP2GS. The training-free multi-view CLIP feature aggregation has been explored by LUDVIG, Dr.Splat, VALA. The prototype-based language storage comes from Dr.Splat. So the major novelty falls in the alternately update manner and the Bernoulli entropy, while in ablation they show minor improvement to the accuracy.

  2. The paper claims low overhead without reporting the reconstruction time and memory requirements compared to other methods. The storage comparison is only done with LangSplatV2, but not other compact methods e.g. Dr.Splat.Therefore, there is not enough evidence for the low overhead claim.

Questions for authors
  • Related to W1. Please fully discuss the related literature and fairly measure the unique contribution and novelty.

  • Related to W2. What are the per-scene reconstruction time and training memory of the full language-branch construction (aggregations, opacity optimization, k-means), compared with LangSplatV2, Occam's LGS and Dr. Splat?

  • Is semantic-opacity guided pruning better than RGB-guided in terms of accuracy? It is claimed so(l. 425–427). However, the RGB-visibility setting also uses RGB opacity for compositing. Without any pruning, the two settings already reach 60.17 vs. 63.26 mIoU (a 3.09 gap, Tab. 9). After pruning 83.25% of the primitives with each criterion, they reach 60.83 vs. 63.69 (a 2.86 gap). The gap therefore comes from the compositing opacity, whose benefit has already been shown by prior work such as 3D VL-GS. It does not come from the pruning criterion: at the same pruning ratio, the gap shrinks rather than grows. The experiment does not isolate the advantage of semantic-visibility-guided pruning.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.