IISc’s AdaptiveSplat Uses Texture to Make 3D Scene Representations Smaller
Highlighted by IISc on 23 September, the ECCV 2026 work selectively removes redundant Gaussian primitives. Its compact scene representation should not be confused with an equivalent reduction in total GPU memory.
A smooth wall and a detailed object do not need the same density of elements in a digital scene. Researchers associated with IISc's Vision and AI Lab use that distinction in AdaptiveSplat, a method for reducing the number of Gaussian primitives used to represent a three-dimensional scene reconstructed from images.
IISc highlighted the work on 23 September 2026. The author project identifies it as ECCV 2026 work; the linked preprint was first submitted on 5 July. The September highlight brings attention to that research, rather than announcing a new paper submission this week.
Allocating detail where the scene needs it
Gaussian splatting represents a scene with many small, overlapping primitives whose appearance can be rendered from different viewpoints. AdaptiveSplat starts from a feed-forward reconstruction pipeline and groups related points into regions called SuperClusters.
Its texture analysis uses a discrete wavelet transform to identify image detail. The method removes more redundant primitives from regions judged to have little texture, while an adaptive prediction component adjusts the retained representation. Users can control the pruning level instead of accepting a fixed allocation throughout the scene.
The project reports maintaining reconstruction quality with up to 80% of Gaussians removed in its evaluated settings. That is a claim about the representation and its tested quality-efficiency trade-off. It is not a universal promise that any scene or device will achieve the same result.
Smaller representations are only one resource measure
The distinction between a scene's primitive count and the resources needed to produce it is important. The paper reports the same GPU memory load as the base model in its memory comparison. An 80% pruning result therefore must not be rewritten as an 80% reduction in total GPU memory.
There are also visual limits. The authors identify difficult cases involving detailed patterns on smooth surfaces and specular highlights. Lighting-induced high-frequency detail can mislead the texture-based allocation, degrading reconstruction at aggressive pruning levels.
For applications that need to transmit, store or render reconstructed scenes, selectively retaining useful detail is a relevant research direction. The next practical question is how the quality and resource trade-offs hold across target scenes and hardware. The evidence here is benchmark research, rather than a demonstrated deployment on every mobile device or a guarantee of real-time performance.
Reading the schematic
Figure 3 attribution: Badrinath Singhal, Srihari K G, Sreehari Iyer, Ankit Dhiman and R. Venkatesh Babu, AdaptiveSplat. Source paper, CC BY-SA 4.0. The cover shows a cropped detail of this schematic; the linked original is unmodified. Open the full-resolution figure.