acceptodds
Withdrawn from ICLR 2027

FlexiCoord: Flexible Spatial Representations via Function Reparameterization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial Representations form the foundation of various modern Machine Learning systems in Computer Vision and Graphics, such as image encoding, data compression, and 3D reconstruction. Despite their differences, these methods share a common assumption: spatial functions are evaluated in a fixed coordinate system. We argue that this assumption fundamentally limits representation efficiency, as complex geometric structures must be encoded by increasingly expressive basis functions, which can introduce large memory overhead but marginal performance gain. We introduce FlexiCoord, a general framework that augments a spatial representation via learned function reparameterization, transforming fixed basis functions into adaptive spatial primitives that model complex geometric structures via local coordinate deformation. Unlike architecture-specific improvements, FlexiCoord operates as a universal enhancement that can be attached to arbitrary spatial functions. We validate this hypothesis on three fundamentally different representations: Gaussian kernels for splatting-based novel-view synthesis, multi-resolution hash grids for image encoding, and vector-matrix tensor decompositions for volumetric medical data compression. Across all settings, FlexiCoord yields more efficient representations to capture sharp features with complex geometry. These findings highlight function reparameterization as a broadly applicable mechanism for increasing representational power and suggest a new direction for the design of spatial representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spatial representation, gaussian splatting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Z7qcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework that augments existing spatial representations with learned coordinate reparameterization/ deformation fields. The method remaps query coordinates before evaluating the underlying spatial function, aiming to improve the representation of sharp boundaries, thin structures, and fine details. The framework is instantiated with Gaussian kernels for novel-view synthesis, multiresolution hash grids for image encoding, and vector–matrix decompositions for medical-volume compression. In the Gaussian variant, per-primitive displacement textures modify the kernel’s spatial weighting within its local tangent plane. The hash-grid and tensor variants instead use learned offsets to query their feature representations. Experiments evaluate Gaussian splatting at fixed primitive budgets and image and volume encoding under approximately matched memory budgets. The reported results show improved reconstruction quality over the corresponding base representations. The Gaussian variant achieves rendering speeds comparable to 2DGS and faster than the evaluated textured Gaussian methods, while requiring additional storage relative to 2DGS for its displacement fields.

Critical strengths
  1. The paper formulates coordinate reparameterization as a simple function composition and demonstrates its application to Gaussian kernels, multiresolution hash grids, and VM decompositions. These instantiations provide evidence that the idea can benefit multiple representation families and signal types.

  2. The proposed method alleviates the representative efficiency issue of prepresentation in scene reconstruction in a simple and novel way. The pipeline is sound.

Critical weaknesses
  1. The main GS comparisons fix the number of primitives, although FlexiCoord introduces additional learnable displacement textures. Table 3 reports that its models require approximately 1.6–1.8x the storage of 2DGS at the 10% primitive budget. While the quality and rendering-speed improvements are useful, the experiments do not establish whether allocating this additional capacity to deformation is more effective than allocating the same storage to additional ordinary Gaussians. Comparisons under matched total storage, together with a parameter breakdown, would better substantiate the representation-efficiency claim.

  2. The GS experiments compare against 2DGS, TexturedGS, and FACT-GS, but do not include Mini-Splatting, which improves Gaussian spatial distribution and simplifies the representation to operate with fewer primitives. This omission is relevant because the paper’s main GS evaluation emphasizes reconstruction quality at restricted primitive budgets. The existing comparisons demonstrate the benefit of deformation over the evaluated baselines, but leave unclear how it compares with strategies that improve the placement and selection of conventional Gaussians. Including Mini-Splatting at matched primitive counts, with reconstruction quality, total model size, and rendering speed reported together, would provide a more informative assessment of the proposed quality–efficiency trade-off.

Questions for authors

How sensitive are the results to the displacement-field resolution and deformation strength?

The paper mentions implementing a custom CUDA kernel based on gsplat. Could the authors clarify how the rasterization pipeline handles the learned coordinate offsets, particularly how screen-space bounding boxes and tile assignments are computed? Since coordinate reparameterization can extend a kernel’s effective support beyond the original Gaussian footprint, how do you ensure that valid contributions are not excluded during candidate selection or early culling?

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

The paper proposes FlexiCoord, which improves spatial representations through learned coordinate reparameterization. It instantiates this idea as local Gaussian kernel deformation, warped hash-grid image encoding, and warped vector–matrix volume encoding. Experiments show improved reconstruction at fixed Gaussian counts and approximately matched storage budgets for image and volume encoding, while preserving rendering speed close to 2DGS.

Critical strengths
  • A practical improvement to Gaussian expressiveness. The local kernel deformation improves reconstruction under restricted primitive budgets while retaining rendering speed close to 2DGS.
  • Evidence across distinct representations. Improvements on Gaussian splatting, hash grids, and tensor-factorized volumes support the usefulness of the approach beyond a single implementation. The approximately matched-memory NGP and VM comparisons are particularly helpful.
Critical weaknesses
  • The conceptual novelty needs better clarification. Learning a coordinate mapping jointly with a spatial field has substantial precedent, including General Neural Gauge Fields [1] and Factor Fields [2]. Factor Fields formulates spatial representations as fields evaluated on transformed input coordinates and already covers Instant-NGP and TensoRF, two of the three representations instantiated here. The paper does not sufficiently distinguish its general formulation from this literature.
  • The paper contains no ablation studies, leaving its key design choices unsupported.

[1] General Neural Gauge Fields. ICLR 2023

[2] Dictionary Fields: Learning a Neural Basis Decomposition. SIGGRAPH 2023

Questions for authors
  • What is the precise methodological distinction between the general FlexiCoord formulation and prior jointly learned coordinate transformations, including General Neural Gauge Fields and Factor Fields?
  • Does the 2DGS baseline receive the same total optimization budget as the two-stage FlexiCoord-GS procedure, including an equivalent additional refinement stage?
  • FlexiCoord-GS uses more memory than plain 2DGS at the same primitive count (Table 3). How does it compare with 2DGS using additional primitives under similar storage, or with a per-primitive opacity texture of matched resolution? Does the quality and speed advantage still remain?
  • Could the authors provide a sensitivity analysis of λ, and support the single-level offset-grid and nested two-layer offset-VM designs with ablations?
Reviewer_sPKmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FlexiCoord, a coordinate reparameterization strategy that adapts the target signal for various spatial representations, such as Gaussian kernels, multi-resolution hash grids, and vector-matrix (VM) tensor decompositions. Specifically, it adopts a learned coordinate reparameterization field that warps input coordinates to better capture complex signals. Based on this idea, FlexiCoord is applied to 2D Gaussian Splatting (2DGS) for novel view synthesis (NVS), Instant-NGP for ultra-resolution image encoding, and VM decomposition for volumetric data compression. Experimental results on each task demonstrate that the proposed FlexiCoord improves the quality of baseline methods under the same memory budget. In particular, FlexiCoord achieves better rendering quality than existing GS methods while maintaining high rendering speed comparable to 2DGS.

Critical strengths
  • The learned coordinate reparameterization field provides a general strategy applicable to various spatial representations, consistently improving their representation quality. In this paper, the proposed framework is applied to three representative spatial representations, demonstrating improved performance on each task. Beyond these representations, it can also be applied to other coordinate-based representations.
  • The manuscript clearly describes the differences between the proposed approach and existing methods, such as texture-based representations. In particular, while existing methods focus on increasing the representation capacity of appearance through texture mapping for color and opacity, the proposed technique enhances the flexibility of the underlying spatial representation by reparameterizing its coordinate domain.
Critical weaknesses
  • The conceptual novelty of FlexiCoord appears limited.
    • Evaluating a spatial function at learned, signal-adaptive coordinates has already been explored in various implicit and grid-based representations. For example, deformation fields have been widely adopted for representing dynamic scenes in NeRF and 3DGS. More directly, DefGrid [1] learns to deform a fixed grid according to the target signal, suggesting that a similar idea has already been explored for grid-based representations such as FlexiCoord-NGP and FlexiCoord-VM.
    • Furthermore, FlexiCoord-GS is also closely related to FACT-GS, with the learned deformation applied to the Gaussian kernel rather than the texture. The paper should better distinguish its conceptual contribution from these prior coordinate-reparameterization approaches.
  • Although the paper demonstrates FlexiCoord on three different tasks, the current evaluation appears insufficient to fully support its claim as a general strategy.
    • Several experimental settings and analysis for results are not sufficiently described. For NVS, it is unclear how the N% primitive settings are obtained and why the full 100% setting is not evaluated. Similarly, the ultra-resolution image experiments rely on images collected from the Internet without clearly specifying their sources and detailed specifications. Furthermore, the substantial FPS improvement also requires further explanation to clarify whether it stems from the representation itself or from the custom implementation.
    • The NVS evaluation is limited to comparisons with 2DGS variants. Comparisons with 3DGS variants, such as standard 3DGS and Scaffold-GS, would make the evaluation more comprehensive.
    • Since both NGP and VM are grid-based representations, evaluating FlexiCoord on a substantially different representation family, such as an implicit representation, would provide stronger evidence for its claimed generality.

[1] Gao et al., Beyond Fixed Grid: Learning Geometric Image Representation with a Deformable Grid, ECCV 2020

Questions for authors
  • Why is FlexiCoord not applied to standard 3DGS, which is one of the most widely used Gaussian representations? Also, can FlexiCoord be applied to coordinate-based implicit representations? To support the claimed generality of the framework, it would be more important to demonstrate its applicability across diverse representation families rather than simply across different tasks.

  • As discussed in the weaknesses, please clarify the conceptual novelty of the proposed method compared with existing coordinate deformation approaches.

  • The FPS improvement of FlexiCoord-GS is substantial, although it is unclear what causes such a large speedup. Please clarify whether this improvement mainly comes from the representation itself, the pruning strategy, or implementation-level optimizations.

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

Related papers

Discussion (0)

Sign in to comment.