acceptodds
Under review as a conference paper at ICLR 2027

RevTet: Adaptive and Reversible Tetrahedral Neural Representations

Abstract

Novel-view synthesis enables immersive exploration of scenes reconstructed from images, making radiance fields an important representation for visual computing. However, directly learned volumetric attributes remain difficult to combine with movable geometric support and reversible resolution changes on a conforming mesh. We introduce RevTet, a tetrahedral radiance representation that couples constrained vertex optimization with compatible newest-vertex bisection, explicit attribute transfer, and checked inverse events for capacity reuse and levels of detail. Experiments show that importance-guided refinement and vertex motion improve reconstruction in controlled continuations, while hierarchical coarsening retains much of the learned image quality at reduced cell counts without retraining. These results support hierarchical tetrahedral adaptation as a means of controlling representation capacity while retaining direct field parameters and explicit connectivity, with lossy coarsening providing a tunable quality–size trade-off.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
novel view synthesis, radiance fields, differentiable rendering, 3D scene representations, adaptive representations, tetrahedral meshes

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 4 reviews, above 27% of the others.
Reviewer_KB6LOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a radiance field representation for noel view synthesis, based on tetrahedralization of space, where the vertices carry density and directionally-varying radiance. This allows for an efficient radiance estimation along any ray. Similar approaches exist (Radiance MEshes, RadiantFoam, SVRaster); the novelty here is in the reversible subdivision operations that automatically adapt the vertex density to important regions.

Critical strengths

Overall, this is a useful direction; alternative approaches to Gaussian splatting or NeRF could become very important if they are shown to be more robust and/or amenable to clean adaptiveness heuristics. The metrics reported are promising.

Critical weaknesses

The fundamental issue with the current version of the paper is that it shows very few visual results, and that it reports no results (visual or numerical) on real scenes reconstructed from photographs. It is impossible to judge the kinds of artifacts the method could lead to in more complex cases, from so few experimental data. In my opinion, the viability of the paper for publication completely depends on resolving these omissions. The writing could also be improved, but that is not the critical issue.

Questions for authors

Was real data (object or full scene) tested with this method?

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

This paper presents RevTet, a tetrahedral radiance-field representation that integrates directly learned attributes, constrained vertex optimization, conforming newest-vertex bisection, geometric repair, and validity-checked inverse operations. The main goal is to support adaptive scene reconstruction and post-training levels of detail within a persistent tetrahedral hierarchy.

The paper demonstrates competitive reconstruction quality on NeRF Synthetic, improvements from importance-guided refinement and movable vertices in a controlled continuation, improved worst-cell geometry through repair, and quality-preserving offline coarsening.

Critical strengths
  1. Meaningful problem and coherent system design. Combining direct radiance attributes, movable geometry, and conforming hierarchical adaptation is a useful direction. The integrated system is technically nontrivial.

  2. Reasonable mathematical motivation and implementation checks. The moving-basis analysis, affine refinement inheritance, and geometric validity checks provide a clear foundation for the proposed operators.

  3. Useful experimental evidence for individual components. The controlled Lego continuation demonstrates benefits from adaptive allocation and vertex optimization. The repair ablation substantially reduces severe cell compression while maintaining nearly unchanged image quality.

Critical weaknesses
  1. The central claim of adaptive reversibility is only partially validated.

Although reversible hierarchical adaptation is a central contribution, merging is disabled in the main full-training experiments. The coarsening results primarily evaluate offline LOD generation rather than repeated refinement, learning, coarsening, and capacity reallocation during training.

Moreover, coarsening is lossy, and inverse operations may become unavailable because of validity constraints. The current evidence establishes that some topology operations can be reversed under specific conditions, but does not establish the practical effectiveness of reversible capacity management throughout training.

  1. The novelty relative to existing tetrahedral representations and mesh-adaptation methods needs clearer justification.

Several key components, including compatible NVB, affine coefficient inheritance, constrained mesh motion, and the ALE-inspired moving-basis formulation, build on established techniques.

The integration is potentially valuable, but the paper does not sufficiently isolate what is algorithmically new or what distinctive capabilities are enabled by the integration.

In addition, Tetra-NeRF (ICCV 2023) appears to be relevant prior work but is not discussed. The relationship to Radiance Meshes and Radiant Foam is explained qualitatively without a direct quantitative comparison of relevant capabilities.

  1. The empirical comparisons do not sufficiently establish practical efficiency advantages.

RevTet achieves competitive but not superior full-resolution reconstruction quality relative to the reported baselines. More importantly, the experiments do not provide matched end-to-end training costs, rendering FPS, or quality-versus-memory comparisons against the most relevant methods.

The LOD comparison uses percentages of each method's own primitives, although tetrahedra and voxels are not equivalent in representation capacity or storage. Reducing RevTet to approximately 50% of its active tetrahedra reduces its persistent tensor inventory from 852.41 MiB to 615.00 MiB, illustrating the nontrivial hierarchy overhead.

Therefore, the reported quality-retention curves do not establish superiority at matched computational or storage budgets.

Questions for authors
  1. What are the rendering FPS, end-to-end training time, and actual memory requirements of RevTet compared with relevant existing methods at matched image quality or resource budgets?

  2. Can the authors clarify the novel contributions relative to Tetra-NeRF and provide a more targeted comparison with Radiance Meshes or Radiant Foam?

  3. Can the authors extend the key adaptive-refinement ablation beyond Lego, or provide additional evidence regarding its robustness across scenes and random seeds?

  4. How much of the benefit comes from the conforming NVB hierarchy itself, rather than from importance-based allocation, vertex optimization, or the initialization procedure?

Reviewer_pPz6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RevTet is a per-scene radiance-field representation on a conforming tetrahedral mesh. Density pre-activations and spherical-harmonic (SH) coefficients are stored directly at the vertices, with per-cell color offsets, and are interpolated barycentrically; rays traverse shared faces. The mesh is seeded from a trained voxel field (five tetrahedra per voxel, then two uniform bisection rounds) and trained for 30k iterations under a 4M-cell ceiling. Four operators act on it: (a) Constrained vertex learning. Interior vertices move together with the attributes, under a fixed boundary and a per-cell determinant floor, with rollback on failure. (b) Refinement. Image-guided newest-vertex bisection (NVB) with closure keeps the mesh conforming. Midpoint averaging reproduces the parent's affine field exactly (Prop. 1). (c) Merging. Recorded parent–child events are inverted after validity checks, which yields offline levels of detail (LOD). Merging is lossy whenever the midpoint carries learned detail. (d) Repair. Recorded midpoints are replayed and local vertex corrections are applied to degenerate cells. A moving-basis (ALE) identity (Eq. 4) explains why moving vertices with fixed attributes also moves the field. Extensive experiments have demonstrated the effectiveness of the proposed method.

Critical strengths

It is a principled and carefully verified operator framework. Each edit pairs a mesh change with an explicit attribute-transfer rule and a validity check:

  • NVB refinement with closure reproduces the parent's affine field exactly (Prop. 1).
  • Inverse events restore recorded parents only after face-incidence, boundary and determinant checks.
  • Vertex updates are constrained to a feasible set (Eq. 8) and rolled back on failure.
  • Eq. 4 separates moving the field from moving the mesh. It makes clear why repair without remapping changes the world-space interpolant, and Table 5 shows how far bounded remapping reduces that error.

The verification is thorough: fifteen CUDA operator assertions, a float64 reference renderer, and zero traversal failures across all 88 LOD states and all final meshes. This matters because a reversible hierarchy over directly stored attributes is the paper's main contribution, and none of the closest methods (SVRaster, Radiant Foam, Radiance Meshes) reports one.

Critical weaknesses

W1. The evidence that the operators matter is narrow. All operator ablations are single-seed and partial:

  • Refinement and motion. The only test is a 6k-step Lego continuation from iteration 1k, ending at 28.0–30.6 dB versus 35.18 dB for the full model. At a saturated checkpoint, all four fixed/movable and fixed/reallocated arms differ by less than 0.002 dB (App. C.3). There is no eight-scene full-training ablation (motion off, random refinement).
  • Repair. Repair improves the worst cells with no visible downstream effect: PSNR changes by −0.006 dB, and both arms already pass every certificate with zero traversal failures. It costs +21% training time and +14% memory, and leaves Ficus essentially unchanged.
  • LOD ranking. At 50%, 76% of the mean gain over RandomEvent comes from Ship (+1.69 dB). Without Ship the gain is 0.08 dB; RandomEvent is ahead on Chair (0.26 dB) and level on Drums, Ficus and Mic.

This weighs most in my rating: apart from LOD, no operator is shown to improve a converged model.

W2. The practical value relative to existing methods is not established.

  • Quality. RevTet trails 3DGS and SVRaster (32.99 vs. 33.33 / 33.25 dB; LPIPS 14% worse than local SVRaster), despite 30k iterations on top of a pre-trained voxel field.
  • No comparison with the closest methods. The motivation against Radiance Meshes and Radiant Foam is avoiding neural attribute queries and Delaunay retriangulation, yet neither is compared quantitatively; Table 6 is only structural.
  • No rendering speed, real scenes or compact output. No FPS is reported, and all scenes are bounded and synthetic. At 4M tetrahedra, persistent tensors average 852 MiB per object, with no render-only size given; halving the cells saves only 28% because the hierarchy remains.

Without these, it is unclear when RevTet should be preferred.

Questions for authors
  1. How do Radiance Meshes and Radiant Foam compare on the same scenes (PSNR/LPIPS, training time, rendering FPS, memory)? This would test the claimed advantage of avoiding neural queries and retriangulation.
  2. What is RevTet's rendering FPS at 100% and 50% retained cells? How large is a render-only export without the hierarchy, compared with SVRaster and 3DGS?
Reviewer_CsXPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RevTet, a per-scene radiance-field representation on a conforming tetrahedral mesh. Density pre-activations and spherical-harmonic coefficients are stored at vertices and interpolated barycentrically, and rays traverse cells through shared faces. Starting from a trained voxel field, the method jointly optimizes vertex positions and attributes under fixed connectivity with positive-volume constraints. It refines the mesh by image-guided newest-vertex bisection (NVB) with exact inheritance of the affine field, repairs degenerate cells by replaying recorded midpoints, and uses the recorded parent–child relations to coarsen the mesh after training for levels of detail. On NeRF Synthetic, RevTet reaches 32.99 dB mean PSNR (local SVRaster: 33.25 dB). Importance-guided refinement and vertex motion improve a single-scene continuation, repair raises worst-cell quality at unchanged PSNR, and offline coarsening to 50% of the tetrahedra retains 31.48 dB.

Critical strengths
  • The formulation is careful. It separates field learning from mesh maintenance, states exactly what each edit preserves or loses (Eq. 4, Prop. 1), and checks validity explicitly; all reported states pass the geometry certificates with zero traversal failures.
  • Coarsening without retraining degrades gracefully: at 50% of the tetrahedra, PSNR drops by 1.5 dB, compared with 5.5 dB for SVRaster's native pruning at 50% of its voxels (Table 9), although primitive counts are not matched.
  • The paper is transparent about the limits of its evidence.
Critical weaknesses
  1. No comparison with the closest methods. The paper positions RevTet against Radiance Meshes, whose neural attribute field and Delaunay maintenance it aims to avoid, and Radiant Foam (p.2), but compares with neither quantitatively; Table 6 is structural only. The only quantitative baseline is SVRaster, against which RevTet is slightly lower (32.99 vs. 33.25 dB PSNR, 0.0441 vs. 0.0387 LPIPS; Table 1). No rendering speed is reported for any method, and the reported timings exclude the voxel-field initialization (p.6).

  2. The benefits of vertex motion and repair are not shown at convergence or in downstream terms.

    • Importance-guided refinement and vertex motion are evaluated only in one 6,000-step continuation of Lego from an iteration-1,000 checkpoint (+0.60/1.40 dB and +0.68/0.69 dB; App. D.4). On a converged checkpoint, the fixed/movable and fixed/reallocated arms differ by less than 0.002 dB (App. C.3, p.16).
    • Repair raises the worst-cell quality Q_min by 23.4% at unchanged PSNR, at the cost of 21.0% more training time and 14.2% more memory (Table 2). However, the no-repair runs also have no nonpositive cells and zero traversal failures (p.8), and the paper states that rendering speedups and fewer traversal failures remain unverified (p.9). What the better worst cells buy is therefore not demonstrated.
  3. The practical value of the level-of-detail hierarchy is not measured. Retaining 50% of the tetrahedra reduces the persistent tensor inventory only from 852.41 to 615.00 MiB (−27.9%), because the hierarchy itself is large (p.9, App. D.5), and no rendering time is reported at any tier. Importance-ranked coarsening improves over random selection by 0.11–0.33 dB on average, but at the 50% tier it wins on only 4 of 8 scenes, and Ship accounts for about three quarters of the mean gain (Table 9). The SVRaster comparison uses each method's own primitive fraction, which, as the paper notes, does not equalize storage or compute (Fig. 5).

  4. The evaluation scope is narrow. All experiments use the eight bounded NeRF Synthetic scenes with a single seed; the repair study covers four scenes and the motion study one scene (p.9). There are no real-world or unbounded scenes.

  5. Closely related work is missing. Tetra-NeRF (Kulhanek & Sattler, ICCV 2023) already represents radiance fields on tetrahedra with barycentric interpolation of vertex features. DefTet (Gao et al., NeurIPS 2020) and DMTet (Shen et al., NeurIPS 2021) optimize vertex positions of deformable tetrahedral grids, and DMTet combines this with subdivision. Coarsening of newest-vertex-bisection meshes has also been studied in adaptive finite elements (e.g., Chen & Zhang, 2010; Funken et al., 2011), whereas the paper cites only the refinement side (Maubach, 1995; Stevenson, 2008). Positioning against these works would clarify that the contribution lies mainly in combining these operators with explicit attribute transfer for radiance fields.

Questions for authors
  1. Which voxel model initializes RevTet, how long does that stage take, and what is its PSNR before tetrahedral training? This would show how much of the final quality comes from the 30k tetrahedral iterations.
  2. Could you compare with Radiance Meshes and Radiant Foam on the same scenes, reporting PSNR, training time, rendering FPS, and memory?
  3. Given that the converged-checkpoint control shows less than 0.002 dB difference, does vertex motion improve the final model under full training? A full-length frozen-versus-movable comparison on all eight scenes with several seeds would answer this.
  4. For each LOD tier, what are the size of a render-only export (without the hierarchy) and the rendering time? How does this compare at matched bytes with SVRaster, or with LOD-oriented representations such as Octree-GS or hierarchical 3DGS?
  5. Does repair have any measurable downstream effect, for example on traversal robustness under longer training, numerical stability, or rendering speed?
  6. Strict NVB is 0.31 dB below your earlier longest-edge variant (Table 8). Does conformity cost reconstruction quality under a matched recipe?
  7. How would the method handle unbounded real scenes (e.g., Mip-NeRF 360), given that boundary vertices are fixed (Eq. 8)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.