DecomVoxel: Harnessing 3D-Native Priors with Guided In-situ Denoising Optimization for Decompositional Scene Reconstruction

• Occlusions and limited viewpoints leave decompositional methods with incomplete geometry and incoherent appearance, while 2D priors lack 3D awareness and 3D-native priors can drift in complex scenes.The cited 2D-prior work includes [Ni et al. 2025; Yang et al. 2025b], while the cited 3D-native-prior work includes [Chen et al. 2025; Xia et al. 2026; Yu et al. 2025].Source: §1 · p. 2
• DecomVoxel formulates missing-region completion as guided in-situ denoising that integrates 3D-native priors with neural scene reconstruction.Its epsilon-based distillation loss for flow matching improves convergence stability and detailed geometry recovery over velocity-based optimization.Source: §1 · p. 2
• Adaptive spatial guidance uses occupied and vacant anchors with temporal annealing to constrain generative priors, suppress hallucinations, and mitigate spatial drift.Occupied anchors preserve coherence with observations; vacant anchors discourage generation where the object must not exist.Source: §1 · p. 2
• DecomVoxel consistently outperforms state-of-the-art baselines on geometry and appearance, especially under heavy occlusion, while preserving layout, structural fidelity, and style-consistent texture.It produces high-fidelity textured meshes with clean topology and complete geometry.Source: §1 · p. 2

2.1 3D Scene Representation and Reconstruction

• Sparse voxel-based differentiable rendering offers structured scene representations, while decompositional reconstruction seeks separate object and background representations but still struggles with occluded geometry.The paper integrates a 3D-native prior into voxel-based neural reconstruction to recover missing structures in situ while preserving scale and spatial alignment.Source: §2.1 · p. 3

2.2 Generative Priors for Decompositional Reconstruction

• 2D diffusion completion can produce multi-view inconsistency and Janus artifacts, while decoupled 3D generation often requires post-hoc alignment that accumulates scene-coordinate errors.The cited 2D-prior discussion includes [Ni et al. 2025; Xia et al. 2025; Yang et al. 2025b] and [Lin et al. 2023; Shi et al. 2023]; cited examples of 3D generation include MetaScenes [Yu et al. 2025] and SimRecon [Xia et al. 2026].Source: §2.2 · p. 3

2.3 Guidance for Denoising Optimization

• Prior implicit-feature guidance can retain alignment errors, while feed-forward point-cloud conditioning lacks per-scene precision; DecomVoxel instead uses explicit spatial guidance within iterative in-situ denoising.The related approaches include ReconViaGen [Chang et al. 2025], Geometry Forcing [Wu et al. 2025a], CUPID [Huang et al. 2025a], SAM3D [Chen et al. 2025], and ShapeR [Siddiqui et al. 2026].Source: §2.3 · p. 3

3.1 Preliminaries

• Sparse voxels connect scene reconstruction to TRELLIS's 3D-native prior, which jointly models geometry and appearance with multi-view consistency.The structured voxel representation also provides a basis for direct geometric anchoring during optimization.Source: §3.1 · p. 3

3.2 In-situ Denoising Optimization

• The epsilon-based objective replaces uniform velocity-loss gradients with timestep-dependent scaling, suppressing high-noise updates and strengthening fine refinement.Its gradient scales by (1 − t): it is suppressed as t approaches 1 and increases as t approaches 0.Source: §3.2 · pp. 4–5

3.3 Adaptive Spatial Guidance

• Occupied anchors preserve low-uncertainty voxels, while vacant anchors use redistributed pruning toward a near-zero void prior instead of zeroing features.Vacant regions combine raycast free space with areas occupied by the background or other objects; forcing their features to zero can destabilize optimization and cause rendering artifacts.Source: §3.3 · p. 5
• Temporal annealing begins with high-noise generation and strong anchors, then decays timestep bounds and guidance, finally deactivating structural preservation and vacant-zone pruning.This schedule first anchors high-confidence observed segments and suppresses vacant zones, then relaxes those constraints in the final iterations.Source: §3.3 · pp. 5–6

3.4 Optimization Pipeline

• The pipeline partitions an initial voxel scene, completes the background with inpainted color and plane-based depth, then refines object geometry and appearance in situ.The background mesh is extracted with Marching Cubes; object refinement proceeds through geometry completion followed by texture synthesis.Source: §3.4 · pp. 6–7

4.1 Settings

• The evaluation uses eight synthetic Replica scenes and six real-world ScanNet++ scenes, with every tenth view held out for testing.It assesses geometry, rendering, human-rated object geometry and appearance, and spatial layout against six decompositional reconstruction baselines; ShapeR is evaluated only on geometry.Source: §4.1 · p. 7, §4 · p. 7

4.2 Results

• DecomVoxel completes geometry while preserving the original spatial layout, and human evaluators favor its reconstruction quality and spatial alignment.The evaluation also reports sensitivity of point-cloud-conditioned methods to input density and incomplete geometry on complex real-world scenes such as ScanNet++.Source: §4.2 · p. 7
• Its appearance completion produces style-consistent textures in unobserved regions, maintains consistency with visible images, and receives the strongest expert assessment for plausibility.The passage contrasts this with 3D-native baselines whose textures deviate from source style and have inferior rendering metrics.Source: §4.2 · pp. 7–8

4.3 Ablation Studies

• Epsilon-based denoising, linear noise sampling, re-distributed pruning, and cosine-power preservation-weight scheduling yield stable convergence or higher reconstruction fidelity than the examined alternatives.The reported comparisons also find that fixed noise is less stable, hard pruning offers negligible improvement over no pruning, and velocity-based loss has higher late-stage error.Source: §4.2 · p. 8, Figure 5 · p. 8
• On Replica, the method trades longer runtime for better geometric completeness, texture consistency, and spatial alignment while maintaining comparable average memory.Table 4 reports runtime in hours and GPU memory in GB.Source: §4.2 · p. 8, Table 4 · p. 8
• Geometry and appearance priors improve reconstruction and rendering, while adaptive spatial guidance adds consistency; the full model performs best across all metrics.The ablations report that unguided geometry priors can introduce geometric deviations and unguided appearance priors can slightly reduce rendering performance.Source: §4.3 · p. 8, Table 3 · p. 8

5 Conclusion

• DecomVoxel bridges 3D-native generative priors and neural scene reconstruction through guided in-situ denoising optimization.The framework combines a reformulated epsilon-based distillation loss with adaptive spatial guidance to suppress generative hallucinations and mitigate spatial drift.Source: §5 · p. 8
• Evaluations on Replica and ScanNet++ show that DecomVoxel preserves structural fidelity and spatial layout while delivering textured meshes with clean topology, geometry, and appearance.The paper presents the framework as a robust solution for recovering high-fidelity object entities within complex scenes, with potential relevance to embodied AI and spatial intelligence.Source: §5 · p. 8
• The framework is presented as having potential to enhance physical-world perception and interaction in embodied AI and spatial intelligence.Source: §5 · p. 8

Supplementary Material

• The supplement presents the loss derivation, pipeline and implementation details, evaluation protocols, and limitations and future directions.These materials are organized into Sections A through D.Source: Supplementary Material · p. 14

A Epsilon-Based Distillation for Flow Matching

• Epsilon-based refinement minimizes the discrepancy between predicted and target noise residuals along the denoising path.The path is defined as x_t = (1 − t)x′ + tε, and the predicted noise estimate is derived from the flow model.Source: A Epsilon-Based Distillation for Flow… · p. 14
• The epsilon- and velocity-based objectives have equivalent gradient directions under the simplified SDS derivation, differing by a time-dependent scaling factor.The derivation omits the model Jacobian, which the passage characterizes as computationally expensive and often unstable.Source: A Epsilon-Based Distillation for Flow… · p. 14
• The epsilon-based loss reweights guidance by (1 − t), weakening gradients at high noise and yielding more robust convergence and superior geometric fidelity than velocity-based optimization.The passage describes conservative guidance at high noise and precise guidance at low noise.Source: A Epsilon-Based Distillation for Flow… · p. 14

B.1 Adaptive Spatial Guidance

• Occupied and vacant anchors are progressively relaxed alongside the denoising timestep, transitioning from preserving reliable observations to prior-driven refinement.Vacant anchors suppress occupancy in invalid regions.Source: §B.1 · p. 14
• Occupied-anchor certainty is derived from voxel geometric uncertainty, producing a soft grid used to preserve structural and appearance evidence.Uncertainty depends on voxel edge length and octree level; the certainty grid is smoothed and sampled for appearance refinement.Source: §B.1 · pp. 14–15
• Vacant anchors combine multi-view visibility evidence with regions occupied by the background or another object, giving occupied anchors precedence in overlap.A cell is treated as high-confidence vacant when it lies in front of observed surfaces in at least three views.Source: §B.1 · p. 15
• Re-distributed pruning shifts vacant-cell features toward a void prior rather than setting them to zero, discouraging persistent occupancy without collapsing features.The strategy resamples vacant-cell features and applies redistribution during early geometry optimization.Source: §B.1 · p. 15
• Timestep ranges expand toward lower noise while occupied-anchor weights decay, relaxing spatial constraints over geometry and appearance refinement.Geometry scheduling spans the first 4,000 iterations and appearance scheduling the first 2,000; the appearance-preservation weight starts at 2 and decays over 3,000 iterations.Source: §B.1 · p. 15

B.2 Optimization Parameters

• Geometry optimization runs for 5,000 iterations and appearance optimization for 3,000, using Adam with cosine learning-rate decay and frozen TRELLIS modules.The table caption identifies Table 5 as listing principal optimization parameters.Source: §B.2 · p. 15, Table 5 · p. 15

C. Evaluation Details

• The evaluation protocol covers datasets, metrics, baseline settings, and the user-study procedure.Source: C Evaluation Details · p. 15

C.1 Datasets

• The evaluation includes eight Replica scenes with synthetic reference geometry and controlled renderings, and six ScanNet++ scenes with incomplete, noisy real captures.Source: §C.1 · p. 15
• All methods use the same scene split, camera poses, and object masks; ScanNet++ masks may be incomplete, and the real-world qualitative examples use SAM2-generated masks.The settings evaluate moderate mask noise and occlusion without manually repairing every object mask.Source: §C.1 · p. 16

C.2 Reconstruction Metrics

• CD, reported in cm, F-score at a 5 cm threshold, and NC are used to evaluate 3D scene and object reconstruction.The metrics are defined in Table 6.Source: §C.2 · p. 16
• Table 6 states that P and P* are point clouds sampled from the predicted and ground-truth meshes.Source: Table 6 · p. 16
• Table 6 identifies n_p as the normal vector at point p.Source: Table 6 · p. 16

C.3 Rendering Metrics

• Held-out-camera renders are evaluated within benchmark object or scene masks, using unit-range PSNR and SSIM and VGG LPIPS inputs scaled to [−1, 1].All methods use the same camera parameters and evaluation code.Source: §C.3 · p. 16

C.4 Baseline Protocol

• Baseline protocols use official implementations and parameters, with shared view selection for SimRecon and SAM3D and method-specific input or alignment procedures for other baselines.ReconViaGen and CUPID outputs are aligned to scene coordinates using FoundationPose.Source: §C.4 · p. 16

C.5 User Study Protocol

• The user study collected 44 questionnaires from unique expert raters, with 25 questions covering object geometry, appearance, and scene layout.Questions were sampled from eight Replica and six ScanNet++ scenes; the study was blind and randomized and used a five-point scale.Source: §C.5 · p. 16
• Preference tests were significant for scene layout, object geometry, and object appearance, with p-values 8.45 × 10−75, 1.50 × 10−58, and 4.26 × 10−36, respectively.The study used two-sided exact binomial tests against a null preference probability of 0.5.Source: §C.5 · p. 16

D Limitations and Future Work

• DecomVoxel depends on an initial GeoSVR reconstruction, object masks, and a pretrained 3D prior.Severely inaccurate masks, poor initialization, or extreme occlusion can leave too little reliable evidence for occupied anchors; out-of-distribution objects may not be faithfully completed.Source: D Limitations and Future Work · p. 16
• The current implementation uses the object-centric TRELLIS prior and is evaluated primarily on indoor scenes with well-defined object instances.Street-scene extension is proposed provided reliable initial reconstruction and object decomposition are available; non-object regions such as sky or grassland remain outside scope.Source: D Limitations and Future Work · p. 16
• Per-object geometry and appearance optimization is slower than a single feedforward pass, although multiple objects can be optimized in parallel.Source: D Limitations and Future Work · p. 17
• Future work could reduce latency by distilling the optimization process or developing more efficient guidance schedules while preserving reconstruction quality and spatial fidelity.Source: D Limitations and Future Work · p. 17