acceptodds
Under review as a conference paper at ICLR 2027

RRA-Avatar: Representation Responsibility Allocation for Gaussian Avatars with Loose Garments

Abstract

Animatable Gaussian avatars typically use a single body-guided representation for both body-supported surfaces and off-body structures such as loose garments, hair, and accessories. This creates a representation-responsibility ambiguity: regions with substantially different levels of body-prior support are forced to share geometry, deformation, and optimization capacity. We introduce RRA-Avatar, a support-aware dual-layer Gaussian avatar framework that assigns representation responsibility according to whether foreground geometry is adequately explained by the fitted body prior, rather than according to garment semantics. The body-aligned layer preserves stable SMPL-guided modeling, while the residual-surface layer provides dedicated capacity for unexplained off-body structures. To instantiate this decomposition without garment templates or semantic labels, we recover residual-surface seeds from temporally fused multi-view foreground evidence. We further introduce a responsibility-allocation mechanism that coordinates layer-specific supervision, gradient flow, and topology evolution, reducing inter-layer competition while retaining joint refinement near uncertain boundaries. Experiments on DNA-Rendering, I3D-Human, and ZJU-MoCap demonstrate substantially improved reconstruction of loose and off-body structures while preserving competitive body-region and novel-pose performance. On six DNA-Rendering subjects, RRA-Avatar reduces LPIPS from 30.91 to 22.98 relative to the strongest compared baseline under the same protocol.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Animatable Human Avatars, 3D Gaussian Splatting, Neural Rendering, Loose-Garment Reconstruction, Multi-View 3D Reconstruction, Human Motion Modeling, Layered Representations, Novel-View Synthesis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PwuWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary
  1. RRA-Avatar reconstructs animatable Gaussian avatars from multi-view videos, allocating body-aligned and residual-surface representations according to body-prior support rather than garment semantics.
  2. It initializes residual Gaussians using SMPL-guided shell sampling, multi-view foreground evidence, and temporal fusion, and uses separate deformation heads for the two layers.
  3. A staged training mechanism coordinates supervision, gradient scaling, and Gaussian topology updates to reduce inter-layer competition. Experiments on three datasets and component ablations support improved reconstruction of loose garments and off-body structures.
Critical strengths
  1. The dual-layer representation improves the rendering of loose garments and off-body structures, preserving clearer boundaries and details in the qualitative comparisons. On DNA-Rendering, Off-LPIPS decreases from 0.3962 for SeqAvatar to 0.3452 (Fig. 3 and Table 1).
  2. The residual initialization offers a practical way to identify off-body structures from multi-view masks and fitted body models, without requiring garment templates, semantic clothing partitions, or direct 3D supervision. Temporal fusion helps filter out candidates with insufficient cross-frame support (Sec. 3.3).
  3. Coordinating supervision, gradient updates, and topology growth provides a clear mechanism for maintaining layer specialization during training. The ablations support its effectiveness: adding responsibility allocation to the model with residual discovery and separate deformation heads reduces LPIPS* from 27.62 to 22.98 (Table 3).
Critical weaknesses
  1. The paper discusses several layered avatar methods, such as LayGA, D3GA, and Gaussian Wardrobe, but does not include direct experimental comparisons with them. This limits the evidence that the proposed support-based partition and responsibility-allocation mechanism offer advantages over existing layered representations.
  2. The residual layer still undergoes SMPL-guided coarse deformation, although loose garments can move substantially differently from the underlying body. This may constrain the modeling of weakly body-coupled clothing motion, and the suitability of this deformation assumption for challenging loose garments is insufficiently validated (Sec. 3.2).
  3. The training configuration varies substantially across datasets: residual guidance lasts 5,000 iterations on DNA-Rendering but only 100 on the other two datasets, and its loss weight changes from 1.0 to 0.001. Without sensitivity analysis or clear selection guidelines, the tuning effort and reliability when training on new datasets remain unclear (Appendices A.3 and B.1).
Questions for authors
  1. Why are the layered or hybrid methods discussed in the paper, including LayGA, D3GA, Gaussian Wardrobe, MonoCloth, ReLoo, and FreeCloth, absent from the comparisons? Given available public implementations, could you include suitable baselines under a matched protocol?
  2. Despite separate local deformation heads, why do both layers use SMPL-guided coarse deformation? Could you justify this choice for loose garments and compare it with a more independent residual-layer deformation mechanism?
3 comments

We thank the reviewer for the helpful comments.

[Layered baselines] We agree that direct comparisons with these layered methods would be informative. However, the public resources we inspected have input or implementation requirements that prevent direct evaluation under our matched RGB protocol.

  • FreeCloth requires scanned point-cloud and normal supervision and does not provide a native learned RGB appearance model; D3GA requires clothing-part segmentation and manually prepared cage assets.
  • LayGA and MonoCloth do not provide released training implementations in the official resources we inspected.
  • Gaussian Wardrobe and ReLoo provide training code but omit parts of the preprocessing workflows needed for custom datasets.

We therefore report quantitative RGB comparisons under a matched evaluation protocol. We also recognize the contributions of the layered and hybrid approaches discussed above and explain their methodological relationship to RRA-Avatar in the related work (Sec. 2, p. 3, lines 127–139).

[SMPL-guided residual deformation] We agree that loose garments can move differently from the body. However, articulated SMPL skinning can still provide a useful coarse motion reference. To test this, we compared the full model with a root-only residual variant that retains global root rotation and translation and removes articulated skinning. Its deformation head predicts non-rigid motion in the root coordinate frame. The body layer, initialization, network capacity, input features, training budget, losses, and topology settings remain unchanged.

Under the same novel-view protocol on six DNA-Rendering subjects:

Table R1. Residual-layer deformation on DNA-Rendering.

Residual deformationPSNR↑SSIM↑LPIPS∗↓Off-LPIPS∗↓
Root-only32.170.974424.42358.7
SMPL-guided, full model32.520.976022.98345.2

Retaining articulated SMPL guidance improves PSNR by 0.35 dB and reduces LPIPS∗ and Off-LPIPS∗ by 5.9% and 3.8%. SMPL-guided skinning provides a coarse reference for global motion and first-order articulation, while the independent residual head learns clothing-specific deviations from this reference. Thus, sharing the coarse prior does not impose identical motion on the two layers.

[Sensitivity and parameter selection] We added a sensitivity analysis in Appendix C.11, summarized in Sec. 4.3. We sweep λR∈10−3,10−2,10−1,1\lambda_{\mathrm R}\in\\{10^{-3},10^{-2},10^{-1},1\\} and halve or double the default guidance duration, varying one parameter at a time and keeping total training iterations fixed. Across three datasets, the largest within-dataset ranges are 0.07 dB PSNR, 0.0003 SSIM, 0.47 LPIPS∗, and 7.6 Off-LPIPS∗, indicating limited sensitivity over the tested ranges. Dataset-level defaults are fixed across subjects without test-metric tuning. For new datasets, the configurations in Appendix B.1 provide practical starting points: stronger residual supervision on DNA-Rendering (λR=1\lambda_{\mathrm R}=1) and weaker supervision on I3D-Human and ZJU-MoCap (λR=10−3\lambda_{\mathrm R}=10^{-3}). We recommend assessing the extent and reliability of training-view residual evidence: stronger, longer guidance for large, consistent residual regions and weaker, shorter guidance for smaller or less reliable regions.

Table R3. Sensitivity to residual-guidance weight and duration (paper Table 15).

DatasetλR\lambda_{\mathrm R}DPSNR↑SSIM↑LPIPS∗↓Off-LPIPS∗↓
DNA-Rendering10⁻³5,00032.500.976022.96346.0
10⁻²5,00032.530.976023.00342.4
10⁻¹5,00032.500.976022.87343.6
15,00032.520.976022.98345.2
12,50032.540.976222.68338.8
110,00032.530.976222.60338.4
I3D-Human10⁻³10032.350.967129.28281.9
10⁻²10032.310.967129.31280.7
10⁻¹10032.280.967029.30281.4
110032.330.967129.19280.0
10⁻³5032.320.967129.22280.8
10⁻³20032.280.967129.22280.6
ZJU-MoCap10⁻³10031.060.962327.47274.4
10⁻²10031.060.962327.42274.7
10⁻¹10031.010.962127.65274.9
110031.030.962227.68275.6
10⁻³5031.010.962127.62275.6
10⁻³20031.030.962127.89276.0

Thank you for the detailed response and the additional experiments. With only about five hours left in the discussion period, I understand that running further experiments takes considerable time and effort. The sensitivity study and your analysis of SMPL-guided deformation address my questions about the residual-guidance settings and the role of the SMPL prior. I also understand why the limitations of the publicly available implementations and preprocessing pipelines make a fair comparison with existing layered methods difficult under the same protocol.

I had lowered my score to 2 because my questions were still open on the final day of the rebuttal period. After reading your response, I am restoring my original score of 3. Thank you for taking the time to address my concerns.

We are very pleased that our rebuttal has addressed your concerns. Sincerely thank you for your valuable time.

Reviewer_MXjXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a dual-layer Gaussian avatar framework for reconstructing animatable humans with loose garments. The key idea is to separate body-aligned and residual-surface Gaussians according to their support from the fitted body prior. The method initializes residual Gaussians using multi-view silhouette evidence and introduces a responsibility-allocation mechanism that coordinates supervision, gradient updates, and Gaussian densification. Built upon SeqAvatar's motion modeling architecture, RRA-Avatar achieves improved reconstruction quality, particularly for loose garments and off-body regions.

Critical strengths
  1. The paper addresses a practical limitation of SMPL-guided Gaussian avatars, where loose garments and off-body structures are difficult to represent accurately. Separating body-supported and residual regions is an intuitive and well-motivated approach.

  2. The method achieves consistent improvements across multiple datasets, with particularly notable gains on DNA-Rendering. The quantitative results and visual comparisons demonstrate improved reconstruction quality for loose garments while maintaining competitive performance on novel-pose animation.

Critical weaknesses
  1. The proposed framework is largely built upon SeqAvatar, with additional residual Gaussian initialization, layer-specific deformation heads, and staged optimization. Related works such as FreeCloth and LayGA have already explored region-specific modeling and layered Gaussian representations. Although the proposed combination is reasonable, its fundamental distinction from existing approaches remains somewhat limited.

  2. The paper introduces a coordinated mechanism involving supervision, gradient scaling, and topology control, but its effectiveness is not thoroughly analyzed. In particular, the ablation results show that removing any individual component leads to only marginal degradation, despite the substantial improvement attributed to the complete mechanism. Additional visualizations of the allocation process and sensitivity analyses of the training schedules and loss weights would help substantiate the claimed contributions.

Questions for authors
  1. The responsibility-allocation schedules and loss weights vary considerably across datasets (Appendix A.3 and B.1). For instance, the residual-opacity loss weight differs by three orders of magnitude, and the guidance duration ranges from 100 to 5,000 iterations. How sensitive is the method to these choices, and how are the hyperparameters determined for a new dataset? A sensitivity analysis would help evaluate the robustness and generalizability of the proposed mechanism.

  2. The proposed framework builds upon SeqAvatar and shares similar motivations with FreeCloth and LayGA regarding region-specific modeling and layered representations. Could the authors clarify the key methodological novelty of the proposed responsibility-allocation mechanism beyond combining body-prior-aware partitioning, layer-specific deformation, and staged optimization?

3 comments

We thank the reviewer for the constructive suggestions.

[Methodological novelty] We acknowledge that we build on SeqAvatar. Its hierarchical skeletal and point-wise motion encoder is effective. However, it fits body-supported and off-body structures in one Gaussian set. We retain its encoder and use multi-view evidence to split Gaussians by body support. This split selects deformation heads and directs supervision, updates, and topology. On six DNA-Rendering subjects, LPIPS∗/Off-LPIPS∗ fall from 30.91/396.2 to 22.98/345.2 (Table 1).

In Sec. 2 (p. 3), we discuss LayGA and FreeCloth (lines 132–134) and distinguish our approach (137–139).

  • Region-specific modeling. FreeCloth routes points via clothing-cut maps to copying, LBS deformation, or free-form generation, using scanned points and normals as supervision. We fuse multi-view foreground evidence over time to seed residual Gaussians beyond the fitted body, then render and optimize both layers jointly from RGB without garment labels or scans.
  • Layered representation. LayGA reconstructs and segments an avatar, then uses reconstructed clothing geometry to supervise separate body and clothing Gaussian maps for transfer. Our layers follow body-prior support. The supported waist and off-body hem of one garment can thus occupy different layers.
  • Responsibility allocation. FreeCloth uses branch-specific losses and LayGA retains segmentation constraints. During guidance, one evidence-derived assignment governs all three controls. We supervise residual opacity in assigned regions, reduce body-layer and shared-module updates, and restrict topology changes to the residual layer. Joint refinement follows.

Table R1 compares our model with single-set, random-split, and distance-based controls.

Table R1. Controlled comparisons on DNA-Rendering (paper Table 5).

VariantFinal GaussiansPSNR↑SSIM↑LPIPS∗↓
Single set (matched budget)110,95732.400.974926.80
Random two-layer split111,79332.410.974926.76
Distance-based split111,12932.370.974826.90
ours130,95232.520.976022.98

[Allocation components] Full allocation lowers whole-image LPIPS∗ from 27.62 to 22.98. Each ablation keeps two controls. LPIPS∗ measures whole-image perceptual error. Off-LPIPS∗ averages spatial LPIPS over foreground pixels beyond a margin around the projected SMPL mask. Blurred hem folds can raise Off-LPIPS∗ with little effect on the whole-image score. We agree that each ablation changes LPIPS∗ only slightly. However, removing residual-opacity supervision, topology gating, or gradient scaling raises Off-LPIPS∗ by 5.5%, 6.9%, and 4.8%, respectively. Each control helps off-body rendering. At similar counts (130,632 vs. 130,952), removing gradient scaling raises LPIPS∗ by 0.04 and Off-LPIPS∗ by 16.6. Fig. 6 shows that full allocation cuts residual-opacity leakage inside the SMPL mask from 51.56% to 32.86%.

Table R2. Ablation (paper Table 4).

VariantFinal GaussiansLPIPS∗↓Off-LPIPS∗↓
(c) Residual discovery + layer-specific deformation heads110,28427.62379.4
(d₁)=(d) w/o supervision127,65823.27364.3
(d₂)=(d) w/o topology gating124,44623.16368.9
(d₃)=(d) w/o gradient scaling130,63223.02361.8
(d) Full allocation130,95222.98345.2

[Sensitivity and parameter selection] We vary residual-opacity weight λ or halve/double default duration D, one at a time with total steps fixed (Sec. 4.3, Appendix C.11). The largest within-dataset ranges are 0.07 dB PSNR, 0.0003 SSIM, 0.47 LPIPS∗, and 7.6 Off-LPIPS∗, indicating limited sensitivity. For new datasets, we recommend the Appendix B.1 starting point that best matches the extent and reliability of training-view residuals. Large, consistent regions favor stronger, longer guidance (λ_R=1). Smaller or less reliable regions favor weaker, shorter guidance (λ_R=10⁻³).

Table R3. Guidance sensitivity (paper Table 15).

Datasetλ_RDPSNR↑SSIM↑LPIPS∗↓Off-LPIPS∗↓
DNA-Rendering10⁻³5,00032.500.976022.96346.0
10⁻²5,00032.530.976023.00342.4
10⁻¹5,00032.500.976022.87343.6
15,00032.520.976022.98345.2
12,50032.540.976222.68338.8
110,00032.530.976222.60338.4
I3D-Human10⁻³10032.350.967129.28281.9
10⁻²10032.310.967129.31280.7
10⁻¹10032.280.967029.30281.4
110032.330.967129.19280.0
10⁻³5032.320.967129.22280.8
10⁻³20032.280.967129.22280.6
ZJU-MoCap10⁻³10031.060.962327.47274.4
10⁻²10031.060.962327.42274.7
10⁻¹10031.010.962127.65274.9
110031.030.962227.68275.6
10⁻³5031.010.962127.62275.6
10⁻³20031.030.962127.89276.0

[Ethics] We collected no new human appearance data. Our Ethics Statement covers data sources, privacy, consent, misuse (unauthorized reconstruction and impersonation), licenses, access agreements, and restrictions on sharing identity-specific avatars.

Thanks for authors responses. It seems that the performance shows relative robust on different guidance strength. And the response partly resolves my concern. I decide to raise my rating to 3.

We are very pleased that our rebuttal has addressed your concerns. Thank you for your valuable feedback and time.

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

RRA-Avatar proposes a dual-layer Gaussian avatar reconstructed from calibrated multi-view videos. Body-aligned Gaussians represent surfaces supported by a fitted SMPL/SMPL-X prior, while residual-surface Gaussians are initialized through shell sampling, multi-view silhouette-residual voting, and temporal fusion. The layers share a motion encoder but use separate deformation heads. A staged optimization procedure coordinates body and foreground supervision, residual-opacity guidance, update scaling, and layer-specific topology changes. Experiments report substantial novel-view improvements on six DNA-Rendering subjects, smaller improvements on I3D-Human, and competitive ZJU-MoCap results.

Critical strengths
  1. The representation split addresses a concrete limitation of body-guided avatars. Allocating capacity according to observed body-prior support is more closely connected to reconstruction difficulty than assigning an entire garment to one semantic layer. The discovery procedure provides off-body support without garment templates or direct 3D surface supervision, and the method connects this initialization to deformation and optimization. This is a useful design for loose sleeves, hems, and other structures whose geometry is poorly represented by the fitted body.

  2. The principal novel-view result is consistent and visually supported. On DNA-Rendering, LPIPS* improves from 30.91 to 22.98 relative to SeqAvatar, with improvements on all six subjects. Off-LPIPS improves from 0.3962 to 0.3452, and the ranking persists across all four proxy margins in Appendix C.2. The supplied comparison videos show clearer garment boundaries and textile details. The paper also retrains baselines under common data/evaluation protocols and explicitly qualifies Off-LPIPS as appearance quality in an image-space proxy, rather than independent geometric ground truth. These choices strengthen the empirical case.

Critical weaknesses
  1. The evidence does not yet isolate the claimed benefit of coordinating all three allocation controls from the accompanying optimization and capacity choices. Table 3 improves LPIPS* from 27.62 to 22.98 after adding allocation, but removing supervision, topology gating, or update scaling individually gives 23.27, 23.16, and 23.02. This may reflect redundancy or interactions, but it leaves the necessity of the proposed coordination unclear. Appendix C.7 usefully controls initial Gaussian count and extra deformation heads under a stated common training protocol; it does not specify exactly how the loss/regularization and update schedules are instantiated for each control, or report their final Gaussian counts. The full model ends with 1.18 times SeqAvatar's Gaussians. A schedule-matched, comparable-final-capacity control would make the central interpretation more convincing.

  2. The supplied core implementation has material interface inconsistencies that prevent it from substantiating the reproducibility claim. train.py calls render with body_only and free_only arguments, but the supplied renderer accepts neither. The camera loader passes cloth_res_mask and smpl_mask to a Camera constructor that does not accept them. GaussianModel also lacks the layer-label and masked-topology interfaces invoked by training, while the supplied NonrigidDeformer exposes one deformation-head set rather than the stated two. These are gaps in the submitted model integration, beyond installation or missing licensed assets. They may be packaging errors and do not establish that the reported results are incorrect, but a consistent version is needed to assess how the central method and ablations were instantiated.

Questions for authors
  1. For Table 3 and Appendix C.7, which image-loss weights, AIAP schedules, optimizer restarts, and topology settings are shared across controls? Can you provide final Gaussian counts and an available single-set or no-allocation comparison with the same optimization schedule and comparable final capacity?

  2. Can you identify or provide the consistent source version used for Tables 1 and 3, including the actual dual-head model, layer-aware renderer, residual-seed integration, and ablation switches? The interfaces in the current package do not match the calls described above.

  3. The largest loose-garment gains use held-out views of the first 100 DNA-Rendering frames, while novel-pose testing uses the more moderate I3D-Human setting. If the DNA-Rendering novel-pose split already present in the loader has been evaluated, do the off-body gains persist there? This would materially affect the interpretation of loose-garment animation generalization.

1 comment

We thank the reviewer for the detailed comments.

[Allocation components]

Full allocation lowers whole-image LPIPS∗ from 27.62 to 22.98. Each ablation keeps two controls, and whole-image LPIPS∗ can be less sensitive to changes confined to off-body regions. LPIPS∗ measures whole-image perceptual error. Off-LPIPS∗ averages spatial LPIPS over foreground pixels beyond a margin around the projected SMPL mask. For example, if an ablation blurs folds at a loose garment hem while the rest of the image looks similar, whole-image LPIPS∗ may change little because it pools perceptual differences across the image. We agree that each ablation changes LPIPS∗ only slightly. However, removing residual-opacity supervision, topology gating, or gradient scaling raises Off-LPIPS∗ by 5.5%, 6.9%, and 4.8%, respectively. Each control helps off-body rendering. At similar counts (130,632 vs. 130,952), removing gradient scaling raises LPIPS∗ by 0.04 and Off-LPIPS∗ by 16.6. Fig. 6 shows that full allocation cuts residual-opacity leakage inside the SMPL mask from 51.56% to 32.86%.

Table R1. Allocation ablation on DNA-Rendering (paper Table 4).

VariantFinal GaussiansLPIPS∗↓Off-LPIPS∗↓
(c) Residual discovery + layer-specific deformation heads110,28427.62379.4
(d₁) Allocation w/o supervision127,65823.27364.3
(d₂) Allocation w/o topology gating124,44623.16368.9
(d₃) Allocation w/o gradient scaling130,63223.02361.8
(d) Full allocation130,95222.98345.2

[Training protocol and Gaussian counts]

We originally treated three-stage densification as part of our method, so we did not apply it to the single-set, random-split, and radial-split controls. The controls now use the full model's three-stage optimization and densification timings. The table gives the shared image and AIAP loss weights for each stage.

StageL1 / opacity / SSIM / LPIPSAIAP position / covariance
11 / 0.1 / 0.01 / 0.011 / 100
21 / 0.1 / 0.01 / 0.010.25 / 25
31 / 0.1 / 0.1 / 0.11 / 100

Optimizers and learning-rate schedulers restart at the same stage boundaries. Variants differ in the representation or allocation component tested. The three controls start with a fixed Gaussian budget matched to RRA-Avatar's dataset-average initial count. We agree that comparable final counts would help assess the role of model capacity. Matching the initial budget and schedule gives a common basis for comparison. However, final counts can differ because cloning, splitting, and pruning depend on learned gradients and opacities, and topology gating determines which Gaussians are eligible for these operations. Tables R1–R2 give the measured counts (paper Tables 4–5). The schedule-matched single-set control has an LPIPS∗ of 26.80 with 110,957 Gaussians, compared with 22.98 and 130,952 for the full model.

Table R2. Controlled comparisons on DNA-Rendering (paper Table 5).

VariantFinal GaussiansPSNR↑SSIM↑LPIPS∗↓
Single set (matched budget)110,95732.400.974926.80
Random two-layer split111,79332.410.974926.76
Distance-based split111,12932.370.974826.90
RRA-Avatar130,95232.520.976022.98

[Code consistency] While preparing the supplement, we inadvertently packaged different versions of the training script and several supporting modules, which caused the interface mismatches. The corrected Code Supplement integrates the dual-head model, layer-aware renderer, camera-mask interfaces, masked topology operations, residual-seed initialization, and explicit ablation entry points. We will also publicly release the source code upon acceptance.

[DNA novel-pose generalization]

We originally followed SeqAvatar's protocol and evaluated novel poses on I3D-Human, which provides separate motion sequences for the same subject and outfit. DNA-Rendering lacks equivalent paired sequences, so we originally used it for novel-view evaluation. Following your suggestion, we now test all six DNA-Rendering subjects on held-out poses. We train on frames 0–99 and test on 20 uniformly sampled frames from 100–149, with six uniformly sampled camera views per pose. All methods use the same protocol without test-time optimization. RRA-Avatar leads on all four metrics. Compared with SeqAvatar, LPIPS∗ drops from 52.34 to 49.30 and Off-LPIPS∗ from 440.1 to 422.3. See Sec. 4.2, Table 2, and Fig. 4 for the setup, results, and renderings.

Table R3. Novel-pose results on DNA-Rendering (paper Table 2).

MethodPSNR↑SSIM↑LPIPS∗↓Off-LPIPS∗↓
3DGS-Avatar24.470.930862.75477.0
GART22.660.920785.19590.6
GauHuman25.830.939857.54464.8
ToMiE25.790.939560.29472.5
SeqAvatar25.820.940052.34440.1
RRA-Avatar25.900.940449.30422.3

Comments on the paper

1 comment

Dear ACs,

Thank you for your patience with our late response. We especially thank Reviewer PwuW for following up and confirming that the sensitivity study and the SMPL-guided deformation experiment addressed their concerns.

In response to Reviewer nqyS’s concern about the allocation ablations, we clarified that adding all three controls lowers whole-image LPIPS∗ from 27.62 to 22.98, while removing one control with the other two retained increases it by only 0.2–1.3%. Since whole-image LPIPS∗ can dilute changes confined to off-body regions, we also report Off-LPIPS∗ in the revised Table 4. Removing residual-opacity supervision, topology gating, or gradient scaling increases Off-LPIPS∗ by 5.5%, 6.9%, and 4.8%, respectively.

We have responded to each concern point by point. We would appreciate it if Reviewers nqyS and MXjX could revisit their assessments in light of our responses and the additional evidence.

Thank you again for your time and consideration.

Best regards,
The authors of Submission 59580

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.