acceptodds
Under review as a conference paper at ICLR 2027

Native 3D Human Generation with Geometry-Aligned Soft Rigging from a Single Image

Abstract

Generating high-fidelity, riggable 3D human avatars from a single image remains a fundamental challenge. Native 3D models suffer from limited data and poor generalization. Meanwhile, 2D-inspired methods rely on multi-view or video supervision, which introduces cross-view inconsistency and unstable convergence. To address these limitations, we present NVWA, a native 3D generation framework that reconstructs geometry-aligned, riggable avatars from one image. Our method makes two key innovations. First, rather than supervising with multi-view renderings or Internet videos, we synthesize large-scale explicit 3D human assets for direct native 3D supervision. This eliminates the multi-view inconsistency inherent in distillation-based methods and enables substantial gains in geometric fidelity. Second, we depart from predefined parametric topologies (e.g., SMPL). Such templates constrain outputs to naked-body shapes and cannot model loose clothing or accessories. Instead, we propose soft rigging. Our model jointly generates free-form volumetric geometry and a spatially aligned semantic correspondence field. This field establishes per-vertex correspondence to a standard rigging template without restricting the generated shape. Shape generation is thus decoupled from template constraints while full pose controllability is preserved. To support training, we introduce a scalable pipeline for synthesizing photorealistic explicit 3D humans. We also release NVWA-100K, a dataset of 100,000 high-fidelity textured mesh models. Experiments demonstrate that NVWA achieves state-of-the-art fidelity and rigging quality, significantly advancing single-image 3D human generation.

Primary area:
generative models
Keywords:
native single-image 3D human generation, animatable human avatars

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_vUYQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents NVWA, a single-image 3D human generation framework targeting both high-fidelity reconstruction and animation-ready output. The method has three main ingredients: (1) NVWA-100K, a dataset of 100K explicit textured human meshes synthesized from human images; (2) a geometry-aligned soft-rigging representation that generates free-topology geometry together with per-vertex semantic UV correspondences to SMPL-X, allowing downstream SMPL-X fitting and skinning without forcing the generated surface to adopt SMPL-X topology; and (3) adaptation of the TRELLIS.2 native-3D backbone to the clothed-human domain, followed by real-scan post-training.

Critical strengths
  1. The strongest contribution is the separation of the generated surface topology from the body-template topology. NVWA generates a free-topology mesh while predicting semantic UV coordinates referenced to SMPL-X, after which these correspondences guide SMPL-X fitting and skinning-weight transfer.
  2. This paper performs strongly across several complementary reconstruction measures. In Table 1, NVWA obtains the best reported X-Humans CD and CLIP score, as well as the best reported in-the-wild CLIP and LPIPS . The results are not uniformly dominant—e.g., PSHuman is stronger on X-Humans ULIP-2 and LPIPS—but the aggregate performance is convincing.
Critical weaknesses
  1. A central argument is that directly training on explicit 3D assets avoids the cross-view inconsistency of multi-view supervision. However, NVWA-100K itself is produced by first running PSHuman's multi-view diffusion model to generate six RGB images and normal maps, followed by explicit reconstruction and appearance projection. Thus, the current paper demonstrates that training on the resulting explicit meshes works well, but it does not demonstrate that multi-view inconsistencies have actually been eliminated.
  2. The main soft-rigging contribution lacks quantitative rigging and animation evaluation. For animation, the manuscript reports results on 20 AMASS motion sequences but presents the comparison to IDOL and LHM primarily through qualitative figures.
Questions for authors
  1. Can the authors provide a controlled comparison between explicit 3D supervision and multi-view supervision using the same source identities, TRELLIS.2 initialization, training budget, and image-generation pipeline? For example, one condition could train from the six synthesized PSHuman views while another uses the explicit meshes reconstructed from exactly those views.
  2. Do the authors have quantitative measurements of soft-rigging accuracy or animation quality, such as SMPL-X joint/vertex error, correspondence accuracy.
  3. I think this paper is too similar to Synchuman; please explain in detail the differences between these two papers.
5 comments
AuthorsOpenReview ↗

Reply to Reviewer vUYQ, Part 1

Reply to Reviewer vUYQ, Part 1

We thank you for your time and valuable comments. Our responses to each point are as follows.

Weakness 1 and Question 1. Explicit 3D supervision and cross-view inconsistencies

Thank you for suggesting this paired comparison. We agree that it would help quantify the contribution of explicit 3D supervision. In our TRELLIS.2-based pipeline, the diffusion transformer is trained with a flow-matching objective whose target is constructed from a clean 3D latent encoded from a mesh. The six PSHuman RGB views and their corresponding normal maps provide observations of the surface from different viewpoints, but they do not directly provide the clean 3D latent required by this objective.

A view-only baseline initialized from the same TRELLIS.2 checkpoint would therefore require a different training objective, for example an image-space loss comparing renderings of the predicted asset with the six source views. We have not evaluated this baseline under matched source identities, initialization, image-generation pipeline, and training budget. Our experiments establish that training on reconstructed 3D assets is feasible: each sample provides a single 3D target, so the generator does not have to fit six potentially conflicting 2D targets separately.

We agree with the reviewer that cross-view inconsistencies can still leave artifacts in the reconstructed mesh. To limit their influence on training, we manually inspected multiview renderings of all 80,495 meshes in the random-pose subset and assigned each mesh an overall quality score from 1 to 10 based on visible geometry and appearance defects. We retained meshes scoring 5 or above. Even when the front view appeared intact, a missing hand or foot or severe fragmentation visible from a side view was sufficient for rejection.

The scoring guidelines are described in Appendix A.7 (pp. 24–25). Scores of 1–3 indicate severe defects, such as distorted faces or hands, missing extremities, or fragmented garments and accessories. Scores of 4–7 indicate a largely intact body with localized defects, such as mild limb deformation, hand color projected onto clothing, or streaks and gaps on the legs or shoulders. Scores of 8–10 indicate few visible defects across the inspected views; fused fingers alone could still fall in this range. Fig. 17 (p. 32) shows examples below the retention threshold.

Table 7 (p. 25) reports the following distribution for the random-pose subset:

Quality scoreNumber of meshesPercentage
s ≤ 34,8306.0%
3 < s < 59,65912.0%
s = 59,01511.2%
5 < s ≤ 734,29042.6%
s > 722,70128.2%
Total80,495100.0%

After filtering, 66,006 random-pose meshes (82.0%) met the threshold. Together with the 23,800 A-pose meshes retained separately, this yielded 89,806 meshes for training.

AuthorsOpenReview ↗

Reply to Reviewer vUYQ, Part 2

Reply to Reviewer vUYQ, Part 2

Weakness 2 and Question 2. Quantitative evaluation of soft rigging and animation

We added a controlled comparison on 40 in-the-wild videos. For each video, we used the same NVWA-generated mesh and varied only the rigging method: NVWA’s native soft rigging, UniRig, or Make-It-Animatable (MIA). All three rigs were evaluated with the same target motion, camera, and frames. Native soft rigging achieved the lowest mean normalized 2D body-keypoint error (AKD), 0.01699, compared with 0.01927 for UniRig and 0.01999 for MIA on the same meshes. It also achieved the highest mean PSNR and SSIM and the lowest mean LPIPS (Table 2, p. 9). Figure 16 (p. 31) overlays the target and rendered body keypoints to illustrate the differences in pose alignment.

The broader evaluation on the same 40 videos includes 12 animation pipelines (Appendix A.6, pp. 23–24; Table 5, p. 24). Three are the NVWA-mesh pipelines described above: native soft rigging, NVWA mesh + UniRig, and NVWA mesh + MIA. Three are animatable human reconstruction methods: IDOL, LHM, and LHM++. The remaining six pair each of three mesh generators (TRELLIS.2, Pixal3D, and PSHuman) with each of the two automatic rigging methods (UniRig and MIA). Across all 12 pipelines, NVWA with native soft rigging achieved the lowest LPIPS (0.06168) and AKD (0.01699). IDOL achieves the highest PSNR, and LHM++ achieves the highest SSIM. Fig. 10 (p. 23) illustrates a limitation of whole-frame PSNR and SSIM. Although NVWA scores lower than LHM++ and IDOL on both metrics for the frames shown, it reproduces the subject’s overall appearance more faithfully.

In a separate user study, 24 participants rated complete clips from four methods (NVWA, IDOL, LHM, and LHM++) on the same 10 cases. NVWA received the highest mean deformation-plausibility rating (4.52; Table 6, p. 24). These evaluations quantify downstream animation quality, with AKD measuring pose agreement in the image plane. We do not report 3D SMPL-X joint or vertex errors, and the generated free-topology meshes lack independently annotated per-vertex correspondences for measuring correspondence accuracy directly.

For a visual comparison, please see our anonymous 12-case animation comparison of NVWA, IDOL, LHM, and LHM++.

AuthorsOpenReview ↗

Reply to Reviewer vUYQ, Part 3

Reply to Reviewer vUYQ, Part 3

Question 3. Differences between NVWA and SyncHuman

Thank you for raising this comparison. Both methods use a native 3D generative prior to reconstruct a textured, free-topology clothed human from a single image. SyncHuman focuses on recovering the geometry and appearance of the person in the observed pose. NVWA additionally learns how the generated surface corresponds to a canonical human body, so that the same clothed mesh can be rigged and driven in new poses.

Network architecture and prediction targets. SyncHuman jointly generates RGB images and normal maps from four viewpoints together with a sparse 3D structure. Bidirectional attention aligns multiview image features with 3D voxel features, and the decoder uses this alignment to recover surface detail (SyncHuman, Secs. 3.1–3.2, Figs. 3–4). NVWA first generates a free-topology clothed mesh and uses its geometry to condition two prediction branches. The appearance branch predicts surface texture. The rigging branch adopts the same DiT architecture to predict our geometry-aligned soft rigging representation. This representation assigns each mesh vertex a semantic UV coordinate in the canonical SMPL-X atlas, providing the surface-to-body correspondences that guide SMPL-X fitting and skinning-weight transfer (Secs. 3.2–3.3, Fig. 2, p. 5).

Rigging and animation. SyncHuman outputs a static textured mesh and does not describe a native rigging module. Driving its output with a skeleton therefore requires a separate rigging step, for example with an external tool, to assign skinning weights to the mesh vertices. NVWA predicts our geometry-aligned soft rigging representation during generation and uses it to fit SMPL-X to the clothed mesh. We then interpolate the fitted body’s skinning weights in the canonical SMPL-X UV atlas and transfer them to the original mesh vertices. The skeleton can therefore drive the generated surface while preserving its vertices, faces, and connectivity (Appendix A.1, pp. 17–19). On identical NVWA meshes across the 40 evaluated videos, native rigging achieved a mean normalized 2D body-keypoint error of 0.01699, compared with 0.01927 for UniRig and 0.01999 for Make-It-Animatable (Table 2, p. 9).

Training data and the role of multiview generation. SyncHuman trains on existing human scans and their rendered views. The authors identify training scale as a limitation, noting that they fine-tuned their multiview generator on approximately 5,000 scans (SyncHuman, Secs. 4.1 and 5). We developed a scalable dataset construction pipeline that incorporates a refined reconstruction stage based on PSHuman. We used this pipeline to build NVWA-100K, a collection of approximately 100,000 explicit textured human meshes (Sec. 3.1, Fig. 1, p. 4). We retained 89,806 meshes for training and subsequently post-trained on a mixture of real scans and synthetic assets. Thus, multiview synthesis is part of dataset construction in NVWA. For a new input image, NVWA directly generates geometry, appearance, and soft rigging, whereas SyncHuman runs its synchronized 2D and 3D branches for each reconstruction.

Reviewer_J8oXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper presents NVWA, a native 3D human generation framework for reconstructing textured, animation-ready human avatars from a single image. The proposed approach builds upon TRELLIS.2 and introduces two main contributions.

First, the authors construct NVWA-100K, a dataset of 100,000 synthetic, textured 3D human meshes. The dataset is generated using a multi-view diffusion and explicit mesh reconstruction pipeline, providing native 3D supervision for adapting a general-purpose 3D generation model to clothed humans.

Second, the paper proposes geometry-aligned soft rigging, where the model generates free-topology human geometry together with a semantic UV correspondence field referenced to SMPL-X. The predicted correspondences are subsequently used for parametric body fitting and skinning-weight transfer, enabling animation without constraining the generated geometry to the SMPL-X topology.

The framework also incorporates real-scan post-training and an optional image-space A-pose editing step to improve geometric fidelity and animation robustness. Experiments on X-Humans and in-the-wild images demonstrate promising reconstruction quality and qualitative animation results.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper addresses a practically important problem: generating high-fidelity, animatable 3D humans with free-topology geometry that can represent clothing, hairstyles, and accessories beyond fixed body models.
  • NVWA-100K offers a potentially valuable large-scale collection of textured human meshes for native 3D generation, although its quality, diversity, and redistribution terms require clear documentation.
  • The method achieves competitive geometric and perceptual quality, including the lowest reported Chamfer Distance on X-Humans and strong semantic consistency on in-the-wild images.
  • The semantic correspondence representation provides a practical interface between native 3D generation and SMPL-X animation, supported by detailed procedures for correspondence assignment, parametric fitting, and skinning-weight transfer.
Critical weaknesses
  • The claimed benefit of native 3D supervision is weakened because NVWA-100K is constructed using PSHuman’s multi-view diffusion and reconstruction pipeline. The resulting meshes may inherit synthesis artifacts, so dataset quality, artifact frequency, and comparisons with real scans should be quantified.

  • The novelty of geometry-aligned soft rigging requires stronger justification. Although it supports free-topology meshes, it still relies on SMPL-X correspondence and linear blend skinning. Comparisons with Make-It-Animatable (CVPR 2025), UniRig (SIGGRAPH 2025), and HumanRig (CVPR 2025) are needed.

  • Rigging and animation quality are insufficiently validated. The paper lacks quantitative evaluation of correspondence accuracy, skinning weights, deformation under challenging poses, clothing-body penetration, and large articulations. Nearest-vertex UV transfer may also be unreliable for loose garments, hair, accessories, and overlapping regions.

  • The contributions of A-pose editing, native 3D training, real-scan post-training, and soft rigging are not adequately disentangled. Real-scan post-training improves Chamfer Distance only from 1.75 to 1.74, while external A-pose editing may independently affect topology, identity, and animation stability. Controlled ablations under matched inputs are needed.

  • Comparisons with recent animatable human reconstruction methods are incomplete. Beyond IDOL and LHM, the paper should discuss and, where possible, compare with LHM++, AniGS (CVPR 2025), SinGS (CVPR 2025), and Disco4D (CVPR 2025).

  • Evaluation, efficiency, and reproducibility require clarification. The evaluation covers only 100 X-Humans scans and 50 in-the-wild images, while PSHuman performs better on X-Humans ULIP-2. The paper should report end-to-end runtime and memory, training budgets, post-filtering dataset size, uncertainty estimates, and the number and licensing conditions of releasable NVWA-100K assets.

Questions for authors
  • Since NVWA-100K is constructed using PSHuman’s multi-view diffusion and reconstruction pipeline, how do the authors verify that view inconsistencies and hallucinations are not inherited by the meshes? Please report quality statistics, filtering criteria, and comparisons with real scans.

  • How does the semantic UV correspondence field differ from existing correspondence-transfer and automatic-rigging techniques? Comparisons with Make-It-Animatable, UniRig, or similar methods on identical meshes would better establish its novelty and effectiveness.

  • Could the authors quantitatively evaluate correspondence accuracy, skinning-weight quality, and deformation under unseen poses? Please also analyze challenging geometry such as skirts, long coats, loose garments, and accessories far from the SMPL-X surface.

  • Could controlled ablations separately assess A-pose editing, semantic UV generation, and real-scan post-training? Applying the same A-pose editing to LHM, IDOL, and other baselines would clarify its contribution to animation stability.

  • Please discuss and, where applicable, compare with LHM++, AniGS, SinGS, and Disco4D to clarify NVWA’s reconstruction, animation, and representation advantages.

  • What are the end-to-end runtime and peak memory for A-pose editing, native 3D generation, correspondence prediction, SMPL-X fitting, and skinning transfer? Please also report the post-filtering dataset size and how many of the 100,000 assets can legally be redistributed.

  • Could the authors provide confidence intervals for reconstruction metrics and user-study results and evaluate more identities? Please also analyze Chamfer Distance sensitivity to ICP registration, normalization, and visible-surface evaluation.

9 comments
AuthorsOpenReview ↗

Reply to Reviewer J8oX, Part 1

Reply to Reviewer J8oX, Part 1

We are grateful for your highly constructive comments, which have been very helpful. Our responses are detailed below:

Weakness 1 and Question 1. Quality of NVWA-100K and artifacts inherited from multiview generation

We thank the reviewer for raising this concern. We agree that inconsistencies in PSHuman’s generated views can carry over into the reconstructed mesh. For example, a mesh may look intact from the front, while a side view reveals a missing hand or a fragmented garment. We apologize that the original manuscript did not explain how we checked for these defects.

To limit their influence on training, we manually inspected multiview renderings of all 80,495 meshes in the random-pose subset and assigned each mesh an overall quality score from 1 to 10 based on visible geometry and appearance defects. We retained meshes scoring 5 or above. Even when the front view appeared intact, a missing hand or foot or severe fragmentation visible from a side view was sufficient for rejection.

The scoring guidelines are described in Appendix A.7 (pp. 24–25). Scores of 1–3 indicate severe defects, such as distorted faces or hands, missing extremities, or fragmented garments and accessories. Scores of 4–7 indicate a largely intact body with localized defects, such as mild limb deformation, hand color projected onto clothing, or streaks and gaps on the legs or shoulders. Scores of 8–10 indicate few visible defects across the inspected views; fused fingers alone could still fall in this range. Fig. 17 (p. 32) shows examples below the retention threshold.

Table 7 (p. 25) reports the following distribution for the random-pose subset:

Quality scoreNumber of meshesPercentage
s ≤ 34,8306.0%
3 < s < 59,65912.0%
s = 59,01511.2%
5 < s ≤ 734,29042.6%
s > 722,70128.2%
Total80,495100.0%

After filtering, 66,006 random-pose meshes (82.0%) met the threshold. Together with the 23,800 A-pose meshes retained separately, this yielded 89,806 meshes for training.

Weakness 2 and Question 2. How the semantic UV field differs from automatic-rigging

Thank you for suggesting a comparison on identical meshes. NVWA uses SMPL-X as its semantic reference and standard linear blend skinning (LBS) for animation. Our contribution lies in how the surface-to-body correspondence is obtained. As NVWA generates a free-topology human mesh, it also predicts, for each surface vertex, a location in the canonical SMPL-X UV atlas (Fig. 2, p. 5). A garment can therefore have its own geometry and topology while its vertices still carry a correspondence to the body. We use these correspondences first to fit SMPL-X to the generated avatar and then to transfer skinning weights by barycentric interpolation on the UV atlas (Appendix A.1, pp. 17–19). In the automatic-rigging comparison, UniRig and Make-It-Animatable (MIA) instead receive the completed mesh and produce a rig in a separate step.

Following your suggestion, we compared all three methods on the same NVWA-generated mesh for each of 40 in-the-wild videos, using the same target motion, camera, and evaluation frames. Table 2 (p. 9) reports the results.

Rigging method on NVWA meshesPSNR (dB) ↑SSIM ↑LPIPS ↓AKD ↓
NVWA+UniRig19.640.90150.066240.01927
NVWA+MIA19.530.90090.067080.01999
NVWA20.100.90510.061680.01699

With geometry held fixed, NVWA achieves the best mean result on all four metrics. AKD measures the normalized 2D distance between corresponding body keypoints in the rendered and reference frames. Fig. 16 (p. 31) overlays these keypoints so the pose differences can be inspected directly. These results support the effectiveness of the complete correspondence-guided rigging procedure on the tested motions. We also considered HumanRig, as you suggested. Its official repository links to the released dataset, but we could not find official inference code or pretrained weights for rigging our NVWA-generated meshes. We were therefore unable to evaluate HumanRig under the same protocol.

AuthorsOpenReview ↗

Reply to Reviewer J8oX, Part 2

Reply to Reviewer J8oX, Part 2

Weakness 3 and Question 3. Validation of rigging and animation quality for challenging geometry

We agree that the original submission provided insufficient quantitative evidence for rigging and animation quality. We have added a quantitative animation evaluation and analyzed how loose garments and accessories extending far from the SMPL-X surface deform under large body motions.

Quantitative evaluation. As discussed in W2/Q2, we compared NVWA’s native rigging, UniRig, and Make-It-Animatable (MIA) on identical NVWA meshes under the same target motions. Native rigging achieved the lowest AKD and the best mean PSNR, SSIM, and LPIPS of the three (Table 2, p. 9; Fig. 16, p. 31). These three pipelines are also included in our evaluation of 12 pipelines on 40 in-the-wild videos. The other nine comprise IDOL, LHM, and LHM++, plus six combinations obtained by pairing each of three mesh generators (TRELLIS.2, Pixal3D, and PSHuman) with UniRig or MIA. Across all 12 pipelines, NVWA achieved the lowest LPIPS (0.06168) and normalized 2D body-keypoint error (AKD, 0.01699) (Appendix A.6, p. 23; Table 5, p. 24). IDOL achieves the highest PSNR, and LHM++ achieves the highest SSIM. Fig. 10 (p. 23) illustrates a limitation of whole-frame PSNR and SSIM. Although NVWA scores lower than LHM++ and IDOL on both metrics for the frames shown, it reproduces the subject’s overall appearance more faithfully.

To assess deformation over complete motion sequences, we therefore also asked 24 participants to rate deformation plausibility after watching complete clips for the same 10 cases. NVWA received the highest mean rating (4.52), followed by LHM (3.85), IDOL (3.77), and LHM++ (3.72) (Table 6, p. 24).

For a visual comparison, please see our anonymous 12-case animation comparison of NVWA, IDOL, LHM, and LHM++.

These experiments measure the combined effect of correspondence estimation and skinning on the resulting animation. Our generated meshes lack independently annotated per-vertex correspondences and reference skinning weights, so we cannot report direct errors for either quantity. We have also not quantified cloth–body penetration rates.

Geometry far from the SMPL-X surface. We use nearest-vertex lookup in UV space during SMPL-X fitting. To transfer skinning weights, we interpolate weights within SMPL-X UV triangles and propagate them along the mesh from reliable anchors (Appendix A.1, pp. 17–19). Body correspondences can still be ambiguous where garments overlap or hang far from the body.The original manuscript mentioned potential animation failures in the conclusion without showing examples or explaining their causes. In the revision, Fig. 22 (p. 36) shows four such cases involving loose garments, accessories, and hair. As discussed in Appendix A.9 (p. 25), fixed skinning weights make these parts follow the skeleton, limiting their ability to drape or swing independently during motion.

AuthorsOpenReview ↗

Reply to Reviewer J8oX, Part 3

Reply to Reviewer J8oX, Part 3

Weakness 4 and Question 4. Contributions of A-pose editing, semantic UV generation, and real-scan post-training

Thank you for suggesting comparisons under matched inputs. We agree that editing the input pose can itself affect the generated topology and subsequent animation. We clarify below what each experiment establishes.

A-pose editing. Figure 9 (p. 21, Appendix A.3) compares NVWA with and without editing for the same person, target pose, and camera. In the second row, raising the leg stretches a surface connecting the two legs when the original image is used. With A-pose editing, the legs remain separate during the same motion. We also gave IDOL and LHM the same edited A-pose image used by NVWA. The IDOL-Apose and LHM-Apose results in Fig. 12-13 (p. 27-28), described in Appendix A.5 (p. 22), allow us to compare downstream reconstruction and animation under matched preprocessing. For a fair comparison, all 12 pipelines in the quantitative animation evaluation discussed in W3/Q3 used the same edited A-pose reference image for each video (Table 5, p. 24).

Real-scan post-training. Table 4 (p. 10) compares NVWA before and after this stage on X-Humans. We agree that the improvement in CD is small (1.75 to 1.74). The other four metrics also improve, including CLIP from 0.8613 to 0.8749 and Uni3D from 0.3116 to 0.3190. Figure 6 (p. 8) illustrates the changes in local geometry and texture detail. This stage uses 5K real scans together with 2K synthetic assets (Appendix A.2, p. 19), so the comparison measures the effect of adding this mixed-data post-training stage.

Semantic UV generation and soft rigging. The predicted UV field supports both SMPL-X fitting and skinning-weight transfer. Figure 7 (p. 9) compares fitting time and illustrates pose alignment on challenging inputs. As described in W2/Q2, we applied NVWA’s native soft rigging, UniRig, and Make-It-Animatable to the same NVWA-generated mesh for each video. With motion, camera, and evaluation frames fixed, native soft rigging achieved the best mean results on all four metrics across the 40 evaluated videos (Table 2, p. 9).

Weakness 5 and Question 5. Comparison with LHM++, AniGS, SinGS, and Disco4D

Thank you for pointing out that the original comparison with IDOL and LHM omitted more recent animatable-human methods. We added LHM++ to the 40-video quantitative evaluation described in our response to Weakness 3 and Question 3. Appendix A.6 (p. 23) gives the protocol, and Table 5 (p. 24) reports the following results on the same videos.

MethodPSNR (dB) ↑SSIM ↑LPIPS ↓AKD ↓
LHM++20.860.91820.063220.01928
NVWA20.100.90510.061680.01699

LHM++ achieves higher PSNR and SSIM, while NVWA achieves lower perceptual error (LPIPS) and normalized 2D body-keypoint error (AKD). Figure 10 (p. 23) illustrates how these metrics can differ on an individual frame: in the top row, LHM++ achieves higher PSNR and SSIM, while NVWA preserves the outline of the prosthetic lower leg more clearly. For complete clips, the 24 participants in Table 6 (p. 24) gave NVWA higher mean ratings than LHM++ on the same 10 cases for pose agreement (4.61 vs. 4.05), deformation plausibility (4.52 vs. 3.72), and overall visual quality (4.55 vs. 3.73).

Of the four methods you suggested, LHM++ was the one we could evaluate end to end under the same protocol. We also examined the official releases of AniGS, SinGS, and Disco4D to determine whether we could evaluate them on the same 40 videos, starting from the same single-image inputs.

AniGS. We could not locate the avatar reconstruction and animation code in the public release we tested, so we could not evaluate AniGS on the same input images as the other methods. Official repository

SinGS. The public release we tested did not include the original generation model or its pretrained weights, so we could not run the original pipeline from the same single-image inputs. Official repository

Disco4D. We could not locate the renderer for texture projection or the diffusion package specified for appearance completion in the release we tested, so we could not evaluate Disco4D on the same input images.Official repository

Reviewer_Jx9MOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents NVWA, a framework for generating textured, riggable 3D human avatars from a single image. The authors introduce NVWA-100K, a dataset of 100,000 explicit human meshes constructed through multiview generation, normal-guided reconstruction, and appearance projection. Built on TRELLIS.2, NVWA generates free-topology geometry alongside texture and a geometry-aligned semantic UV correspondence field. This soft rigging representation guides subsequent SMPL-X fitting and skinning-weight transfer, enabling animation while accommodating loose garments and accessories. Training combines human-domain adaptation with real-scan post-training, while image-space A-pose editing helps reduce topology artifacts during animation. Experiments evaluate reconstruction quality, fitting efficiency, and animation results against general-purpose and human-specific methods, supplemented by a user study and ablations.

Critical strengths
  1. Soft rigging combines free-topology human generation with skeletal rigging. By generating a geometry-aligned semantic UV correspondence field, the authors preserve complex shapes, including loose garments, hairstyles, and accessories, while providing correspondences for subsequent SMPL-X fitting and skinning-weight transfer. This well-motivated design balances geometric expressiveness and animation readiness, and its effectiveness is supported by the ablation study.

  2. The data construction pipeline provides large-scale explicit supervision for native 3D human generation. By combining multiview generation, normal-guided mesh reconstruction, and appearance projection, the authors construct NVWA-100K, comprising 100,000 textured human meshes spanning diverse poses, body shapes, and clothing styles. This pipeline converts image resources into assets that directly support training 3D generative models, offering a scalable approach to addressing the limited availability of real human scans.

  3. The work presents a comprehensive pipeline and extensive empirical evaluation. The authors carefully develop the data construction pipeline and conduct large-scale training with appropriate fine-tuning and real-scan post-training. The evaluation includes comparisons against a broad range of models on test sets from different domains, a user study, and convincing ablation studies. The proposed method achieves state-of-the-art performance across multiple evaluation metrics.

Critical weaknesses

As the authors acknowledge in the Conclusion, “Current limitations include dependence on external pose editing and linear blend skinning, which may remain unreliable under extreme occlusion or highly non-rigid garments..." The method still handles clothing through linear blend skinning rather than modeling it separately. The authors do not present failure cases or discuss the inherent limitations of an approach that does not model clothing separately.

Questions for authors
  1. Could the authors provide some comparisons with strong closed-source models, such as Rodin, VAST, and Meshy? I am curious whether these models could achieve comparable or even better results.

  2. Are there any quantitative comparisons for human animation?

  3. The weaknesses I am concerned about.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer Jx9M

Reply to Reviewer Jx9M

We thank you for your time and thoughtful feedback. We address your main concern and each of your questions below.

Critical Weakness and Question 3. Failure cases and limitations of clothing animation

We apologize for mentioning potential animation failures in the conclusion of the original manuscript without showing examples or explaining their causes. In the revised manuscript, we present four failure cases involving loose garments, accessories, and hair in Fig. 22 (p. 36) and analyze them in Appendix A.9 (p. 25).

The top row of Fig. 22 shows two garment failures. During a roll, a skirt train lifts with the lower body and stretches into a pointed sheet. As an arm moves, the free end of a sari stretches sideways and then forms a tight loop. The bottom row shows related failures for accessories and hair. A lantern tassel points sideways as the lantern tilts, and a ponytail stays close to the upper back as the torso bends forward.

These examples illustrate the limitation you identified. NVWA can reconstruct the shape of a garment extending beyond the body surface. During animation, however, it uses skinning weights transferred from the fitted SMPL-X body to the entire mesh. Under linear blend skinning (LBS), each vertex follows a fixed weighted combination of skeletal transformations. A free end can therefore hang plausibly at rest, as the sari does here, even though its subsequent response to gravity, contact, and inertia is not independently modeled. The same constraint applies to the tassel and ponytail.

Question 1. Comparison with Rodin, VAST, and Meshy

We compared NVWA with all three systems using the same four input images. Appendix A.8 (p. 25) describes the comparison, and Fig. 20 (p. 35) shows front and back views of each result. On these examples, the three commercial systems produce avatars comparable to NVWA in overall static appearance. For the person in the red robe, for instance, all four methods reproduce the lunge and wide sleeves, while the folds at the back differ.

In Appendix A.8 (p. 25), we also discuss the animation comparison among NVWA, Rodin, and Hunyuan3D on four characters driven by the same motions. Fig. 21 (p. 35) shows representative frames, and the supplementary video shows the corresponding animation sequences.

For a visual comparison, please see our anonymous animation comparison of NVWA, Rodin, Hunyuan3D.

Question 2. Quantitative comparisons for human animation

We evaluated 12 animation pipelines on 40 in-the-wild videos. We compared NVWA with three animatable human reconstruction methods (IDOL, LHM, and LHM++). To include pipelines that generate and rig meshes in separate stages, we paired each of three mesh generators (TRELLIS.2, Pixal3D, and PSHuman) with each of two automatic rigging methods (UniRig and Make-It-Animatable), yielding six additional baselines. We also applied both rigging methods to NVWA’s own meshes to compare them with native soft rigging while holding the generated geometry fixed.

Appendix A.6 (p. 23) describes the evaluation protocol, and Table 5 (p. 24) reports all four metrics. Across the 12 pipelines, NVWA achieves the lowest LPIPS (0.06168) and the lowest normalized 2D body-keypoint error (AKD, 0.01699). IDOL achieves the highest PSNR, and LHM++ achieves the highest SSIM. Fig. 10 (p. 23) illustrates a limitation of whole-frame PSNR and SSIM. Although NVWA scores lower than LHM++ and IDOL on both metrics for the frames shown, it reproduces the subject’s overall appearance more faithfully.

On the same NVWA meshes, native soft rigging achieves the best result among the three rigging methods on all four metrics.

Rigging method on NVWA meshesPSNR (dB) ↑SSIM ↑LPIPS ↓AKD ↓
NVWA+UniRig19.640.90150.066240.01927
NVWA+MIA19.530.90090.067080.01999
NVWA(soft rigging)20.100.90510.061680.01699

To assess deformation over complete motion sequences, we also asked 24 participants to rate clips from NVWA, IDOL, LHM, and LHM++ on the same 10 cases. As reported in Table 6 (p. 24), NVWA received the highest mean ratings for pose agreement (4.61), deformation plausibility (4.52), and overall visual quality (4.55).

For a visual comparison, please see our anonymous 12-case animation comparison of NVWA, IDOL, LHM, and LHM++.

good job, thanks

AuthorsOpenReview ↗

Reply to Reviewer Jx9M

Dear Reviewer Jx9M,

Thank you for your positive follow-up and for raising your score. We appreciate your recognition of our work and the constructive suggestions that helped us improve the paper.

Your comments prompted us to strengthen the quantitative animation evaluation and video-based user study, add comparisons with commercial systems, and expand the analysis of animation failures. In particular, the new examples involving loose garments, accessories, and hair, together with the discussion of linear blend skinning, make the capabilities and limitations of our animation pipeline clearer. We are grateful for your careful review and your contribution to these improvements.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.