- 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
- 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.
- 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
- 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.
- 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
- 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.
- Do the authors have quantitative measurements of soft-rigging accuracy or animation quality, such as SMPL-X joint/vertex error, correspondence accuracy.
- I think this paper is too similar to Synchuman; please explain in detail the differences between these two papers.
5 comments
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++.
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.
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:
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.