acceptodds
Under review as a conference paper at ICLR 2027

UVRelief: A Surface Codec for Avatar-as-Image Generation

Abstract

Generating human avatars as an Avatar-as-Image representation has become an active and rapidly growing research direction, where we map an avatar onto a shared UV domain over a parametric body model and turn 3D generation into a 2D problem. Realizing this potential, however, hinges on how to obtain a compact latent space for Avatar-as-Image generation. We present UVRelief, a surface codec for Avatar-as-Image generation that encodes a human as a UV-native relative shape: given a textured mesh and its fitted SMPL-X model, UVRelief bakes surface geometry and appearance into Relief Maps, where each texel stores a relief displacement relative to the SMPL-X surface and a texture signal. Built on these maps, a VAE encodes them into a compact UV-structured latent space and decodes texel-aligned 3D Gaussian attributes under rendering supervision, and a flow-matching prior is trained in this latent space for image-conditioned 3D avatar generation. Extensive experiments show that UVRelief improves reconstruction fidelity and achieves competitive generation quality compared with existing multi-view and latent avatar generation methods.

Primary area:
generative models
Keywords:
3D avatar generation, UV-space representation, 3D Gaussian Splatting, flow matching

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_Z3xPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UVRelief, a surface codec for representing human avatars in a compact latent space. The proposed method represents clothed human geometry as surface displacements relative to a fitted SMPL-X model in the UV space. A VAE is then trained to encode this representation into a compact latent space, followed by an image-conditioned flow-matching model for avatar generation. Experiments on reconstruction and generation demonstrate the effectiveness of the proposed representation.

Critical strengths
  1. Simple and intuitive human-specific representation. The proposed method effectively leverages a human prior by representing complex clothed human geometry as displacements relative to the SMPL-X surface. Compared with directly modeling absolute 3D positions, UVRelief provides a more structured and compact representation.

  2. Solid performance on both reconstruction and generation. The paper evaluates the proposed representation on both codec-level reconstruction and downstream image-conditioned avatar generation. The proposed codec achieves better reconstruction performance than TRELLIS and its human-specific finetuned variant, while the full generation pipeline also achieves strong quantitative performance, particularly in terms of PSNR and SSIM.

Critical weaknesses
  1. The motivation and scalability of the proposed representation are not fully convincing. One of the main motivations of this work is that existing image-based codecs implicitly need to solve 3D reconstruction when learning the latent representation. However, UVRelief instead assumes access to complete textured 3D meshes and fitted SMPL-X models during training. In particular, part of the training data from MVHumanNet++ is first converted into 3D meshes using a multi-view reconstruction method, followed by SMPL-X fitting with OmniFit. Therefore, the proposed approach appears to move a significant part of the reconstruction difficulty from representation learning to data preprocessing. While this design indeed makes the latent representation easier to learn, it is unclear whether this trade-off is favorable in terms of data scalability.

  2. The experiments do not sufficiently demonstrate the advantage of UVRelief itself for generative modeling. The paper argues that the proposed relative representation leads to a more learnable latent space. However, in the controlled encoder-side ablation, the reconstruction performance of the position-map baseline is already very close to that of UVRelief (23.35 vs. 23.46 PSNR). The much larger improvement is observed over the SIGMAN-like cross-attention baseline, which needs to obtain its latent representation from multi-view images, whereas UVRelief is constructed from complete 3D geometry. Therefore, this comparison does not fully isolate the benefit of the proposed relative representation. More importantly, a controlled quantitative generation comparison between UVRelief and the position-map representation is missing. Such an experiment would provide more direct evidence that UVRelief indeed results in a substantially more learnable latent space for downstream generative modeling.

Questions for authors
  1. Could the authors provide more details about how TRELLIS was fine-tuned on the human-specific dataset, including the training setup and supervision?

  2. Could the authors provide a quantitative generation comparison between UVRelief and the position-map baseline?

  3. How sensitive is UVRelief to errors in SMPL-X fitting?

3 comments

Response to Reviewer Z3xP

We thank the reviewer for the constructive report and for pointing at the exact experiment that was missing. We have run it, together with an out-of-distribution evaluation and a fit-sensitivity study.

W1. Reconstruction-based preprocessing: does it just move the difficulty?

The reviewer is right that our pipeline needs an explicit surface before learning the latent space, and that this is a design choice with a cost. We would argue it is a favourable trade-off, for a reason that can be measured.

An explicit surface supplies geometric supervision: the latent is tied to a plausible 3D structure rather than to view-dependent appearance. A codec trained directly from multi-view images without such a reconstruction has no constraint on the unobserved geometry, so it tends to fit the input views and may produce shapes that are only consistent from those views. Our approach moves part of the difficulty into preprocessing, but the resulting latent is well defined in 3D, and the generation prior only has to model a bounded residual over the SMPL-X manifold.

On scalability, we note that the reconstruction step is exactly what recent human-specific datasets provide at scale, and that the cost is one-off per asset rather than per training run. We will state this trade-off explicitly in the limitations, including the option of using the million-scale reconstructed bodies released with SIGMAN.

W2 / Q2. The controlled generation comparison

Agreed, and now provided. One flow-matching prior per encoding, identical architecture, data, supervision and steps; only the encoder input changes:

SettingMetricOursposition map
X-Humans VAEPSNR / SSIM / LPIPS / FID28.12 / 0.9770 / 0.0354 / 13.3027.72 / 0.9753 / 0.0364 / 13.44
X-Humans GenerationPSNR / SSIM / LPIPS / FID27.08 / 0.9657 / 0.0442 / 18.3325.99 / 0.9607 / 0.0465 / 18.57
4D-DRESS VAEPSNR / SSIM / LPIPS / FID25.65 / 0.9647 / 0.0453 / 31.1423.14 / 0.9511 / 0.0522 / 39.22
4D-DRESS GenerationPSNR / SSIM / LPIPS / FID20.32 / 0.9222 / 0.0721 / 69.4620.04 / 0.9227 / 0.0738 / 71.11

The reviewer's observation that "the position map is already close" is exactly what we see on X-Humans (27.08 vs 25.99 for generation, 28.12 vs 27.72 for reconstruction). The new 4D-DRESS set shows where the two diverge: +2.51 dB for reconstruction and better FID for generation, on frames whose surface leaves the volume covered by the training bodies (0.82% of the surface for held-out X-Humans bodies versus 9.19% for the selected 4D-DRESS frames). This is the direct evidence that the relative encoding gives a more learnable latent space once the query leaves the training distribution.

We also agree that the cross-attention comparison confounds two factors (it must recover geometry from images), which is why we do not use it as the main evidence for the representation claim; it is reported as a different design point.

Q1. How TRELLIS was fine-tuned

We fine-tune only the second-stage VAE of TRELLIS, i.e. the sparse-voxel decoder that produces voxel-attached 3DGS from aggregated multi-view DINOv2 features. The first stage is left untouched: the voxel grid that conditions the decoder is computed directly from the GT geometry, exactly as in the public release. The fine-tuning uses the same 8k human meshes, the same renderer and the same image supervision as our model. The baseline therefore receives perfect geometry at its input, which is a favourable setting for it, and any remaining gap is due to the latent representation rather than to the input. We will state the optimisation details (steps, batch size, learning rate) in the appendix so that the comparison is reproducible.

Q3. Sensitivity to SMPL-X fitting errors

We perturb the fit at three levels, rebuild the body consistently, re-bake the maps and re-run reconstruction on the 100 X-Humans subjects:

levelshape sigmapose sigmaPSNRSSIMLPIPS
0----28.120.97700.0354
10.31.5 deg26.220.97190.0435
20.63.0 deg23.470.96040.0557
31.05.0 deg20.340.94340.0736

The codec indeed inherits the accuracy of the fit — a body-relative representation cannot do otherwise — and we now quantify this instead of only stating it in the limitations. Note that both the relief map and the position map receive the same perturbed body, so the comparison between encodings is unaffected by this dependence.

Ethics

We will add a short data statement: all scans come from public research datasets with consent for research use; our method produces avatars of the scanned subjects and could be misused for impersonation, so we will discuss the need for consent when applying it to new identities.

Changes promised for the revision

  1. Sec. 4.1: baseline grouping, SMPL-X provenance, TRELLIS fine-tuning details.
  2. Sec. 4.5: the controlled generation comparison.
  3. New section: out-of-distribution evaluation with voxel statistics.
  4. Limitations: preprocessing trade-off, data scale, sensitivity to fit errors.
  5. Ethics statement.

It's a little late for response, but it addresses my several concerns. Thanks for authors' effort on these extra experiments, and I decide to raise my rating to 3. I hope this method can be applied to larger dataset.

Reviewer_qGW5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces a UV-native relative representation (UVRelief) to address the challenge of latent representation design in "Avatar-as-Image" generation tasks. This representation encodes relative displacement and texture information at each texel based on a fitted SMPL-X model, thereby representing 3D human geometry as regular Relief Maps. Subsequently, the authors train a VAE to compress these Relief Maps into a compact latent code that preserves UV structure, which is then decoded—under rendering supervision—into 3D Gaussian attributes aligned with the texels. Leveraging the latent space learned by the VAE, the authors further train an image-conditioned flow-matching generative model for single-image avatar generation. They validate their key contributions in representation learning and downstream generation through VAE reconstruction and image-to-avatar generation tasks, respectively; additionally, they conduct an encoder-side ablation study within the VAE reconstruction setting, comparing UVRelief against other methods such as absolute 3D position maps and cross-attention-based encoding. Experimental results demonstrate that this method enhances reconstruction fidelity and achieves competitive generation performance compared to the baseline methods.

Critical strengths

The problem formulation and the proposed insight regarding representation are valuable. The paper specifically addresses the core question of how 3D human geometry should be encoded into a latent space suitable for generation, proposing the construction of a "UV-native relative representation" that leverages SMPL-X priors. By representing human geometry as small residuals relative to SMPL-X, the method achieves properties of being pose-agnostic and spatially regular. A VAE is trained to learn the latent space using this representation, and a subsequent image-conditioned flow-matching model learns to generate within this space; the overall methodological design is sound. This insight offers valuable reference for "Avatar-as-Image" latent generation.

Critical weaknesses

The core claim has not been sufficiently or directly validated in the context of downstream generation. The authors assert that the proposed "UVRelief" relative representation is easier for downstream generators to learn, whereas position maps impose a heavier learning burden; however, the key validation experiments were conducted within a VAE reconstruction setting. Moreover, the reconstruction metrics for UVRelief and position maps in Table 3 are actually very similar. While qualitative experiments for downstream tasks were performed using "in-the-wild" data, these do not fully demonstrate the advantages of the proposed representation for such tasks.

Questions for authors
  1. Please provide additional controlled quantitative experiments for the core claim regarding downstream generation tasks. The authors are requested to quantitatively compare the performance of UVRelief and position maps on downstream tasks while keeping the generative prior architecture, decoder architecture, dataset, supervision, and training setup constant, varying only the encoder-side representation.
  2. Please clarify the source of the posed SMPL-X data used for quantitative testing on X-Humans. The paper does not specify how the posed SMPL-X data used by the decoder for the X-Humans dataset evaluation was obtained—specifically, whether it was estimated from the conditioning image or derived via ground-truth (GT) scan fitting. Given that this parameter provides a strong human body geometry prior for the proposed method, this detail is crucial for assessing the fairness of the comparison in the single-image generation task.
  3. Could the authors explain why TRELLIS.2 was not included in the image-to-avatar generation benchmark comparison? Since the authors have already included TRELLIS as a baseline, and given that TRELLIS.2 also employs a general 3D latent generation paradigm—offering a meaningful point of comparison with the latent representation used in this paper—the inclusion of supplementary experiments or a justification for its exclusion would influence the assessment of the current baseline coverage.
1 comment

Response to Reviewer qGW5

We thank the reviewer for the positive assessment and for identifying the missing experiment precisely. We have now run it, and we also clarify the two protocol questions.

Summary of new evidence

  1. Controlled generation comparison (relief vs. position map, identical prior/decoder/data/supervision/schedule).
  2. A second test set from 4D-DRESS, deliberately out of distribution, with a voxel-coverage analysis of why it is OOD.
  3. SMPL-X perturbation study (three levels) quantifying sensitivity to fitting errors.
  4. Qualitative figures on both sets.

Q1 / W1. Controlled quantitative comparison for downstream generation

This is the central request and we agree it was missing. We train one flow-matching prior per encoding with the same architecture, data, supervision and number of steps; only the encoder input differs. Results:

SettingMetricOursposition map
X-Humans VAEPSNR / SSIM / LPIPS / FID28.12 / 0.9770 / 0.0354 / 13.3027.72 / 0.9753 / 0.0364 / 13.44
X-Humans GenerationPSNR / SSIM / LPIPS / FID27.08 / 0.9657 / 0.0442 / 18.3325.99 / 0.9607 / 0.0465 / 18.57
4D-DRESS VAEPSNR / SSIM / LPIPS / FID25.65 / 0.9647 / 0.0453 / 31.1423.14 / 0.9511 / 0.0522 / 39.22
4D-DRESS GenerationPSNR / SSIM / LPIPS / FID20.32 / 0.9222 / 0.0721 / 69.4620.04 / 0.9227 / 0.0738 / 71.11

Two conclusions. First, generation with the relative encoding is better on both benchmarks (+1.09 dB on X-Humans, +0.28 dB on 4D-DRESS, with lower FID in both cases), so the claim is no longer supported only by reconstruction. Second, the margin grows on the out-of-distribution set (+2.51 dB for reconstruction versus +0.40 dB on X-Humans), which is consistent with the explanation we give in the paper: an absolute position must be memorised and has no support outside the training volume, whereas a relief displacement is a bounded residual w.r.t. SMPL-X.

Q2. Provenance of the posed SMPL-X for X-Humans

Both in training and in evaluation the SMPL-X is obtained by fitting the model to the scan with OmniFit, not by estimating it from the conditioning image. All compared methods in Table 2 receive the same body, so no method is advantaged by a different body prior. This is now stated explicitly in Sec. 4.1. We also quantify what happens when the fit is imperfect (next paragraph), since a reader may reasonably ask how much of our result depends on the body prior.

Sensitivity to fit errors (context for the fairness question)

As a sanity check for the fairness of the body prior: perturbing the fit at a mild level (shape sigma 0.3, pose sigma 1.5 degrees) already reduces reconstruction from 28.12 to 26.22 dB, so the codec clearly inherits the quality of the fit. Both variants receive exactly the same perturbed body, which is why this dependence does not affect the comparison between encodings.

Q3. Why TRELLIS.2 is not in the generation benchmark

TRELLIS.2 is a newer version of the same family. We use TRELLIS, the more widely adopted release, to represent this category; the family also contains several Hunyuan variants, and it is not feasible to include every member. Including one representative per family is what we do consistently for the other groups (SyncHuman for multi-view generation, IDOL for single-image reconstruction). We will add one sentence making this explicit so that the coverage of the benchmark is not ambiguous.

Changes promised for the revision

  1. Sec. 4.1: SMPL-X provenance and the baseline grouping rationale.
  2. Sec. 4.5: the controlled generation table above and its discussion.
  3. New section: out-of-distribution evaluation (voxel analysis + figures).
  4. Limitations: sensitivity numbers and the data-scale discussion.
Reviewer_5HBmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UVRelief, a dedicated human codec for image-conditioned 3D avatar generation. Given a textured human mesh and a fitted SMPL-X model, the method constructs UV maps containing scalar displacements along the surface normal direction and RGB appearance. A VAE compresses these maps into a UV-structured latent space and decodes them into UV-aligned 3D Gaussian attributes under rendering supervision. A conditional flow matching model is then trained on these latents for single-image generation. Its main contribution lies in using mesh-derived, body-relative geometric information as the codec input. The validation includes VAE reconstruction and image-conditioned generation on the X-Humans dataset, qualitative comparisons on in-the-wild images, and ablation studies on the encoding strategy and decoder design.

Critical strengths

Rather than directly predicting UV attributes from images, this paper first learns a human-specific human VAE based on geometry and appearance information encoded from a textured mesh, and then trains an image-conditioned generative prior. Representing this input relative to SMPL-X provides a concrete way to incorporate anatomical structure into representation learning, without requiring the codec's encoder to recover geometry from rendered views. This is, in my view, the main positive contribution of the paper: the design is conceptually interesting and offers a valuable direction for building structured latent spaces for avatar generation.

Critical weaknesses

Limited flexibility of the geometric encoding. A single height value along each template normal direction is overly restrictive for complex, loose, unconventional, or layered clothing. It is currently unclear how multiple intersections are handled, for example when a single ray hits both the back of the head and a ponytail, or both the thigh skin and a skirt. Which surface is retained? How is the other surface handled? The encoding also depends on accurate SMPL-X fitting: fitting errors shift the ray origins and directions, yet their impact is not quantified. These issues cast doubt on the reliability of the codec when dealing with challenging clothing.

The claimed generative advantage is insufficiently demonstrated. Table 1 fails to disentangle the benefit of the height (relief) encoding from the effect of the SMPL-X prior, even with a TRELLIS model fine-tuned on human data. In Table 3, relative encoding and absolute position encoding yield very close reconstruction results (PSNR of 23.46 dB and 23.35 dB, respectively), as also shown in Fig. 6. More importantly, the core claim of the paper is that relative encoding makes generation easier, yet the quantitative ablation measures VAE reconstruction instead. Supporting this core claim requires a controlled comparison of generation quality.

Generalization to in-the-wild images is unconvincing. In the first example of Fig. 5, the red underwear clearly visible in the input image is missing from the output. The loss of such a salient observed detail (rather than an unobserved region) makes the claimed high fidelity on real photographs difficult to believe.

Lack of comparison with recent baselines. Comparisons with related human reconstruction methods such as LHM and FISHuman are missing. Such comparisons would help establish the practical value of the method relative to other recent task-specific solutions.

Questions for authors

How does the codec handle rays with multiple intersections, such as those passing through both the scalp and the ponytail surface, or through both skin and clothing? Which intersection is retained? How are rays with no intersection handled?

In the X-Humans generation experiments, are the SMPL-X parameters estimated from the input image or fitted to the ground-truth scans? Do the baseline methods receive equivalent body information? How sensitive are the results to fitting errors?

Could you report generation results for relative encoding versus absolute position encoding under the same denoiser architecture, the same data, and the same training budget? Since the VAE reconstruction scores of the two are very close, a quantitative comparison of generation quality is essential to evaluate the claimed advantage of relative encoding.

Could you provide broader evaluations on loose or layered clothing and on real photographs, ideally comparing with LHM or FISHuman under identical input conditions? This would help determine whether failure cases such as the missing red garment in Fig. 5 are isolated incidents or reflect a broader limitation.

2 comments

We thank the reviewer for the careful reading and for recognising the core idea (body-relative geometry as the codec input). Below we address each concern with new experiments; the corresponding text changes are marked and will be applied in the revision.

Summary of new evidence added for the rebuttal

  1. Controlled generation comparison between the relief map and the position map with everything else fixed (same denoiser, data, schedule, steps).
  2. A second, out-of-distribution test set built from 4D-DRESS (20 frames with large motion and loose garments), with a voxel-coverage analysis showing why it is out of distribution.
  3. A sensitivity study that perturbs the SMPL-X fit at three levels and measures the reconstruction drop.
  4. Qualitative comparisons on both sets (two figures).

W1 / Q1. Height-field limits and multiple intersections

We agree that a single displacement per texel cannot represent layered garments, and we now state this assumption explicitly. The current codec is a single-layer height field: for a ray with several intersections (scalp/ponytail, thigh/skirt) we keep the first hit along +n inside a |l| <= 0.30 m band, and texels whose ray leaves the band are marked empty and decoded as zero displacement (i.e. the point falls back to the SMPL-X surface). The cast is implemented by rasterising the UV grid, taking each texel's barycentric point on the SMPL-X surface as the ray origin and its interpolated normal as the direction, querying a BVH of the scan, and rejecting hits outside the band or whose surface normal deviates by more than 90 degrees.

Extending to multi-layer clothing is a natural next step: store several hits per texel (multi-hit ray cast) and let the decoder predict a small set of depths per texel instead of one. We will add this discussion to Sec. 3.1 and to the limitations.

W2 / Q3. The generative advantage was not demonstrated

This was the most important gap, and we have now run the missing experiment. We train one flow-matching prior per encoding with identical architecture, data, augmentation, schedule and number of steps; only the encoder input changes. Results (100 X-Humans scans; the 20 selected 4D-DRESS frames of the new section):

SettingMetricOursposition map
X-Humans VAEPSNR / SSIM / LPIPS / FID28.12 / 0.9770 / 0.0354 / 13.3027.72 / 0.9753 / 0.0364 / 13.44
X-Humans GenPSNR / SSIM / LPIPS / FID27.08 / 0.9657 / 0.0442 / 18.3325.99 / 0.9607 / 0.0465 / 18.57
4D-DRESS VAEPSNR / SSIM / LPIPS / FID25.65 / 0.9647 / 0.0453 / 31.1423.14 / 0.9511 / 0.0522 / 39.22
4D-DRESS GenPSNR / SSIM / LPIPS / FID20.32 / 0.9222 / 0.0721 / 69.4620.04 / 0.9227 / 0.0738 / 71.11

The picture is consistent: on the in-distribution benchmark the two encodings are close (this is what Table 3 shows), but on the out-of-distribution set the reconstruction gap widens from +0.40 to +2.51 dB, while generation, which is +1.09 dB ahead on X-Humans, stays ahead by +0.28 dB. In other words the advantage of the relative encoding is largest where the absolute position of a texel loses training support, which is the mechanism we claim. The smaller generation margin is consistent with what we write in the limitations: the flow prior is trained on about 8k avatars and has no 2D generative prior, so it saturates before the representation does.

W3 / Q4. In-the-wild example with the missing red garment

We agree: this is a genuine failure in detail preservation, and we should not have presented it as a success. Our prior is trained on about 8k reconstructed avatars, and we do not use any 2D generative prior. Multi-view methods such as SyncHuman inherit a 2D generative model trained on millions of images, which is what produces sharp textures, while 3D-native methods such as TRELLIS are trained on millions of objects. Two remedies follow, and we will state them as future work: (i) train the codec on the million-scale reconstructed bodies released with SIGMAN; (ii) because our representation is a 2D map, initialise the prior from a 2D generative model and fine-tune it instead of training from scratch. We will also soften the claim of "high fidelity on real photographs" accordingly.

On the broader evaluation the reviewer asks for: the new 20-frame set is our evaluation on loose garments, and on it the two encodings separate clearly (see W2). Layered clothing is not covered by the current codec, since it stores a single displacement per texel (W1); multi-hit casting is the natural extension.

W4. Missing baselines (LHM, FISHuman)

We group baselines by problem setting and include one representative per group, because the groups share the same failure modes and adding more members does not change the conclusion:

  • Multi-view generation (FISHuman, SyncHuman, PSHuman): we use SyncHuman, which is open-source and reports the strongest numbers. FISHuman is not released, so it cannot be run under identical inputs. The shared weakness of this group is cross-view inconsistency and visible seams.
  • Single-image reconstruction (LHM, IDOL): we use IDOL, which also aligns an image to a UV latent. The shared weakness is the lack of a generative prior, so unobserved regions are blurry.

We are happy to add a paragraph making this grouping explicit in Sec. 4.1 and to cite LHM/FISHuman there.

Q2. SMPL-X source and sensitivity to fitting errors

Source: for both training and evaluation, the SMPL-X is fitted to the scan with OmniFit, not estimated from the conditioning image; every compared method receives exactly the same body, so the comparison is fair for methods that also consume a body prior.

Sensitivity: we perturb the fit at three levels, rebuilding the body consistently, re-baking the maps and re-running reconstruction on the 100 X-Humans subjects:

levelshape sigma (beta)pose sigmaPSNRSSIMLPIPS
0 (original fit)----28.120.97700.0354
1 (mild)0.31.5 deg26.220.97190.0435
2 (moderate)0.63.0 deg23.470.96040.0557
3 (strong)1.05.0 deg20.340.94340.0736

The degradation is graceful and monotone, and level 1 already corresponds to a perturbation that is visible on the body outline, which confirms that the codec inherits the accuracy of the fit, as stated in the limitations, and quantifies it for the first time. This also shows that the comparison with the position map is not an artefact of a particularly good fit: both variants receive the same body.

Changes promised for the revision

  1. Sec. 3.1: single-layer assumption, multi-intersection rule, implementation.
  2. Sec. 4.1: explicit baseline grouping and SMPL-X provenance.
  3. Sec. 4.5: the controlled generation comparison above.
  4. New section: out-of-distribution evaluation with the voxel analysis.
  5. Limitations: detail fidelity / data scale / 2D prior, and the SMPL-X sensitivity numbers.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Note on the revised manuscript

The revision is now in the paper, marked in blue: a new section Additional Evidence: Out-of-Distribution Evaluation (p. 10) that reports a controlled comparison of the relief map and the position map for both reconstruction and generation, on X-Humans and on a 20-frame 4D-DRESS set (Table 4, Figs. 7–8), with the selection and evaluation protocol in Appendix E.5. Our point-by-point responses are posted in the three review threads below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.