acceptodds
Under review as a conference paper at ICLR 2027

TOPOS: High-Fidelity and Efficient Industry-Grade 3D Head Generation

Abstract

Industrial 3D head generation requires fixed reference topology for production-level rigging, skinning and animation, yet existing 3D generative models typically produce inconsistent mesh topologies which are incompatible with production pipelines. In this paper, we present TOPOS, a single image conditioned framework that jointly generates high-fidelity geometry and appearance under a unified industry-standard topology. To model heads under this unified topology, we proposed a novel variational autoencoder structure, termed TOPOS-VAE. Inspired by multi-modal large language models (MLLMs), TOPOS-VAE leverages the Perceiver Resampler to convert input pointclouds sampled from head meshes of diverse topologies into the target reference topology. Building upon TOPOS-VAE's structured latent space, we train a rectified flow transformer, TOPOS-DiT, to efficiently generate high-fidelity head meshes from a single image. We further present TOPOS-Texture, an end-to-end module that produces relightable and geometry-aligned UV texture maps from the same portrait image via fine-tuning a multi-modal image generative model. Extensive experiments demonstrate that TOPOS achieves state-of-the-art performance on 3D head generation, surpassing both classical face reconstruction methods and general 3D object generative models, highlighting its effectiveness for digital human creation. We will release our code and trained models to facilitate future research.

Primary area:
generative models
Keywords:
3D Head Generation, Geometry Modeling, Texture Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ifJwOpenReview ↗
Rating
3
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents TOPOS for single-image-conditioned 3D head generation with uniform topology and relightable texture maps. TOPOS-VAE maps unstructured point clouds to fixed-topology meshes, TOPOS-DiT generates its latent tokens conditioned on portrait images, and TOPOS-Texture fine-tunes Qwen-Image-Edit via LoRA to produce UV textures. The reported results show improved geometric accuracy and strong texture-generation performance compared with the evaluated methods.

Critical strengths

1.TOPOS directly generates head meshes with consistent MetaHuman topology, supporting downstream rigging and character animation. The demonstrated rig-driven facial expressions show the practical value of this design for production pipelines. 2.TOPOS-Texture produces relightable and geometry-consistent UV texture maps from a single portrait image. The comparisons under different lighting conditions show high-fidelity skin textures and accurate color reproduction.

Critical weaknesses

1.The benefit of using diverse-topology meshes is not directly validated. TOPOS-VAE is trained on DMD_M and evaluated on MetaHuman-topology meshes, without a separate cross-topology reconstruction test. It could be helpful to separately report the geometric fidelity of converting inputs with other source topologies to the target topology, and compare the results of TOPOS-DiT trained only on DMD_M with those of TOPOS-DiT trained on the full DD.

2.The texture evaluation does not fully isolate the contribution of TOPOS-Texture. Table 3 compares renderings of each method’s generated geometry and textures, although geometry quality already differs substantially in Table 2. Generating textures for the same mesh using different methods could help assess the contribution of TOPOS-Texture more directly.

3.TOPOS-Texture is trained on the full DD, but the paper does not explain whether these assets need to have UV layouts compatible with the target MetaHuman layout. If so, how such compatibility is established.

4.There may be a minor presentation issue with the numerical formatting in the tables. or example, NC is reported as 0.6546 for MeshAnythingV2 and 0.818 for TreeMeshGPT, using different numbers of significant figures.

Questions for authors

1.How many input points are used, and how are point features extracted or encoded from the point cloud? This is not fully discussed in the paper.

2.The paper claims that its graph convolution update is more effective than previous graph convolutions with MLP update functions, but does not report a direct comparison between these two formulations. Therefore it does not directly validate the advantage of this specific design. How much improvement does this update provide over graph convolutions with MLP update functions?

8 comments
AuthorsOpenReview ↗

Benefits of Diverse-Topology Meshes

We thank the reviewer for the valuable suggestion. We actually conducted cross-topology reconstruction experiments but did not include them in the original submission due to the nine-page limit. We have now included these results in Fig. 12 (page 10) of the revised paper. As shown in the figure, TOPOS-VAE can serve as a topology unifier, converting point clouds sampled from meshes with diverse topologies into the unified MetaHuman topology while preserving geometric details and facial identity. Furthermore, we use the full dataset DD instead of only DMD_M to increase the scale and diversity of the training data, thereby improving the generative capability of TOPOS-DiT.

AuthorsOpenReview ↗

Evaluation of TOPOS-Texture

We thank the reviewer for the suggestion. We agree that the rendered comparisons in Table 3 reflect the combined quality of geometry and textures. However, different texture generation methods, especially 3DMM-based approaches, are designed for specific mesh topologies and rely on their predefined UV parameterizations. Therefore, directly generating textures for the same mesh across these methods is not straightforward without additional topology conversion or texture transfer, which could introduce extra errors and compromise the fairness of the comparison.

To better isolate the contribution of TOPOS-Texture, we also provide ablation studies in Table 3 and Fig. 10, where the geometry generation pipeline remains unchanged. These results demonstrate the effectiveness of our texture-specific designs, including the removal of VL image tokens and the gradual resolution training strategy.

AuthorsOpenReview ↗

UV layout compatibility of the training dataset D with the MetaHuman layout

Yes, all training textures in DD share the same canonical MetaHuman UV layout. Although the meshes have different topologies, their textures were originally authored to follow this unified MetaHuman UV layout. Since TOPOS-Texture directly learns the mapping from 2D images to UV textures without explicit 3D geometry conditioning, its training is independent of mesh connectivity and only requires texture maps aligned with the target UV layout. In principle, meshes with other topologies can also be used by converting their textures to the MetaHuman UV layout offline.

Reviewer_1wcDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents TOPOS, a single-image 3D head generation framework that produces high-fidelity geometry and relightable textures with a consistent, industry-standard mesh topology. The framework consists of three components: First, TOPOS-VAE, which combines a Perceiver Resampler encoder and a GNN decoder to learn a structured latent representation of fixed-topology head meshes. Second TOPOS-DiT, which employs rectified flow to generate head geometry conditioned on a portrait image. Last, TOPOS-Texture, which fine-tunes Qwen-Image-Edit to generate geometry-aligned UV textures. Experiments on synthetic renderings and in-the-wild portraits demonstrate competitive reconstruction quality, geometric consistency, and texture fidelity compared with existing 3D face reconstruction and generation methods.

Critical strengths
  • High-quality 3D head generation with consistent topology. The proposed framework generates detailed 3D head meshes with a unified, industry-standard topology, making the outputs directly compatible with downstream applications such as rigging and facial animation.

  • Comprehensive experimental evaluation. The paper provides extensive quantitative and qualitative comparisons with both traditional face reconstruction methods and recent 3D generative models. The experiments cover geometry reconstruction, image-conditioned generation, and texture quality, with additional ablation studies supporting the proposed design choices.

Critical weaknesses
  • The proposed framework primarily integrates existing techniques, including the Perceiver Resampler, GNN decoder, rectified flow transformer, and Qwen-Image-Edit. Although the overall framework is effective, the technical novelty of the individual components beyond their integration appears somewhat limited.

  • The experiments rely largely on internal training and evaluation datasets. In particular, the in-the-wild test set contains only 159 images, without a complete description of their sources or selection criteria. The lack of publicly available evaluation data limits reproducibility and makes it difficult to assess generalization across diverse identities and challenging conditions. Furthermore, mesh quality is evaluated only on the test set, making it difficult to determine whether the method generalizes to different mesh topologies (e.g., FLAME and ICT-FaceKit). Additionally, the visual quality of the generated textures does not appear superior to that of the baselines, even though the quantitative results are substantially better.

Questions for authors

It would be much better to make image larger and present in the appendix for better visualization, for all comparisons.

5 comments
AuthorsOpenReview ↗

Technical Novelty

We thank the reviewer for raising this concern. While TOPOS builds on established components, including the Perceiver Resampler, GNNs, and rectified flow, we believe its novelty goes beyond combining these techniques.

First, TOPOS introduces a generative framework that directly produces high-fidelity 3D heads with a predefined industry-standard topology. Existing 3D mesh generative models generally produce meshes with instance-dependent connectivity, while traditional fixed-topology face models are often constrained by parametric shape spaces or require iterative template fitting. To our knowledge, TOPOS is the first single-image-conditioned flow-based framework that generates head meshes with the exact MetaHuman topology, preserving vertex correspondence and facial edge loops across different identities. This makes the generated meshes directly compatible with downstream rigging, skinning, and facial animation, as also highlighted by Reviewers 3a53 and ifJw.

Second, TOPOS-VAE is specifically designed to make such structured generation possible, rather than simply combining a Perceiver encoder with a GNN decoder. We construct a hierarchical graph from the reference mesh and explicitly align the Perceiver latent tokens with vertices of the coarsest graph. A hierarchical GNN then progressively reconstructs the full-resolution mesh through graph convolutions and predefined unpooling mappings. Together with geometry-aware supervision, this design preserves fine geometric details while maintaining the target connectivity. Importantly, the encoder accepts unstructured point clouds from meshes with different topologies, allowing them to be represented under the same reference topology (Fig. 12).

The effectiveness of these design choices is supported by our ablations in Table 1. Replacing our encoder with VecSet-FPS or VecSet-Learn reduces the F-score from 0.915 to 0.502 and 0.156, respectively, while removing geometric supervision reduces it to 0.636.

Beyond geometry, TOPOS-Texture learns a direct mapping from a single portrait to a canonical UV texture, avoiding geometry-dependent unprojection and multi-stage inpainting. Its training design further addresses UV alignment and relightability.

AuthorsOpenReview ↗

Dataset Availability and Generalization

We thank the reviewer for raising this concern. Since there are currently no publicly available training datasets specifically designed for this task, we had to construct our own datasets for training and evaluation. To improve reproducibility and facilitate future research, we commit to publicly releasing our training and evaluation dataset after acceptance of the paper. Regarding the 159 in-the-wild images, as already described in Sec. 4.5, our in-the-wild images consists of 159 uncurated portrait images (82 males and 77 females) collected from web sources, academic datasets and AI generation tools to cover a diverse range of identities and appearances.

AuthorsOpenReview ↗

Visual quality compared to baselines

We thank the reviewer for this comment. We agree that some baselines can produce visually similar results under the input viewpoint. However, we would like to emphasize that our goal is not merely to reproduce the input appearance, but to generate production-ready textures with a consistent industry-standard UV layout and relightable albedo.

First, many existing face reconstruction methods recover textures through color projection, warping, or UV completion, which can closely reproduce the input appearance but often retain the original illumination. On the other hand, general 3D generative models typically produce meshes with inconsistent topology and instance-dependent UV layouts, lacking the semantic correspondence required by industrial pipelines. In contrast, our method directly generates textures aligned with the predefined MetaHuman UV layout, enabling consistent texture editing and downstream material workflows.

Second, industrial applications require albedo textures rather than textures with baked-in illumination. While directly reproducing input colors may yield visually appealing results under the original lighting, the retained shadows and highlights can introduce artifacts under novel illumination. Our method instead learns to generate canonical albedo, supporting consistent relighting and further material editing.

These advantages are illustrated in Figs. 8 and 9 (please zoom in for a closer inspection) and further supported by the quantitative results in Table 3.

Reviewer_3a53OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel framework for 3D head mesh generation with UV texture. Unlike many existing method, it can produce meshes with fixed topology with geometry-aligned textures from inference portrait images, which is compatible for graphics pipeline and many downstream tasks such as animation. There are three main contributions:

  1. TOPOS-VAE, a reconstruction module from input point clouds to mesh output leverage Graph Neural Network.
  2. TOPOS-Texture, a texture generation pipeline fine-tuned from a pretrained multimodal image generative model.
  3. TOPOS, a high-fidelity 3D head mesh generation framework that simultaneously preserve facial identity and conform to an industry-standard topology.

The authors conduct comparisons to many existing methods regarding geometry and visual quality.

Critical strengths

I appreciate the proposed geometry generation framework. Using a GNN as the decoder provides a natural and effective way to reconstruct meshes with fixed topology, making the generated geometry well-suited for downstream applications. Also, the authors provide extensive comparisons with a wide range of existing methods, including both quantitative and qualitative evaluations, which help demonstrate the effectiveness of the proposed approach.

Critical weaknesses
  1. Lack of important technical details or clarifications, such as UV coordinate generation, rendering setting during evaluation and for visualization. (See questions below)

  2. Risk of texture inconsistency due to seams. (See questions below)

  3. Fairness of comparisons. (See questions below)

Questions for authors
  1. Consistency of parameterization. I was wondering how the proposed framework generates UV coordinates for each vertex. I understand that the training data are meshes with the same mesh topology, meaning that they have the same number of vertices and the same triangulation. Are the parameterizations also fixed? In another word, do the vertices with the same index have the same uv coordinates (if vertex-based texture mapping is used)? Or Is each mesh face assigned predefined UV coordinates shared across all generated meshes (if face-based texture mapping is used)? If so, the author need to point it out in the paper to avoid misunderstanding. If not, the authors need to clarify how the proposed framework generates UV coordinate and how to guarantee the consistency among vertex, UV coordinates and the color information. This would be important for the soundness of the proposed method.

  2. Potential texture inconsistency near the boundary. From the UV texture examples presented in the paper, it appears that the meshes are parameterized by introducing a seam extending from the lower boundary of the back of the head to the top of the head. This introduces additional boundaries into the mesh, where each vertex along the seam is assigned two different UV coordinates despite sharing the same 3D position (or additional actual vertices with the same 3D position and different UV coordinates. Depend on actual implementation). Of course this is a common strategy in computer graphics. However, it may introduce two potential issues when training a generative model to produce textures in this UV space. First, since the two sides of the seam are represented at different locations in the UV texture, it is difficult to guarantee that the generated colors remain consistent across the seam. This may lead to visible texture discontinuities when the generated textures are mapped back onto the 3D meshes and no related processing was mentioned in the paper. Second, the UV parameterization does not cover the entire ([0,1]^2) domain, leaving invalid regions in the texture maps represented as the black pixels in the upper portion of the teture images. When a generative model is trained on such texture maps, it may be difficult to completely prevent colors from these invalid regions from bleeding into nearby valid regions, potentially introducing dark artifacts along the texture boundaries. I was wondering if this happened in this work and if there's any used strategy to deal with seam consistency. I also suggest that authors could presents additional figures to show the quality of the seam regions.

  3. The shading methods used for rendering should be consistent across different methods to ensure fair visual comparisons and quantitative evaluations. For example, in Figure 5, the results of "Ours" and "GT" appear to be rendered with smooth shading, whereas those of "SparseFlex" and "TRELLIS.2" appear to use flat shading. Similarly, in Figure 7, "Ours" and "GT" appear to use smooth shading, while "FFHQ-UV", "Pixel3DMM" and "FLAME" appear to use flat shading. Since smooth shading can improve the perceived visual quality of rendered meshes, these inconsistencies may introduce bias into both the qualitative comparisons and quantitative evaluations. I suggest that the authors clarify the shading settings and ensure consistent rendering conditions across all methods.

  4. The implementation of the baseline methods requires further clarification. Methods such as DreamFace enhance the visual geometric details of low-resolution meshes by applying high-resolution normal maps, without modifying the underlying mesh geometry. However, based on the results shown in Figures 7 and 9, it appears that the authors may not have applied the normal maps as originally proposed in DreamFace. Given that one of the main objectives of this work is to achieve high-fidelity and efficient 3D head generation, it is important to ensure that the baseline methods are faithfully implemented.

  5. Although efficiency is emphasized as one of the main contributions of this work, the experimental evaluation of computational efficiency appears insufficient. I suggest reporting inference time peak GPU memory consumption under comparable settings.

  6. If applicable, I recommend the authors upload some generated meshes with texutres of different methods for better comparisons.

8 comments
AuthorsOpenReview ↗

Consistency of parameterization

We thank the reviewer for pointing this out. The UV parameterization is fixed across all generated meshes. Specifically, all generated meshes share exactly the same predefined UV coordinates and face-to-UV mapping as the MetaHuman reference mesh, including the mapping at UV seams. The UV coordinates are inherited from the reference mesh rather than generated by our framework, ensuring consistent correspondence between the geometry and UV textures. We will clarify this in the revised paper to avoid misunderstanding.

AuthorsOpenReview ↗

Seam consistency across UV boundaries and potential black-pixel bleeding from invalid regions

We thank the reviewer for this insightful, graphics-grounded observation. In our framework, both cross-seam discontinuity and black-pixel boundary bleeding are effectively mitigated through supervision from natively extended, seamless ground-truth assets (UV texture padding) rather than post-processing:

  1. Data-Driven Seam Consistency via Natively Extended Assets: Our training textures are not clipped strictly at the active UV island boundaries, nor do we use naive 2D pixel extrapolation. Instead, all 356 base texture maps come from artist-crafted, natively seamless 3D assets where the texture is continuously mapped and extended beyond the physical UV seam coordinates. Because the ground-truth data encodes cross-seam color and texture continuity (with extended margins around UV boundaries), the diffusion model learns to generate consistent skin/scalp tones across UV seams in a purely data-driven manner. Moreover, since the UV seam is primarily located at the back of the head, where skin appearance is relatively smooth, visible color discontinuities are less likely to occur.

  2. Extended Texture Buffer Against Black-Pixel Bleeding: This same extended texture region serves as a spatial buffer zone between the valid UV parameterization and the black invalid regions in the ([0,1]^2) domain. Because the valid UV boundary used by the 3D mesh is recessed inside the extended texture margin, it does not directly border the black background pixels. Consequently, potential dark-pixel bleeding introduced during the diffusion process or VAE downsampling/upsampling can be largely absorbed by the buffer zone, reducing its impact on the active UV coordinates sampled during 3D rendering.

To further address this concern, we provide additional back-view renderings of our generated 3D heads on page 10 (Fig. 11) of the revised paper. These results visually demonstrate that our method produces coherent textures across the UV seams without noticeable color discontinuities or artifacts.

AuthorsOpenReview ↗

Shading strategy

We thank the reviewer for pointing this out. In Fig. 5, the results of SparseFlex and TRELLIS.2 already use smooth shading, consistent with Ours and GT. However, the default shading strategy for some 3DMM-based face meshes in Blender is flat shading. To ensure fair visual comparisons, we have updated the renderings of all 3DMM-based methods in Figs. 7 and 9 to use smooth shading in the revised paper, thereby ensuring consistent shading settings across the compared methods. We also clarify that our quantitative geometry metrics are computed directly from the aligned 3D meshes and are therefore unaffected by the shading settings used for visualization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.