acceptodds
Under review as a conference paper at ICLR 2027

FaceFlow: Native Artist-like Mesh Generation with Face Flow Diffusion

Abstract

Diffusion-based artist-like mesh generation has recently emerged as an important direction in 3D content creation. However, the discrete nature of mesh topology poses a fundamental challenge for continuous diffusion models. In this paper, we further investigate the parallel generation of complete discrete mesh sequences. To this end, we propose FaceFlow, a diffusion-based framework for artist-like mesh generation. FaceFlow employs a Face VAE to encode mesh faces into compact face latents and decode them into complete mesh sequences in a single step. However, conventional sorting strategies cannot resolve the assignment ambiguity caused by the variable cardinality and non-uniform spatial distribution of face latents. Consequently, the face diffusion model cannot reliably determine which target face latent each noisy latent should represent. To address this issue, we introduce a Point Sink Flow model that transports a dense noisy point cloud into a sparse set of face centers. These centers specify the 3D locations of the corresponding face latents and are injected into the Face DiT as positional embeddings, explicitly establishing spatial correspondences between noisy and target latents. This spatial guidance enables the Face DiT to learn a more accurate transformation toward the face-latent distribution, thereby improving the quality of the generated mesh sequences. Experiments on Toys4K demonstrate that FaceFlow achieves state-of-the-art performance in both mesh reconstruction and generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Native Generation, artist-like mesh generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_J9EAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FaceFlow, a two-stage diffusion framework for generating artist-like triangular meshes directly in face space. The authors identify the assignment ambiguity arising from variable face counts, permutation invariance, and non-uniform spatial distributions as a key challenge in face-based diffusion models.

To address this issue, the proposed Point Sink Flow first generates a sparse set of face centers from a dense point representation using a VAE and a flow-matching DiT. These predicted centers provide both the number and spatial locations of face tokens. In the second stage, Face Flow incorporates the predicted centers as spatial positional embeddings to guide face-latent generation. A Face VAE then decodes the generated latents into discrete vertex coordinates, enabling parallel mesh generation.

The method is trained on approximately 230K meshes containing fewer than 5K faces and evaluated primarily on 210 Toys4K samples. The reported results demonstrate improved geometric reconstruction and point-conditioned generation performance over several existing approaches, including autoregressive and diffusion-based mesh generation methods.

Critical strengths

Strength 1: Effective spatial guidance for parallel face generation

The paper identifies a meaningful challenge in face-based diffusion: establishing correspondences between noisy face latents and target faces when their spatial distributions are non-uniform. The proposed spatial guidance provides a practical solution. Table 3 shows that incorporating face-center positional information substantially improves performance over ZYX sorting, reducing CD-L1 from 11.80 to 7.73 and HD from 8.24 to 2.74 in the reported scaled units. This suggests that explicit spatial information is useful for learning face-latent distributions and represents a meaningful improvement over purely sorting-based alignment.

Strength 2: Strong geometric reconstruction and generation results

The proposed Face VAE achieves competitive reconstruction performance, outperforming the evaluated baselines in Table 1. The complete FaceFlow framework also achieves the best reported CD, HD, and NC scores among the compared methods on Toys4K. These results demonstrate the potential of face-level latent diffusion for generating geometrically accurate meshes without autoregressive token-by-token decoding.

Critical weaknesses

W1. Questionable architectural design and insufficient validation

The necessity of introducing face centers as an intermediate representation is not sufficiently justified. For typical triangular meshes, the number of faces is approximately twice the number of vertices, potentially making face-level diffusion less computationally efficient than vertex-based approaches such as Nexus.

Furthermore, Point Sink VAE uses a fixed set of 10,240 sampled points to represent a variable number of face centers, raising concerns about its scalability across different mesh densities. The paper also lacks direct evaluation of face-center prediction accuracy, including Precision, Recall, F1-score, and face-count errors for both the VAE and DiT.

Another fundamental concern is topological continuity. The Face VAE independently predicts vertex coordinates for each face without explicitly enforcing shared-vertex consistency. Even with discrete coordinate prediction, adjacent faces may fail to connect, resulting in cracks, disconnected components, or invalid topology. The authors acknowledge this limitation, but provide insufficient quantitative evaluation. Unlike connectivity-aware representations such as SpaceMesh or Nexus, the proposed decoder offers no explicit guarantee of consistent face connectivity.

W2. Limited scalability and insufficient topological consistency evaluation

The experiments are restricted to meshes with fewer than 5K faces, primarily evaluated on only 210 Toys4K samples. In contrast, Nexus has already demonstrated generation of meshes with up to 20K faces. The paper provides no convincing evidence that FaceFlow can scale to higher face counts, nor does it report inference time or memory consumption across different mesh complexities.

More importantly, the practical significance of the proposed framework remains unclear. Two compelling applications would be retopology of dense meshes containing hundreds of thousands or millions of faces (e.g., meshes generated by Tripo) and image-conditioned artist-like mesh generation. However, the current experiments demonstrate neither capability, relying instead on point clouds sampled from ground-truth meshes with normals. Without stronger scalability experiments or more realistic application scenarios, the advantages of this relatively complex architecture over existing approaches remain unclear.

Questions for authors

Q1. Necessity of Face-center Representation.

What are the advantages of face-center-based generation over vertex-based approaches such as Nexus? Given that triangular meshes typically contain approximately twice as many faces as vertices, could the authors provide a direct comparison under matched conditions to justify the computational efficiency and necessity of this representation?

Q2. Point Sink Flow Accuracy and Fixed-length Representation.

Could the authors report Precision, Recall, F1-score, and face-count prediction errors for the generated face centers, evaluating the Point Sink VAE and DiT separately? How does the fixed 10,240-point representation affect performance across meshes with substantially different face counts and densities?

Q3. Topological Consistency.

Since the Face VAE independently predicts vertex coordinates for each face, how does the model ensure connectivity between adjacent faces? Could the authors report quantitative metrics on cracks, disconnected components, and non-manifold edges, and clarify whether additional post-processing is required?

Q4. Scalability and Practical Applications.

Could the authors demonstrate generation on denser meshes (e.g., 10K–20K faces or more), including inference time and GPU memory consumption? Additionally, could they demonstrate dense-mesh retopology (e.g., from meshes generated by Tripo) or image-conditioned generation to establish the practical value beyond the current GT point-cloud setting?

Q5. LOD and Face-count Controllability.

Can FaceFlow generate different levels of detail for the same input geometry without retraining? Specifically, how does changing the initial noise latent sequence length affect the number and spatial distribution of predicted face centers? Can users explicitly control the target face count?

2 comments

Thank you for the detailed rebuttal and additional experiments. I appreciate the efforts to address my concerns.

However, my main concerns regarding scalability and practical significance remain unresolved. The newly added image-conditioned results involve only very simple shapes, and the quality appears to degrade with denser conditioning inputs. Moreover, the 10K–20K face scalability experiments only evaluate the Face Flow stage with predefined face centers, rather than demonstrating successful end-to-end generation of complex meshes.

The authors also acknowledge that explicit LOD and face-count control are not currently supported. Additionally, the updated topology statistics still reveal noticeable connectivity issues.

Overall, while the additional experiments are helpful, they do not sufficiently demonstrate the generalization and practical scalability of the proposed architecture.

Thank you for your response and for acknowledging our rebuttal. We hope that the following response addresses your remaining concerns.

First, regarding the newly added image-conditioned results, as emphasized in the manuscript, there is a gap in vertex and normal distributions between the conditioning inputs used during training and those generated by Tripo3D during inference. This gap is particularly pronounced in the lack of smoothness of the normals. Specifically, watertight meshes typically contain 500K–2M faces, and the normals of point clouds sampled from their surfaces exhibit more irregular orientations. Without targeted engineering optimizations, such as introducing normal perturbations during training to simulate the normal distributions of watertight meshes, inference quality deteriorates as the number of conditioning points increases. This is because sampling more points leads to a larger gap in normal distributions. These observations also indicate that deploying FaceFlow in practical applications requires additional time and corresponding engineering optimizations. The current experiment is used to indicate that FaceFlow will be able to support image-based generation and retopology with engineering efforts. We will incorporate these optimizations as part of our future open-source efforts to enable such applications. Nevertheless, we believe that these engineering challenges do not diminish the value of FaceFlow in enabling parallel inference for mesh sequences.

Second, regarding the scalability tests at 10K–20K faces, it is difficult to guarantee that the Point Sink Flow stage generates 10K–20K face centers. Therefore, we o. Meanwhile, Point Sink Flow uses a DiT with a fixed latent sequence length of 1,024 and has the same number of parameters as the Face Flow stage. Therefore, if the Face Flow stage can run with 20K faces, Point Sink Flow with a sequence length of 1,024 must also be able to run. We therefore manually construct a specified number of face centers and use 81,920 conditioning points to evaluate the Face Flow stage. To address this concern, we additionally test the inference time and peak memory consumption of the full pipeline on a mesh containing 40K faces.

Input VInput FCondition PointsGenerated Face CentersPoint Sink TimeFace Generation TimeTotal Inference TimePeak Memory
56,92439,96281,9209,07627.23 s23.35 s50.58 s14.42 GiB
Reviewer_zKrWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FaceFlow generates artist-like triangle meshes in two stages. Point Sink Flow predicts discrete face centers and deduplicates them into a variable-size set. These centers provide positional guidance to Face DiT, which generates face latents in parallel. A Face VAE then decodes each face into three vertex coordinates. Experiments on Toys4K show improvements in mesh reconstruction and point-cloud-conditioned generation.

Critical strengths
  • Point Sink Flow is well motivated. Predicting and deduplicating discrete face centers provides an adaptive candidate face count, spatial locations, and local face density information. This reduces the correspondence ambiguity between noise tokens and target faces left by sequence sorting.
  • Positional augmentation targets shifted, duplicated, and off-surface face centers. It addresses the gap between ground-truth centers used in training and predicted centers used at inference. The ablation in Table 6 supports its effectiveness.
  • The method achieves the best reported geometric metrics among the compared methods on Toys4K for both reconstruction and point-cloud-conditioned generation. Some ablations are convincing. In particular, spatial conditioning outperforms a sorting baseline with the same DiT architecture and ground-truth face count (Table 3), showing benefits beyond predicting the face count.
Critical weaknesses
  • The handling of vertex order within each triangle is unclear. Face VAE concatenates the positional encodings of three vertices, but the paper does not explain how it handles orientation-preserving cyclic permutations or reversed winding. Face-center conditioning does not resolve this ambiguity. Please clarify whether canonical ordering, a permutation-aware loss, or another strategy is used, and how face orientation is kept consistent.
  • The benefits of direct face generation over vertex-then-connectivity prediction, as used in Nexus and LATO, lack quantitative analysis. The results support the full system but do not establish that the face representation itself explains the gains. Please discuss how directly generating faces compares with generating vertices and their connectivity in terms of the difficulty of learning topology.
  • Shared vertices rely on identical discrete coordinate predictions across faces. Predictions in different quantization bins can cause disconnections or cracks. The fixed 512-level quantization may also merge distinct vertices or face centers, collapsing or omitting fine structures. Quantitative statistics on disconnections, holes, and non-manifold artifacts would be useful.
  • Validation on complex structures remains limited. Although the appendix includes more examples, evaluation focuses on Toys4K and meshes with few faces and simple topology. Results on detailed meshes with higher face counts would help assess robustness on difficult structures.
  • The paper motivates its approach by the slow inference of autoregressive methods and emphasizes efficient parallel generation, but reports no end-to-end runtime, peak memory, or scaling with face count.
  • Images are mentioned as possible conditioning inputs, but FaceFlow's main experiments use point clouds. Image-conditioned generation is not evaluated.
Questions for authors
  1. What evidence supports the benefit of directly generating faces over predicting vertices then connectivity?
  2. Can you quantify the topology artifacts, such as holes or non-manifold connections, caused by relying solely on matching quantized coordinates to merge vertices across faces?
  3. Can you provide additional results on detailed and complex meshes, and report end-to-end runtime and peak memory as face count increases?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

[Face orientation]

As in autoregressive mesh-generation methods, each face is represented as a sequence of nine tokens corresponding to the 3D coordinates of its three vertices. The ordering of these vertices inherently encodes the face winding and thus its orientation. Therefore, we do not need to additionally estimate vertex normals to determine face orientation.

[Benefits of direct face generation over vertex-then-connectivity prediction]

Mesh-sequence generation is a longstanding task in artist-like mesh generation, and most prior methods directly generate faces using autoregressive models. The primary objective of this work is to resolve the assignment ambiguity that arises in parallel mesh-sequence generation because of the variable cardinality and non-uniform spatial distribution of face latents. We therefore do not position FaceFlow as universally superior to approaches that first predict vertices and then infer their connectivity. Instead, we view the two paradigms as having complementary strengths and limitations. On the one hand, vertex-first approaches may require predicting an O(N2)O(N^2) edge-connectivity matrix, which introduces substantial computational and memory costs during supervision. Their training efficiency may therefore deteriorate as the number of vertices increases. FaceFlow avoids explicitly constructing such a connectivity matrix. On the other hand, directly generating faces provides a more natural representation of mesh primitives and achieves higher reconstruction accuracy in our experiments (see Table 1). Recovering faces from predicted edge connectivity is inherently ambiguous: identifying faces solely through closed triangular loops can introduce duplicate or overlapping faces. Direct face generation instead delegates face construction to the network and avoids additional post-processing for resolving this ambiguity. Nevertheless, as discussed in our limitations, the current FaceFlow model does not explicitly guarantee consistent vertex sharing across adjacent faces. We consider this an open problem that we aim to address in future iterations.

[Quantitative statistics on disconnections, holes, and non-Manifold artifacts]

We provide the corresponding quantitative statistics in the revised manuscript. Please refer to Table 7 for details.

[More complex cases]

Because we restrict the training data to meshes with fewer than 5,000 faces, the current version cannot reliably generate substantially more complex, higher-face-count examples. The face-count distribution of our training data is provided below. Approximately 70% of the training meshes contain fewer than 3,000 faces; therefore, the complexity of Toys4K is well aligned with FaceFlow’s current generation capacity. In future iterations, we plan to incorporate more complex training data to improve the model’s ability to generate meshes with higher geometric and topological complexity.

Face Num< 500< 1000< 2000< 3000
Ratio12.20%25.93%49.68%69.04%

[Efficiency]

We have added a detailed efficiency comparison to the revised manuscript, including inference times across meshes with different face counts. As shown in Table 9, FaceFlow consistently outperforms all autoregressive methods in inference efficiency.

[Image Conditioning]

Regarding image conditioning, please refer to our common response. We have included image-conditioned generation results in the revised manuscript (see Figure 10), demonstrating FaceFlow’s ability to support this task. In our future open-source release, we will further improve the implementation and engineering optimizations for image-based and other conditioning modalities.

Dear Reviewer zKrW

As the discussion deadline is coming to a close, we would greatly appreciate your feedback on whether these additional clarifications and experiments have addressed your remaining concerns. If any questions remain, we would be happy to provide further clarification.

Thank you very much for your time and consideration.

Thanks to the authors for the detailed response. I lean toward accepting this paper and maintain my recommendation

Dear Reviewer zKrW:

Thank you for your follow-up and continued support. We are glad that these revisions address your concerns and remain committed to supporting other conditioning inputs in our release code.

Reviewer_QP8uOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FaceFlow is a two-stage flow-matching pipeline for generating artist-like triangle meshes from a point cloud. Stage 1 (Point Sink Flow) is a latent flow model whose VAE encodes a dense surface point cloud (10,240 points) and decodes, for every point, a quantized face-center coordinate; deduplicating these predictions yields a sparse, variable-size set of face centers. Stage 2 (Face Flow) generates one face latent per center and injects the centers as positional embeddings (absolute PE or 3D RoPE), so that each noisy latent is tied to a spatial location and the assignment ambiguity of unordered, variable-cardinality face sets is removed; a Face VAE then decodes every latent into three quantized vertices. Both DiTs have ~960M parameters and are trained on ~230K Objaverse/Objaverse-XL meshes with fewer than 5K faces. On 210 Toys4K meshes, the Face VAE improves reconstruction (Table 1) and the full pipeline improves point-conditioned generation (Table 2) over AR and diffusion baselines. Ablations cover latent alignment (zyx sorting vs. PE vs. RoPE, Table 3), discrete vs. continuous centers (Table 4), continuous supervision (Table 5) and positional augmentation (Table 6); Appendix A.2.4 adds boundary-edge, triangle-area and VLM-based topology scores (Table 7).

Critical strengths
  • An explicit spatial scaffold for parallel face generation, backed by a controlled ablation. Predicting a variable-size set of face centers before generating the faces gives the Face DiT both the sequence length and the spatial layout of the target set, while keeping decoding fully parallel. The ablation in Table 3 is the right one: the zyx-sorting baseline (the MeshCraft strategy) uses the same DiT and is even given the ground-truth face count at inference, yet RoPE conditioning reduces CD-L1 from 11.8 to 7.73 and HD from 8.24 to 2.74 (Table 3). This isolates the value of spatial correspondence from simply knowing how many faces to generate, which is the paper's central claim.

  • Consistent geometric gains and a robustness study aimed at the actual train/test gap. FaceFlow is best on all four metrics in both reconstruction (Table 1) and point-conditioned generation (Table 2), with HD and CD margins that are not marginal. More importantly, the Face DiT is trained with ground-truth centers but receives predicted centers at test time; Table 6 shows that augmenting the centers according to the three observed error patterns moves HD from 6.60 to 2.74, and Fig. 8 shows 515 predicted centers collapsing to 359 faces. Section A.3 then identifies center prediction, not face generation, as the current bottleneck (Fig. 10). This is the kind of analysis that makes a two-stage pipeline credible.

Critical weaknesses
  • The "artist-like topology" claim rests on far weaker evidence than the geometric claim. The main metrics (CD, HD, NC on 5,000 sampled points) are blind to connectivity, so the geometric gains do not by themselves support the title claim. The complementary evaluation in A.2.4 / Table 7 is underspecified: no formula is given for the boundary or triangle-area scores or for their normalization to (0, 100); the VLM protocol (prompt, views, rendering, number of samples, scoring rule, variance across runs) is not described; and the visual score comes from a single VLM with no calibration against artist or expert judgment. Boundary-edge counts and area histograms also do not measure what "artist-like" means in practice (edge flow, manifoldness, shared-vertex consistency). Table 7 additionally mixes conditioning modalities (LATO2 from rendered images, TriFlow from SDFs), so those rows are not comparable with the point-conditioned ones. Finally, the topology scores are not reported for the ablation variants of Table 3, so the attribution in Sec. 4.3 of "cleaner connectivity and more regular edge flow" to the Point Sink Flow stage is asserted rather than tested.

  • The efficiency claim is never measured. The introduction motivates the method by the slow token-by-token inference of AR methods, and the third contribution states that "high-quality mesh sequences can be generated efficiently". But the paper reports no sampling-step count, latency, throughput or memory for any stage. The pipeline consists of two ~960M-parameter DiTs, a jointly fine-tuned PoseMaster encoder over 81,920 conditioning points, and two VAE decoders (Sec. 4.1, A.1). Parallel decoding within each stage does not imply an end-to-end speed advantage over BPT, FastMesh or MeshFlow, and the overhead of the additional center-generation stage cannot be assessed.

  • Fig. 1 caption: "We present FaceFlow, diffusion-based framework" → "a diffusion-based framework".

  • Fig. 6 labels the second variant "Accurate PE" while Table 3 and the text call it "Absolute PE"; Fig. 7 uses "Continue" for "Continuous".

  • Sec. 3.2.1: "Further discussion of these two prediction heads are provided" → "is provided".

  • Table 1 reports CD-L2 at ×10⁴ and NC as a fraction, while Tables 2–4 and 6 use ×10³ and percent; the asterisk on "LATO 2" in Table 1 is never explained; "LATO 2 / LATO2 / Lato2" and "Triflow / TriFlow" are used interchangeably.

  • Eq. (7) uses ∣⋅∣22|\cdot|_2^2 where Eq. (4) uses ∥⋅∥22\|\cdot\|_2^2.

  • Fig. 1 suggests 50 sampling steps (STEP 0/20/40/50); if that is the setting, state it in Sec. 4.1.

Questions for authors
  • Topology evaluation (clarification; affects Weakness 1). Please give the exact formulas and aggregation for the boundary and triangle-area scores, the full VLM protocol (prompt, renders per mesh, scoring, variance over repeated queries) and, if possible, any agreement check against human raters. Can you report shared-vertex / non-manifold-edge statistics, and the Table 7 metrics for the three variants in Table 3? Do the conclusions hold when Table 7 is restricted to point-conditioned methods?

  • Efficiency (clarification; affects Weakness 2). For both flow stages together, including conditioning and VAE decoding: number of sampling steps, end-to-end latency and peak memory on a named GPU at a representative face count, compared with BPT, FastMesh and MeshFlow at similar output sizes.

  • Augmentation and deduplication (reproducibility). A.2.2 describes three perturbation types but not their distributions, magnitudes or rates. During training, when several augmented centers map to the same face, how are they paired with target latents? At inference, how are redundant faces removed, and how much of the 515→359 reduction in Fig. 8 comes from the model producing identical faces versus post-hoc filtering (A.1 mentions duplicate-face removal only for preprocessing)? Given the HD change from 6.60 to 2.74, these details matter for reproducibility.

  • Conditioning density (scope). The conditioning cloud is 81,920 oriented points sampled from the target mesh, which is close to the surface itself. How does the Point Sink Flow behave with sparser or noisier clouds, or with the image conditioning mentioned in Sec. 3.2.2 but not evaluated?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

[Topology Evaluation]

In the revised topology evaluation, we report only descriptive mesh statistics and do not introduce any additional derived metrics or formulas. For each of the 210 test cases, we render the input mesh, the meshes generated by each method, and the ground-truth mesh from eight viewpoints evenly spaced at 45∘45^\circ intervals. The evaluation primarily considers structural consistency with the input and whether the generated edge flow conforms to common artist-authored modeling practices. We use the following prompt: Please evaluate the results generated by each method according to two criteria. For each test case, inspect all eight views of the input mesh surface renderings and the corresponding generated mesh to assess how accurately the generated result preserves the input structure and geometric details. Then, evaluate the topological quality using the eight wireframe views of the generated mesh. Specifically, determine whether the edge flow follows common artist-authored low-poly modeling practices, whether the triangulation conforms to the underlying geometry whenever possible, and whether the topology is complete and free of visible holes. Finally, average the scores across all test cases and report the overall score for each method.

[Mixed and Missing Comparisons]

We additionally report mesh-structure statistics and VLM-based topology-evaluation results for the FaceFlow variants presented in Table 3. Compared with the sorting-only variant, FaceFlow with face-center conditioning produces fewer boundary and non-manifold edges. Together with its improved VLM-based topology score and Chamfer distance, these results indicate that the Point Sink stage enables the model to generate more coherent edge flow.

[Efficiency]

We provide a detailed efficiency comparison in Table 9. The results show that FaceFlow achieves substantially faster inference than autoregressive methods.

[Augmentation and Deduplication]

We have added the detailed augmentation strategies and their application probabilities to Section A.2.2. To establish the correspondence between augmented centers and mesh faces, we record the index of the face from which each point is sampled. We then use this index to retrieve the corresponding face and pass it through the Face VAE to obtain its latent representation. Consequently, the input to the Face VAE may contain repeated faces when multiple sampled points are associated with the same face.

[Redundant-Face Removal]

During inference, we only use the vertex-merging functionality provided by trimesh to remove redundant faces. In the current version, we do not apply any additional procedure to explicitly resolve shared-vertex connectivity issues.

[Conditioning Density and Scope]

Regarding the scope of conditioning inputs, please refer to our common response. We use 81,920 input points following the default configuration of PoseMaster’s VAE, which applies a 20×20\times downsampling factor and produces point features of size 4096×10244096 \times 1024. The VAE also supports point clouds of different sizes, producing feature sequences whose lengths are reduced by the same factor of 20. We additionally report results on the Toys4K test set under different point-sampling densities and noise levels. Performance is evaluated using the Chamfer distance between the generated and ground-truth meshes. The noise-robustness experiments are conducted using 10,240 conditioning points. The results demonstrate that FaceFlow is robust to variations in the number of conditioning points, consistent with the findings reported in Section A.2.5.

Points5,12010,24020,48040,96081,920
CD (↓\downarrow)0.008510.007680.008000.007040.00773
Noise0.0010.0030.0100.030
CD (↓\downarrow)0.008290.010500.016260.02575

Dear Reviewer QP8u:

As the discussion deadline is coming to a close, we would greatly appreciate your feedback on whether these additional clarifications and experiments have addressed your remaining concerns. If any questions remain, we would be happy to provide further clarification.

Thank you very much for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Response by Authors

Overall response:

We sincerely thank all three reviewers for their careful evaluation and constructive feedback. To address the shared concerns regarding topology evaluation and the scope of FaceFlow’s conditioning inputs, particularly image-based inputs and dense meshes, we provide a consolidated response together with supporting experiments.

[Topology Evaluation]

Regarding topology evaluation, we first emphasize that it is difficult to define a single quantitative metric that directly and comprehensively reflects topological quality. Valid mesh topology is inherently non-unique, and different users may prefer different edge-flow conventions. Following the reviewers’ suggestions, we have therefore reorganized and expanded our evaluation of topology quality. First, we report a range of mesh-structure statistics, including the numbers of connected components, boundary edges, and non-manifold edges. Cracks and holes are difficult to count directly because they lack unambiguous operational definitions; nevertheless, the number of boundary edges can serve as an indirect reference. We stress, however, that none of these statistics alone reliably indicates overall topological quality. Second, we provide complementary perceptual evaluations through GPT-based scoring and a human-preference user study (see Section A.2.4 of the revised manuscript). These evaluations capture aspects of edge-flow quality and modeling conventions that are difficult to express using handcrafted metrics. Given the inherent ambiguity of topology assessment, we still recommend referring to the qualitative visualizations for the most direct evaluation of the generated topology.

[Conditioning Scope]

First, the primary objective of this work is to enable parallel inference for mesh sequences. We identify a key challenge in applying diffusion models to mesh-sequence generation: the assignment ambiguity caused by the variable cardinality and non-uniform spatial distribution of face latents. FaceFlow is designed specifically to address this challenge. Following the standard experimental setting adopted by most autoregressive mesh-generation methods, we evaluate our model using dense point clouds sampled from the original meshes. Under this setting, FaceFlow demonstrates clear advantages and verifies the feasibility of parallelizing mesh-sequence generation with diffusion models.

Second, we agree that image-conditioned generation and retopology are more closely aligned with practical applications. We therefore include qualitative results for an image-to-dense-mesh-to-artist-like-mesh pipeline in the revised manuscript. Importantly, the current FaceFlow model has not been specifically optimized for this setting. Nevertheless, these preliminary results suggest that FaceFlow could achieve substantially better performance if equipped with task-specific training and adaptation.

Additionally, image-conditioned generation and high-resolution mesh retopology typically require substantial additional resources and engineering effort. We therefore adopt the standard experimental setting most commonly used by existing methods for our main comparisons. For image-conditioned artist-like mesh generation, practical challenges include canonicalizing the orientations of the training meshes and aligning the poses of the input images with their ground-truth meshes. For retopology, a key challenge is mitigating the distribution gap between training and inference (for detailed gaps, please see Sec. A.2.5). For example, the appropriate number of conditioning points must be carefully determined: as shown in Figure 10, sampling more points enlarges the distribution gap and degrades generation quality. Conversely, when only a small number of points are sampled, additional cues such as normal maps may be needed to preserve the high-frequency details of the dense input mesh. We plan to investigate these practical engineering challenges further in our future open-source release.

AuthorsOpenReview ↗

Rebuttal by Authors

[Scalability and Practical Applications]

We have added an efficiency comparison across different face counts to the revised manuscript. Because all meshes in our training set contain fewer than 5K faces, we select meshes of varying complexity from Toys4K, whose meshes contain fewer than 4K faces, for the input-based evaluation. To evaluate scalability beyond this range, we additionally construct face-center sequences of predefined lengths and measure the inference latency and peak memory consumption of the Face Flow stage alone, excluding Point Sink Flow, at higher face counts. We emphasize that these measurements are obtained without dedicated engineering optimization. Nevertheless, the results demonstrate that FaceFlow can scale to meshes with substantially larger face counts.

FacesTimeMemory
5,00010.36 s7.66 GiB
10,00022.44 s8.11 GiB
15,00038.86 s8.53 GiB
20,00059.51 s8.98 GiB

We also present an image-conditioned retopology application in Figure 10. However, substantial distribution gaps exist between the training and inference meshes, particularly in their vertex and normal distributions. Consequently, results obtained without task-specific adaptation or optimization may not fully reflect FaceFlow’s potential for this application. Please refer to our common response for further discussion of image and dense-mesh conditioning.

[LOD and Face-Count Controllability]

In the current version, the number of faces in the generated mesh is inferred by the model from the conditioning point cloud rather than specified explicitly. As shown by the statistics in Table 7, the predicted face counts do not differ substantially from those of the ground-truth meshes. In future work, we plan to introduce explicit face-count control for practical applications, particularly under image and dense-mesh conditioning, thereby enabling controllable levels of detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.