Pith. sign in

REVIEW 4 major objections 5 minor 37 references

StrandDesigner: Towards Practical Strand Generation with Sketch Guidance

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sketch conditioning plus learned coarse-to-fine upsampling lets users generate and edit realistic 3D hair strands, outperforming text- and image-guided methods.

desk verdict A plausible, well-engineered sketch-conditioned strand generator whose headline quantitative claims outrun the evidence: no error bars, a 44-sample split, and an unvalidated latent-space arithmetic step. read the letter →

arxiv 2508.01650 v1 pith:FCKYNPJB submitted 2025-08-03 cs.CV

Na Zhang , Moran Li , Chengming Xu , Han Feng , Xiaobin Hu , Jiangning Zhang , Weijian Cao , Chengjie Wang
show 1 more author
Yanwei Fu
This is my paper · ORCID
classification cs.CV
keywords sketch-basedgeneration3Dhairstrandsmulti-scaleupsamplingnext-scalepredictiondiffusiontransformerlatentVAEDINOv2adaptationcoarse-to-finemodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes StrandDesigner, the first sketch-based model for generating 3D hair strands. It argues that text prompts are too ambiguous and reference photos too inconvenient, while sketches let users specify shape and flow directly. To make sketch conditioning work, the paper replaces fixed upsampling with a learnable multi-scale strategy: hair strands are encoded into residual latent maps at three scales, and a transformer with diffusion heads generates each finer scale conditioned on the previous ones and on the sketch. It also adapts a pretrained visual backbone with per-scale learnable tokens so the model handles sparse and dense sketches alike. On USC-HairSalon and CT2Hair, the method reports higher geometric and semantic fidelity than text-guided and image-guided baselines.

What carries the argument

The load-bearing mechanism is next-scale prediction in strand latent space. A pre-trained strand VAE maps 3D strands to a 64-D-per-strand hair map; max pooling with kernel $2^{K-k}$ produces coarser maps, and Equation (1) defines each finer map as the residual $\hat H_k=H_k-\mathrm{tile}(H_{k-1})$ to remove redundancy. Each scale is encoded by its own latent VAE, and a MAR-style transformer with diffusion heads models $p(\hat h_k\mid \hat h_1,\dots,\hat h_{k-1})$, denoising each scale from the previous ones. Conditioning comes from DINOv2 features adapted by learnable scale-specific tokens, with local patch tokens concatenated to strand tokens and the global class token added directly.

What would settle it

Take a trained StrandDesigner, decode each scale's residual latent map, and check whether the decoded strands match the geometric difference between consecutive ground-truth hair maps; if they do not, or if swapping Equation (1) for a simple concatenation of scales leaves the Table 2 metrics unchanged, the claimed multi-scale mechanism is not responsible for the gains.

Watch

Extended reading notes

Core claim

The central claim is that sketch-conditioned 3D strand generation works by decomposing a full hairstyle into a sequence of latent maps of increasing spatial resolution, first guide strands, then residual maps that add detail, and autoregressively generating each map with diffusion heads conditioned on the sketch. The paper's central mechanism is the residual decomposition $\hat H_1=H_1$ and $\hat H_k=H_k-\mathrm{tile}(H_{k-1})$ for $k\ge 2$, with the coarsest map obtained by max-pooling the hair map, which turns fixed interpolation into a learned prediction task. Alongside it, a multi-scale adaptive conditioning mechanism aligns adapted sketch features with each scale and combines global and local tokens. The paper reports that this combination outperforms existing text- and image-conditional methods on PC-IoU, Chamfer distance, Hausdorff distance, CLIP score, and LPIPS, and improves unconditional distribution quality over the text-guided baseline.

Load-bearing premise

The method assumes that the learned latent space for strand maps is arithmetic in the right way: max-pooling and subtracting latent maps produces coarser or residual maps that still correspond to genuine strand geometry, so that the multi-scale targets are meaningful.

Editorial extensions

If this is right

  • Users can specify hairstyles by drawing binary strokes rather than writing prompts or hunting for reference photos, which gives finer control over shape, length, and flow.
  • Edits to a sketch, such as lengthening hair or adding curls, propagate to the generated 3D strands, making the generation loop iterative and interactive.
  • The model accepts both sparse and dense sketches, so novices and professional designers can use the same pipeline.
  • Because every scale is decodable, users can inspect the guide-strand stage before full generation and correct the condition early.
  • The generated hairstyle distribution is closer to real data, with higher fidelity and diversity than the text-guided baseline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same residual multi-scale decomposition could transfer to other fiber-like 3D structures such as fur, grass, or bristles, where guide-strand control is natural; the paper does not test this.
  • Feeding sketches with stylized hand-drawn strokes at training time, or adding a sketch-style transfer module, could narrow the reported gap between hand-drawn inputs and rendered training sketches.
  • A controlled study varying where the learnable tokens are injected in DINOv2, currently the last third of layers, would show whether deeper conditioning improves fine-detail adherence on dense sketches.
  • Extending the conditioning to multiple sketch views would likely resolve occluded regions and parting ambiguity, which the paper lists as future work but does not demonstrate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes StrandDesigner, a sketch-conditioned 3D hair strand generation model. It decomposes a strand-level latent hair map into K scales via max pooling and residual subtraction (Eq. 1), trains per-scale latent VAEs, and generates scales autoregressively with a MAR-style transformer with diffusion heads, conditioned on DINOv2 features adapted through learnable scale-specific tokens and a dual-level global/local conditioning mechanism. The method is evaluated on USC-HairSalon, CT2Hair, and 24 self-collected hairstyles against HAAR, a reimplemented Sketch+HAAR, and HairStep, using geometric metrics (PC-IoU, CD, Hausdorff) and semantic metrics (CLIP, LPIPS), plus unconditional generation metrics against HAAR (MMD-CD, COV-CD, 1-NNA).

Significance. The problem is well motivated, and the paper is the first to propose a strand-level, sketch-conditioned generation framework; the learnable multi-scale upsampling and adaptive conditioning are sensible design choices, and the qualitative results are visually compelling. The authors also promise code release, which would help reproducibility. If the quantitative claims survive a more rigorous evaluation protocol, this would be a useful practical contribution to hair modeling in computer graphics and vision. At present, however, the evidence for the central 'significantly outperforms' claim is weakened by the small test set and the absence of uncertainty quantification, and a core latent-space decomposition assumption is not validated.

major comments (4)
  1. [§4.2, Table 2] The central claim that StrandDesigner significantly outperforms HAAR, Sketch+HAAR, and HairStep is not supported by the reported statistics. All numbers in Table 2 are single-run point estimates on a 44-hairstyle test split, with no error bars, number of seeds, or significance tests. Because HAAR and Sketch+HAAR use stochastic diffusion generation and HairStep relies on SD3-based image translation, the observed differences (e.g., CD 0.80 vs 1.06) could be run-to-run variation. Please report mean and standard deviation over at least three seeds for all methods and apply a paired significance test (e.g., Wilcoxon signed-rank) on the fixed test set, or explicitly weaken the 'significantly outperforms' wording.
  2. [§4.2, HairStep comparison] The HairStep evaluation is selection-biased. The text says: 'We use SD3 to translate sketches into photorealistic images and retain only successful reconstructions.' If the success filter is applied only to HairStep, then the evaluation set is not identical across methods, and HairStep is evaluated on a potentially easier subset of the test sketches. The comparison should be run on a fixed set of test sketches for all methods, or the same filtering should be applied to every method with the number of retained samples reported per method.
  3. [§3.1.1, Eq. (1)] The multi-scale decomposition assumes that max pooling and vector subtraction in the strand-VAE latent space are geometrically meaningful operations. The residual map is defined as H_k - tile(H_{k-1}), but the paper never validates that this latent-space arithmetic corresponds to actual strand geometry. If the assumption fails, the residual targets are degenerate and the reported gains could reflect the reconstruction pipeline rather than sketch-conditioned generation. Please validate by decoding the residual maps and comparing them with geometric strand differences, and by ablating with an alternative residual formulation (e.g., a learned projection) to confirm that the linear latent arithmetic is essential.
  4. [§4.2 and Supplementary A.1] The CLIP and LPIPS metrics compare the rendered output to the input sketch, which is exactly the conditioning signal the model is trained to follow. While this makes them reasonable measures of conditioning adherence, it also means they are not independent evidence of realism or ground-truth fidelity. The paper should explicitly frame these two metrics as conditioning-adherence checks and lean on the geometric metrics (PC-IoU, CD, Hausdorff) for fidelity claims, or supplement the semantic evaluation with a user study that compares outputs against the input sketch in a pairwise fashion.
minor comments (5)
  1. [§3.2, Eq. (5)] The alignment loss in Eq. (5) is under-specified: the expression σ(ĉ_k) log σ(ĉ_i) does not show a sum over feature dimensions, a sign, or which of the two terms is treated as the target distribution. Please clarify the exact loss formulation and how the original DINO embedding for the target scale is obtained.
  2. [§3.3, dataset description] The 24 self-collected hairstyles are not described in any detail; please report their acquisition method, preprocessing, alignment, and whether they will be released. Data provenance is important for reproducibility and for assessing the reported test-set composition.
  3. [Supplementary Tables 3–5] The ablation tables in the supplementary material also report single-run point estimates without error bars, so the differences among NN, BI, HAAR(mix), and Ours are not distinguished from noise; the same uncertainty-quantification remedy proposed for Table 2 should be applied here.
  4. [Abstract and §2.2] The claim of being the 'first sketch-based strand generation model' should be carefully scoped: DeepSketchHair is sketch-based but generates orientation fields rather than strands, and the text acknowledges this, but the abstract's unqualified phrasing may be misread; please explicitly state 'first sketch-based strand-level generator' or add a scoping clause.
  5. [§4.5, Fig. 8] The hand-drawn generalization result is only qualitative; given that the paper explicitly identifies limited generalization to hand-drawn sketches as a weakness, a small quantitative study on hand-drawn inputs (e.g., CLIP score or a user rating) would make the claimed practical utility more concrete.

Circularity Check

1 steps flagged · score 3.0 of 10

Semantic alignment metrics are self-referential to the sketch conditioning, but the geometric metrics and the core multi-scale derivation are independent.

  1. fitted input called prediction [Sec. 4.2 (Quantitative Comparison), Table 2 paragraph]
    "Semantic alignment between front-view rendered outputs and input sketches is evaluated via CLIP Score [9] and LPIPS [31]."

    The model is trained to generate strands conditioned on the input sketch: the diffusion loss L_diff (Eq. 2) is minimized with sketch embeddings injected through the multi-scale conditioning mechanism (Sec. 3.2), and the alignment loss L_align (Eq. 5) explicitly pushes adapted sketch embeddings toward scale-specific DINOv2 embeddings of the same sketches. The CLIP Score and LPIPS metrics then compare the rendered output to that same conditioning sketch. High values therefore partly reflect satisfaction of the training objective, not an independent discovery; any model that accurately copies its input sketch would score well on these metrics.

full rationale

The derivation chain is otherwise self-contained. The multi-scale residual encoding in Eq. 1 is an explicit additive decomposition, not a hidden reuse of the target result; the next-scale prediction and diffusion-head architecture are openly adapted from VAR and MAR with citations. There is no load-bearing self-citation chain: the pretrained strand VAE comes from HAAR, which is not authored by this paper's team, and comparisons to HairStep are external baselines rather than imported assumptions. The main circular element is limited to the semantic metrics in Tab. 2, which measure adherence to the same sketch used as conditioning and are therefore an expected consequence of the training setup. The independent geometric metrics and the qualitative ablations support the central claim of improved strand accuracy, so the overall circularity is partial and moderate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The contributions rest on several hand-tuned hyperparameters (K, compression rates, loss weight, CFG drop) and on domain assumptions about the strand VAE latent space, DINOv2 features, and synthetic sketches. The most fragile is the latent-space arithmetic assumption in Eq. 1, which is ad hoc to this paper and unvalidated.

free parameters (5)
  • Number of guiding scales K = 3
    Chosen by hand; Tab. 4 shows K=3 beats K=1 and K=2, but no principled or data-driven selection.
  • Per-scale latent compression rates = 1/4, 1/16, 1/16; latent lengths 256, 256, 1024
    Hand-set in Sec. 4.1; determines how much detail each scale can carry and directly affects the residual upsampling behavior.
  • Diffusion noise timestep cap for cumulative error simulation = T < 50
    Chosen ad hoc in Sec. 3.3 to make training robust to upsampling errors; no sensitivity analysis.
  • Alignment loss weighting = not reported
    L_align and L_diff are combined without stating a weight; this balance is a tuned free parameter.
  • CFG condition drop probability = 0.1
    Classifier-free guidance drop rate chosen in Sec. 4.1; standard but still a hyperparameter.
assumptions (4)
  • domain assumption HAAR's pretrained strand VAE faithfully encodes strand geometry in a 64-D latent space.
    Used without retraining in Sec. 3.1.1; every strand map and residual lives in this latent space.
  • ad hoc to paper Max pooling and vector subtraction in the strand latent space are geometrically meaningful.
    Eq. 1 defines H_hat_k via tile and subtraction; no validation that these operations correspond to actual strand downsampling or differences.
  • domain assumption DINOv2 embeddings of binarized sketches capture hair structure at multiple granularities.
    The adaptive conditioning and alignment loss in Sec. 3.2 rely on DINOv2 as the semantic anchor.
  • domain assumption Line-art-extractor renderings of 3D hair approximate user hand-drawn sketches.
    Training sketches are generated from rendered images (Sec. 3.3); hand-drawn generalization is only qualitatively shown in Sec. 4.5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StrandDesigner: Towards Practical Strand Generation with Sketch Guidance." pith.science (2026). https://pith.science/paper/FCKYNPJB

@misc{pith2026250801650,
  author       = {Pith},
  title        = {Pith review of: StrandDesigner: Towards Practical Strand Generation with Sketch Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCKYNPJB}},
  note         = {Machine review of arXiv:2508.01650}
}
read the original abstract

Realistic hair strand generation is crucial for applications like computer graphics and virtual reality. While diffusion models can generate hairstyles from text or images, these inputs lack precision and user-friendliness. Instead, we propose the first sketch-based strand generation model, which offers finer control while remaining user-friendly. Our framework tackles key challenges, such as modeling complex strand interactions and diverse sketch patterns, through two main innovations: a learnable strand upsampling strategy that encodes 3D strands into multi-scale latent spaces, and a multi-scale adaptive conditioning mechanism using a transformer with diffusion heads to ensure consistency across granularity levels. Experiments on several benchmark datasets show our method outperforms existing approaches in realism and precision. Qualitative results further confirm its effectiveness. Code will be released at [GitHub](https://github.com/fighting-Zhang/StrandDesigner).

Figures

Figures reproduced from arXiv: 2508.01650 by the authors.

Figure 1
Figure 1. Text prompts often fail to describe hairstyles pre [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed framework employing a learnable multi-scale upsampling strategy via next-scale prediction [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison across HAAR [23], Sketch+HAAR, HairStep [32], and our method. HAAR often ignores input conditions and defaults to short/medium styles. Sketch+HAAR improves accuracy via visual conditioning but still lacks detail fidelity. HairStep struggles with consistent geometry and produces artifacts such as incorrect back strands. In contrast, our method achieves high fidelity and precise control over div… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Ablation study on upsampling strategy. (a) Comparison to common alternatives (NN, BI, HAAR’s fixed approach) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation of the conditioning for initial strands: [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Hairstyle modification via sketch editing. Changes [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Generalization to hand-drawn sketches. 5 Conclusion We present the first sketch-conditioned framework for generating realistic 3D hair strands, which address the limitations of text and raw image prompts in existing methods, balancing user-friendliness with precise geo…
Figure 9
Figure 9. Figure 9: Sketch generation from rendered hair. Compared to [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Additional qualitative comparisons with HAAR [ [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 23 canonical work pages

  1. [1]

    Gaurav Bhokare, Eisen Montalvo, Elie Diaz, and Cem Yuksel. 2024. Real-time hair rendering with hair meshes. InACM SIGGRAPH 2024 Conference Papers. 1–10

    work page 2024

  2. [2]

    John Canny. 1986. A computational approach to edge detection.IEEE Transactions on pattern analysis and machine intelligence6 (1986), 679–698

    work page 1986

  3. [3]

    Matt Jen-Yuan Chiang, Benedikt Bitterli, Chuck Tappan, and Brent Burley. 2015. A practical and controllable hair and fur model for production path tracing. In ACM SIGGRAPH 2015 Talks. 1–1

    work page 2015

  4. [4]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first international conference on machine learning. StrandDesigner: Towards Practical Strand Generation with Sketch Guidance

    work page 2024

  5. [5]

    Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. 2024. Spectral Image Tokenizer.arXiv preprint arXiv:2412.09607(2024)

    arXiv 2024

  6. [6]

    Luca Fascione, Johannes Hanika, Rob Pieké, Ryusuke Villemin, Christophe Hery, Manuel Gamito, Luke Emrose, and André Mazzone. 2018. Path tracing in pro- duction. InACM SIGGRAPH 2018 Courses. 1–79

    work page 2018

  7. [7]

    2011.Beginning blender: open source 3d modeling, animation, and game design

    Lance Flavell. 2011.Beginning blender: open source 3d modeling, animation, and game design. Apress

    work page 2011

  8. [8]

    Chengan He, Xin Sun, Zhixin Shu, Fujun Luan, Sören Pirk, Jorge Alejan- dro Amador Herrera, Dominik L Michels, Tuanfeng Y Wang, Meng Zhang, Holly Rushmeier, et al. 2024. Perm: A parametric representation for multi-style 3d hair modeling.arXiv preprint arXiv:2407.19451(2024)

    arXiv 2024

Show all 37 references
  1. [9]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi

  2. [10]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  3. [11]

    Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598(2022)

    2022 arXiv

  4. [12]

    Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. 2015. Single-view hair modeling using a hairstyle database.ACM Transactions on Graphics (ToG)34, 4 (2015), 1–9

    2015

  5. [13]

    Zhiyi Kuang, Yiyang Chen, Hongbo Fu, Kun Zhou, and Youyi Zheng. 2022. Deepmvshair: Deep hair modeling from sparse views. InSIGGRAPH Asia 2022 Conference Papers. 1–8

    2022

  6. [14]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning. PMLR, 12888–12900

    2022

  7. [15]

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. 2024. Autoregressive image generation without vector quantization.Advances in Neural Information Processing Systems37 (2024), 56424–56445

    2024

  8. [16]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

    2023

  9. [17]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193(2023)

    2023 arXiv

  10. [18]

    Qiaomu Ren, Haikun Wei, and Yangang Wang. 2021. Hair Salon: A Geometric Example-Based Method to Generate 3D Hair Data. InInternational Conference on Image and Graphics. Springer, 533–544

    2021

  11. [19]

    Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 2018. 3D hair synthesis using volumetric variational autoencoders.ACM Transactions on Graphics (TOG)37, 6 (2018), 1–12

    2018

  12. [20]

    Yuefan Shen, Shunsuke Saito, Ziyan Wang, Olivier Maury, Chenglei Wu, Jessica Hodgins, Youyi Zheng, and Giljoo Nam. 2023. Ct2hair: High-fidelity 3d hair modeling using computed tomography.ACM Transactions on Graphics (TOG)42, 4 (2023), 1–13

    2023

  13. [21]

    Yuefan Shen, Changgeng Zhang, Hongbo Fu, Kun Zhou, and Youyi Zheng. 2020. Deepsketchhair: Deep sketch-based 3d hair modeling.IEEE transactions on visualization and computer graphics27, 7 (2020), 3250–3263

    2020

  14. [22]

    Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. 2023. Neural haircut: Prior-guided strand-based hair reconstruction. InProceedings of the IEEE/CVF International Conference on Computer Vision. 19762–19773

    2023

  15. [23]

    Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J Black, and Justus Thies. 2024. Text-conditioned generative model of 3d strand-based human hairstyles. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4703–4712

    2024

  16. [24]

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. 2024. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems37 (2024), 84839–84865

    2024

  17. [25]

    Tao Wang and Xue Dong Yang. 2004. Hair design based on the hierarchical cluster hair model.Geometric modeling: techniques, applications, systems and tools(2004), 329–359

    2004

  18. [26]

    Haomiao Wu, Alvin Shi, AM Darke, and Theodore Kim. 2024. Curly-Cue: Geo- metric Methods for Highly Coiled Hair. InSIGGRAPH Asia 2024 Conference Papers. 1–11

    2024

  19. [27]

    Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng. 2024. Monohair: High-fidelity hair modeling from a monocular video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 24164–24173

    2024

  20. [28]

    Zhan Xu and Xue Dong Yang. 2001. V-hairstudio: an interactive tool for hair design.IEEE Computer Graphics and Applications21, 3 (2001), 36–43

    2001

  21. [29]

    Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J Black. 2024. TECA: Text-guided generation and editing of compositional 3d avatars. In2024 International Conference on 3D Vision (3DV). IEEE, 1520–1530

    2024

  22. [30]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF inter- national conference on computer vision. 3836–3847

    2023

  23. [31]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  24. [32]

    Yujian Zheng, Zirong Jin, Moran Li, Haibin Huang, Chongyang Ma, Shuguang Cui, and Xiaoguang Han. 2023. Hairstep: Transfer synthetic to real using strand and depth maps for single-view 3d hair modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2023

  25. [33]

    Yujian Zheng, Yuda Qiu, Leyang Jin, Chongyang Ma, Haibin Huang, Di Zhang, Pengfei Wan, and Xiaoguang Han. 2024. Towards Unified 3D Hair Reconstruction from Single-View Portraits. InSIGGRAPH Asia 2024 Conference Papers. 1–11

    2024

  26. [34]

    Yuxiao Zhou, Menglei Chai, Alessandro Pepe, Markus Gross, and Thabo Beeler

  27. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  28. [2021]

    Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718(2021)

    2021 arXiv

  29. [2023]

    Zhang et al

    Groomgen: A high-quality generative hair model using hierarchical latent representations.ACM Transactions on Graphics (TOG)42, 6 (2023), 1–16. Zhang et al. Supplementary Material A Additional Implementation Details A.1 Evaluation Metrics Since there is no established benchmark...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.