CartoonAlive: Towards Expressive Live2D Modeling from Single Portraits

Chao He; Jianjing Xiang; Jianqiang Ren; Xiejie Shen

REVIEW 4 major objections 5 minor 22 references

CartoonAlive: Towards Expressive Live2D Modeling from Single Portraits

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 CartoonAlive claims the first fully automated pipeline that generates a complete, animatable Live2D character from a single portrait photo in under 30 seconds.

desk verdict A plausible first step for automatic single-image-to-Live2D, but the identity claim rests on a trained mapping with no domain-shift test and no quantitative evaluation. read the letter →

arxiv 2507.17327 v1 pith:G4NR4YK3 submitted 2025-07-23 cs.CV

Chao He , Jianqiang Ren , Jianjing Xiang , Xiejie Shen This is my paper

classification cs.CV

keywords Live2Dsingleportraitblendshapesfaciallandmarkregressionavatargenerationcharacteranimationsynthetictrainingdatadigitalhuman

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

CartoonAlive aims to generate a complete, animatable Live2D cartoon character from a single portrait photo, fully automatically and in less than 30 seconds. The paper's core idea is to adapt the blendshape concept from 3D face modeling to 2D Live2D: each facial component gets a set of basis shapes for horizontal position, vertical position, and scale, and a face is a weighted combination of these bases. A neural network maps facial landmarks from the input image to the blendshape weights, trained on 100,000 synthetic rendered faces with artificially marked keypoints. The pipeline also repaints the underlying face to hide artifacts during animation and transfers hair as a separate texture layer. If it works, it would replace manual Live2D rigging with a one-click photo-to-avatar pipeline.

What carries the argument

The core machinery is the Live2D blendshape basis: for each facial component (left eye, right eye, nose, mouth) a set of basis shapes along horizontal shift $x$, vertical shift $y$, and scale, with weights $\omega$ ranging from $-30$ to $30$, so the full face is $F = \bar{F} + \omega_{\text{left eye}} B_{\text{left eye}} + \omega_{\text{right eye}} B_{\text{right eye}} + \omega_{\text{nose}} B_{\text{nose}} + \omega_{\text{mouth}} B_{\text{mouth}}$. This reduces identity to a 12-dimensional parameter vector (three axes per component) that a 4-layer MLP is trained to regress from facial landmark coordinates. The synthetic training protocol, where 100,000 renders have their feature regions blacked out and keypoints drawn as white dots, is what makes the landmark-to-parameter mapping learnable without manual labels; the mask-based repainting completes the system by removing animation artifacts from the underlying face texture.

What would settle it

Take 100 real portraits, have an artist manually rig each into a Live2D model using the same blendshape parameter space, and compare the MLP's predicted parameter vector to the artist's ground truth; if the average per-axis error exceeds visual tolerance (roughly 3–5 pixels on a 1024×1024 canvas) on more than a small fraction of examples, the identity-preservation claim fails.

Watch

Extended reading notes

Core claim

The paper claims that a complete Live2D character, with layered facial components and identity-preserving deformation, can be generated end-to-end from a single portrait. The generation is driven by a linear blendshape model of the face: a base face plus weighted basis shapes for the left eye, right eye, nose, and mouth, each with horizontal, vertical, and scale parameters ranging from −30 to 30. The weights are predicted by a 4-layer MLP that takes normalized facial landmark coordinates as input; the MLP is trained on 100,000 synthetic pairs rendered with a synthetic renderer, where facial feature regions are blacked out and keypoints are drawn as white dots for reliable detection. At inference, the input portrait is aligned to a template, facial features are temporarily removed, landmarks are extracted from the underlying face using a landmark detection pipeline, and the MLP predicts the Live2D parameters. The resulting model is described as faithfully reproducing the input subject's facial identity, is driven by 52 expression controls, and is generated without manual binding.

Load-bearing premise

The MLP is trained only on synthetic renders whose keypoints are white dots, but at inference it must read real-world facial landmarks; if those two landmark distributions are far apart, predicted blendshape weights will not preserve identity.

Editorial extensions

If this is right

A single portrait photo suffices to produce a rigged, animatable Live2D character in under 30 seconds, making manual rigging unnecessary for this class of avatars.
Because the parameter space covers horizontal, vertical, and scale shifts for eyes, nose, and mouth, the generated model can match a wide range of facial identities and is driven by 52 expression controls.
The synthetic-data training recipe (white-dot keypoints on blacked-out features) could be reused for other landmark-to-parameter regression problems in 2D character creation.
The dynamic artifact correction, which repaints the underlying face according to masks rendered from predicted parameters, means the character stays coherent during eye-closing and other expressions.
The method extends beyond cartoon style to realistic and 3D-cartoon portraits, albeit with added difficulty.

Reading between the lines

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

The biggest unverified risk is the synthetic-to-real landmark gap: the MLP only saw white-dot markers in rendered faces, so its behaviour on real-world facial landmarks could be tested by comparing predictions against artist-rigged ground truths on real portraits.
If the domain gap is bridged, the same blendshape formulation could be applied to other 2D character formats beyond Live2D, such as spine or custom layered rigs, by relabeling the parameter space.
The 12-parameter linear model (three axes × four components) may be too rigid to capture iris size, ear shape, or asymmetric features, which the paper itself flags; a natural extension is per-component shape bases learned from real portrait data rather than hand-defined shifts and scales.
The dependence on a landmark detector and a hair-removal model suggests that accuracy will vary with pose, occlusion, and hairstyle; a systematic stress test across poses and occlusions would reveal where the pipeline breaks.

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.

Desk Editor's Note

read the letter

Worth a look if you care about practical avatar generation. The genuinely new piece is the application: take one portrait, output a rigged, animatable Live2D model within 30 seconds, no manual binding. That's real. Pipeline is described carefully: blendshape parameterization borrowed from 3DMM, a 4-layer MLP trained on 100k synthetic renders to map landmarks to weights, alignment via Mediapipe, repainting step, hair transfer. Limitation section honestly names ears, pupils/irises, hair strands. Related work is fine, including their own Textoon.

Soft spots are real and load-bearing. First, no quantitative evaluation anywhere: no metrics, no ablations, no user study, no error bars. The central claims - "faithfully reproduce identity" and "artifact-free animation" - rest entirely on qualitative figures. For an application paper whose value is exactly whether it works across diverse inputs, that is a big gap. Second, the domain-shift issue is exactly what the stress-test note flags. The MLP is trained on 100k PyGame renders where landmarks are white dots on blacked-out features, but inference consumes Mediapipe landmarks from real photos, then landmarks from a rendered underlying face after features are removed. Nothing shows these two landmark distributions match. A 4-layer regression will happily propagate a shift in scale, noise, or semantic correspondence into wrong blendshape weights, breaking identity. This seems more central than the listed limitations, and it isn't mentioned. Also, no code or data. That matters less for a technical report, but combined with missing evaluation, an interested reader cannot independently check the core claim.

I largely agree with the reader's conditional verdict, though I'd put less weight on the "internal evaluation" concern and more on the domain gap. The method is coherent and the authors are honest as far as they go, but this is a promising demo, not a verified system.

Who gets value: people building avatar generation tools or production pipelines for interactive characters. It should go to peer review - it's a novel application with a clear pipeline - but a serious referee should ask for quantitative identity-preservation metrics, an ablation of the MLP, a domain-shift check between synthetic and real landmarks, and ideally a small user study. I'd engage with it, but only after those additions.

Referee Report

4 major / 5 minor

Summary. The paper presents CartoonAlive, an automated pipeline that takes a single portrait photograph and generates a complete Live2D-style 2D cartoon avatar with animatable facial components. The method introduces blendshape parameters (horizontal shift, vertical shift, and scale, each in [-30, 30]) for facial components such as eyes, nose, and mouth, borrowing the shape-basis idea from 3DMM. An MLP is trained on a 100,000-sample synthetic dataset rendered with PyGame, where facial features are blacked out and keypoints are marked as white dots, to regress blendshape weights from facial landmarks. At inference, Mediapipe landmarks from the real input are used, the face is aligned to a template, the underlying face is repainted to avoid animation artifacts, and hair is transferred via segmentation. The paper claims this is the first fully automated end-to-end Live2D generation from a single portrait, achieving identity-faithful results in under 30 seconds. The evaluation in Section 4 consists entirely of qualitative figures (Figures 1, 6, 7, 8), with no quantitative metrics, user study, or ablation.

Significance. If the claims are validated, CartoonAlive would be a practically useful contribution to interactive 2D avatar creation, filling a gap between 3D digital humans and 2D video-based methods. The concept of applying 3DMM-style blendshapes to Live2D is reasonable, and the pipeline is described in enough detail to be plausible. However, the central claim of 'faithfully reproducing the facial identity of the input subject' rests entirely on qualitative visual results. The training/inference domain gap between synthetic white-dot landmarks and real Mediapipe landmarks is a critical correctness risk that is not addressed. The paper also provides no quantitative evidence for the claimed speed, identity preservation, or artifact-free animation. The potential significance is therefore conditional: the core idea is promising, but the current evidence is insufficient to support the stated contributions.

major comments (4)

[Sec. 3.3 and Sec. 3.5] The MLP is trained exclusively on synthetic PyGame renders where keypoints are white dots on a blacked-out face (Sec. 3.3), but during inference it consumes Mediapipe landmarks extracted from a re-rendered image built from the real input portrait (Sec. 3.5). The paper provides no evidence that these two landmark distributions are comparable in terms of scale, noise, semantic correspondence, or typical feature positions. If the distributions differ, the predicted blendshape weights will be biased, directly harming the central claim of identity-preserving generation. I ask the authors to supply a quantitative domain-gap analysis, such as a comparison of landmark coordinate distributions on synthetic versus real inputs, and to report the parameter prediction error on real portraits (e.g., by manual annotation of ground-truth parameter values). Alternatively, they can fine-tune or adapt the MLP with real-landmark data and evaluate the improvement.
[Sec. 4 (Results)] The evaluation is purely qualitative. Figures 1, 6, 7, and 8 show example outputs, but there are no quantitative metrics, error bars, ablations, or user studies supporting the claims of 'high fidelity,' 'identity preservation,' or 'artifact-free animation.' Given that the paper's main contribution is automatic identity-preserving Live2D generation, I require quantitative evidence: e.g., facial landmark distance between input and generated avatar, a face-verification similarity score, perceptual similarity metrics, a user study on identity preservation and animation quality, and a measurement of the claimed 'less than 30 seconds' runtime under a controlled setting. Without such evidence, the central claims are unsupported.
[Sec. 3.2, Eq. (1)] The blendshape equation is formally incorrect as printed: it reads 'F = F + ω_left_eye B_left_eye + ...', with the same symbol F on both sides. The text says 'F denotes the base face with all parameters set to zero,' but the notation is inconsistent, and the equation makes the linear combination degenerate. The formulation should be written with a distinct base-face symbol, e.g., F = F0 + Σω_i B_i, and the meaning of the ω ranges ([-30, 30]) should be defined relative to the actual displacement or scale change they induce. This matters because the synthetic data generation in Sec. 3.3 samples these ω values, and an undefined or erroneous formulation undermines the reproducibility of the training data.
[Sec. 3.3 (Training)] The description of the synthetic dataset and MLP training lacks crucial details: the sampling distribution of ω (uniform over [-30, 30]?), the rendering configuration (camera, template mesh, landmark definitions), the exact input feature vector (which landmarks, how normalized), the network architecture (widths, activation functions), the training loss weighting, and the convergence criterion. Without these details, the reader cannot assess whether the learned mapping is well-posed, whether the parameter space is sufficiently explored, or whether the reported behavior is reproducible. I ask the authors to specify these components, and ideally to release the code and synthetic dataset to allow scrutiny.

minor comments (5)

[Figure 2] The figure is dense and the labels are somewhat jumbled (e.g., 'using facial texture rendered without facial featuresdetect facial keypointsin the rendered image' appears as a run-on). Please re-format the figure with clear arrows and a more readable panel layout.
[Sec. 3.5] The statement 'we use a large bounding box to represent the eyebrow area' is vague; please specify the box size relative to the face template and how it is derived, since this directly affects the robustness claim.
[Abstract and Sec. 4] The phrase 'visually accurate' is undefined. Consider defining an explicit quantitative criterion (e.g., landmark or perceptual similarity) or rephrasing to 'visually plausible' to avoid ambiguity.
[Sec. 5 (Limitation)] The limitation section mentions ears, pupils/iris, and hair strands, but does not mention the synthetic-to-real domain gap or the lack of quantitative evaluation. Both are more central to the paper's claims and should be acknowledged.
[References] Some references are incomplete or informal (e.g., [1] is a URL without a full citation, [13] is a product website). Please provide complete bibliographic entries for all references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the landmark-to-parameter regression is supervised on synthetic renders with known parameters, and the paper's self-citations are contextual rather than load-bearing.

full rationale

The paper's central inference step is the 4-layer MLP that maps facial landmark coordinates to Live2D blendshape parameters (Sec. 3.3). Training data are created by randomly sampling parameter sets, rendering faces with PyGame, and extracting landmark annotations from white-dot keypoints; the network is then trained with MSE loss. This is an ordinary supervised regression setup, not a circular one: the prediction target is the known sampled parameter set, and the inputs are landmarks generated independently of the test portraits. At inference, landmarks from real portraits are fed into the same network, and no fitted quantity is renamed as a prediction. The blendshape formula F = F0 + ΣωB is an explicit modeling definition, not a derived result that reduces to its input. The paper's self-citations, Textoon [8] and Make-A-Character [16], appear only in the related-work discussion and do not carry the derivation; no uniqueness theorem from prior work is invoked. The identity-fidelity claim is partly supported by direct transfer of facial-feature textures from the input, but the parameter prediction itself remains an independent learned mapping. The synthetic-to-real landmark distribution shift highlighted by the reader is a generalization and evaluation risk, not a circularity. Accordingly, no circular step can be identified with the required specificity, and the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The system relies on several unverified domain assumptions, most critically the equivalence between synthetic white-dot landmarks and real Mediapipe landmarks. The free parameters are design choices not justified by experiments. No new physical entities are introduced.

free parameters (4)

Blendshape parameter range [-30, 30] = -30 to 30
Manually chosen bounds for horizontal, vertical, and scale parameters in the blendshape linear model (Sec 3.2); no analysis is given for why this range is sufficient.
MLP layers (4) = 4 layers, hidden sizes not specified
Architecture chosen without ablations; hidden dimensions and activations are not reported (Sec 3.3).
Synthetic dataset size = 100,000 samples
The dataset scale is stated without analysis of whether it covers the parameter space adequately (Sec 3.3).
Eyebrow bounding box size = large
Heuristic to compensate for eyebrow landmark variability; exact size and shape are unspecified (Sec 3.5).

assumptions (3)

domain assumption Mediapipe keypoints detected on real portraits are comparable to the white-dot landmarks rendered in synthetic training images.
The whole parameter prediction pipeline depends on this equivalence (Sec 3.3 and 3.4).
domain assumption PyGame-rendered facial images with blacked-out features and white-dot keypoints provide ground-truth labels suitable for training a regressor that transfers to real photos.
The synthetic dataset is the sole source of supervision (Sec 3.3).
domain assumption HairMapper can remove bangs or occluding hair without changing the underlying facial geometry or appearance.
Used as a preprocessing step when hair occludes eyebrows or eyes (Sec 3.7); no evaluation of failure cases is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CartoonAlive: Towards Expressive Live2D Modeling from Single Portraits." pith.science (2026). https://pith.science/paper/G4NR4YK3

@misc{pith2026250717327,
  author       = {Pith},
  title        = {Pith review of: CartoonAlive: Towards Expressive Live2D Modeling from Single Portraits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4NR4YK3}},
  note         = {Machine review of arXiv:2507.17327}
}

read the original abstract

With the rapid advancement of large foundation models, AIGC, cloud rendering, and real-time motion capture technologies, digital humans are now capable of achieving synchronized facial expressions and body movements, engaging in intelligent dialogues driven by natural language, and enabling the fast creation of personalized avatars. While current mainstream approaches to digital humans primarily focus on 3D models and 2D video-based representations, interactive 2D cartoon-style digital humans have received relatively less attention. Compared to 3D digital humans that require complex modeling and high rendering costs, and 2D video-based solutions that lack flexibility and real-time interactivity, 2D cartoon-style Live2D models offer a more efficient and expressive alternative. By simulating 3D-like motion through layered segmentation without the need for traditional 3D modeling, Live2D enables dynamic and real-time manipulation. In this technical report, we present CartoonAlive, an innovative method for generating high-quality Live2D digital humans from a single input portrait image. CartoonAlive leverages the shape basis concept commonly used in 3D face modeling to construct facial blendshapes suitable for Live2D. It then infers the corresponding blendshape weights based on facial keypoints detected from the input image. This approach allows for the rapid generation of a highly expressive and visually accurate Live2D model that closely resembles the input portrait, within less than half a minute. Our work provides a practical and scalable solution for creating interactive 2D cartoon characters, opening new possibilities in digital content creation and virtual character animation. The project homepage is https://human3daigc.github.io/CartoonAlive_webpage/.

Figures

Figures reproduced from arXiv: 2507.17327 by the authors.

**Figure 2.** Overview of the CartoonAlive pipeline. (a) Facial Feature Alignment: The input portrait is first preprocessed to align the eyes horizontally, ensuring consistent orientation. Then, facial keypoints for the eyes, nose, mouth, eyebrows, and facial contour are individually detected. A transformation is computed between each set of detected keypoints and those of a predefined template model. Based on this correspondence… view at source ↗

**Figure 3.** Live2D facial texture and its binding. 3.1. Preliminary of Live2D A Live2D character is composed of multiple layered components, including the underlying face, eyebrows, eyes, nose, mouth, body, and clothing. Each component is defined by a polygonal mesh that controls its deformation during animation. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Nose blendshape creation with three dimensions: horizontal shift, vertical shift, and scale. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Accurate facial feature keypoints are obtained by detecting the white dots in the rendered images. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: The overall animation effects of the generated Live2D model. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Examples of Live2D cartoon characters created from input portraits. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Generated Live2D models from different artistic styles, including realistic human faces and 3D cartoon-style portraits. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

[1]

https://arkit-face-blendshapes.com/

Arkit face blendshapes. https://arkit-face-blendshapes.com/. 7

work page
[2]

Principal component analysis

Herv´e Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4): 433–459, 2010. 3

work page 2010
[3]

A morphable model for the synthesis of 3d faces

V olker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157–164. 2023. 3

work page 2023
[4]

Emoca: Emotion driven monocular face capture and animation

Radek Danˇeˇcek, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311–20322, 2022. 3

work page 2022
[5]

Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set

Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3

work page 2019
[6]

Semi-supervised 3d face representation learning from unconstrained photo collections

Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 348–349, 2020. 3

work page 2020
[7]

Liveportrait: Efficient portrait animation with stitching and retargeting control

Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4

arXiv 2024
[8]

Textoon: Generating Vivid 2D Cartoon Characters from Text Descriptions

Chao He, Jianqiang Ren, Yuan Dong, Jianjing Xiang, Xiejie Shen, Weihao Yuan, and Liefeng Bo. Textoon: Generating vivid 2d cartoon characters from text descriptions. arXiv preprint arXiv:2501.10020, 2025. 3 9

work page Pith review arXiv 2025

Show all 22 references

[9]

A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images

Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 394...

2023
[10]

Learning a model of facial shape and expression from 4d scans

Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194–1, 2017. 3

2017
[11]

Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks

Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pages 5891–5900, 2020. 3

2020
[12]

Mediapipe: A framework for building perception pipelines

Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172,

1906 arXiv
[13]

Tetsuya Nakajo. Live2d. https://www.live2d.com. 2
[14]

A 3d face model for pose and illumination invariant face recognition

Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance , pages 296–301. Ieee, 2009. 3

2009
[15]

Sdxl: Improving latent diffusion models for high-resolution image synthesis

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3

2023 arXiv
[16]

Make-a-character: High quality text-to-3d character generation within minutes

Jianqiang Ren, Chao He, Lin Liu, Jiahao Chen, Yutong Wang, Yafei Song, Jianfang Li, Tangli Xue, Siqi Hu, Tao Chen, Kunkun Zheng, Jianjing Xiang, and Liefeng Bo. Make-a-character: High quality text-to-3d character generation within minutes. arXiv preprint arXiv:2312.15430, 2023. 3

2023 arXiv
[17]

Learning to regress 3d face shape and expression from an image without 3d supervision

Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7763–7772,
[18]

Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction

Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. InProceedings of the IEEE international conference on compute...

2017
[19]

Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions

Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. In European Conference on Computer Vision, pages 244–260. Springer, 2024. 4

2024
[20]

Beyond regression: New tools for prediction and analysis in the behavioral sciences

Paul Werbos. Beyond regression: New tools for prediction and analysis in the behavioral sciences. PhD thesis, Committee on Applied Mathematics, Harvard University, Cambridge, MA, 1974. 2

1974
[21]

Hairmapper: Removing hair from portraits using gans

Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4227–4236, 2022. 7

2022
[22]

Adding conditional control to text-to-image diffusion models, 2023

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3 10

2023

Pith tools

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

[1] [1]

https://arkit-face-blendshapes.com/

Arkit face blendshapes. https://arkit-face-blendshapes.com/. 7

work page

[2] [2]

Principal component analysis

Herv´e Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4): 433–459, 2010. 3

work page 2010

[3] [3]

A morphable model for the synthesis of 3d faces

V olker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157–164. 2023. 3

work page 2023

[4] [4]

Emoca: Emotion driven monocular face capture and animation

Radek Danˇeˇcek, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311–20322, 2022. 3

work page 2022

[5] [5]

Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set

Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3

work page 2019

[6] [6]

Semi-supervised 3d face representation learning from unconstrained photo collections

Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 348–349, 2020. 3

work page 2020

[7] [7]

Liveportrait: Efficient portrait animation with stitching and retargeting control

Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4

arXiv 2024

[8] [8]

Textoon: Generating Vivid 2D Cartoon Characters from Text Descriptions

Chao He, Jianqiang Ren, Yuan Dong, Jianjing Xiang, Xiejie Shen, Weihao Yuan, and Liefeng Bo. Textoon: Generating vivid 2d cartoon characters from text descriptions. arXiv preprint arXiv:2501.10020, 2025. 3 9

work page Pith review arXiv 2025