From Rigging to Waving: 3D-Guided Diffusion for Natural Animation of Hand-Drawn Characters

Hongbo Fu; Jie Zhou; Linzi Qu; Miu-Ling Lam

REVIEW 6 major objections 7 minor 4 references

From Rigging to Waving: 3D-Guided Diffusion for Natural Animation of Hand-Drawn Characters

T0 review · 6 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A hybrid of skeletal animation and video diffusion animates a single hand-drawn picture in 3D while preserving style and adding natural hair and cloth motion.

desk verdict Plausible hybrid pipeline with real new modules, but the headline claim of natural secondary dynamics is supported only by static texture metrics and an unreported user study. read the letter →

arxiv 2509.06573 v1 pith:JKGN3MYR submitted 2025-09-08 cs.GR

Jie Zhou , Linzi Qu , Miu-Ling Lam , Hongbo Fu This is my paper · ORCID

classification cs.GR

keywords hand-drawncharacteranimationvideodiffusionmodelsskeletalsecondarymotionsingle-image3Dreconstructioninpaintingrigginghair

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

The paper tries to settle a trade-off in hand-drawn character animation: skeletal rigging keeps a drawing's identity stable but renders hair and clothing stiffly, while video diffusion models produce lively motion but drift into real-human artifacts and broken contours. It claims that the two can be combined, in a split of labor, by first retargeting the drawing onto a rough 3D character, rendering coarse pose frames, and then treating the final output as an inpainting problem for a domain-adapted video diffusion model. The diffusion model is asked to refine only what skeletal animation does poorly, and a Secondary Dynamics Injection step blends in latent estimates from a pre-trained human-motion diffusion model so that masked regions such as hair ends and skirts gain realistic secondary dynamics. A Hair Layering Modeling step separates hair from body geometry before rigging so that long-haired characters do not deform unnaturally. If the claims hold, an animator needs only one drawing, a target 3D motion, and optionally a few masks to get stylized, editable animations with natural secondary motion.

What carries the argument

The load-bearing object is the three-part guidance-and-refinement loop built around a rough 3D character: a single-image-to-3D module reconstructs a low-poly proxy with optional hair/body separation via segmentation maps; a domain-adapted video diffusion model conditioned on rendered coarse sequences refines appearance as an inpainting task; and Secondary Dynamics Injection blends v-prediction latent estimates $\hat{z}^{v_\theta}_{n,0}$ and $\hat{z}^{u_\theta}_{n,0}$ through masks $M^{\text{SDI}}_{n,\text{down}}$ during a middle denoising interval, followed by Poisson blending and a full re-denoise. The mechanism works because the denoising trajectory separates structure from detail and secondary motion, so the pre-trained model's real-human motion priors can be injected where secondary dynamics live without letting it redraw the whole character.

What would settle it

Systematically degrade the 3D proxy quality, for example by feeding the pipeline drawings whose reconstruction has known limb-length or joint errors, and measure whether the final animation's pose accuracy and identity preservation degrade in step. If animation quality stays flat while proxy error grows, the 3D guidance is not actually carrying the result; likewise, a held-out set of long-haired back views should show the reported shoulder-texture errors if the hair-body separation is only partial.

Watch

Extended reading notes

Core claim

The central claim is that geometric consistency and expressive dynamics are not competing requirements for hand-drawn characters; they can be assigned to different stages of one pipeline. Skeletal animation on a reconstructed 3D proxy supplies coarse guidance sequences that carry identity, viewpoint, and primary motion, and the diffusion stage is cast as inpainting so it redraws only user-marked regions rather than re-imagining the whole drawing. The paper's key observation is that in v-prediction denoising, early steps fix spatial structure and primary motion while later steps add secondary dynamics; the Secondary Dynamics Injection exploits this by blending the denoised latent estimates of the fine-tuned stylized model and the pre-trained real-human model over a middle denoising interval, then re-denoises from scratch with Poisson-blended inpainted coarse frames. Hair Layering Modeling separates hair from body in the implicit field using segmentation maps, so the reconstructed character has separate hair and body geometry instead of one fused low-poly mesh. The authors report that this outperforms both skeletal-only and diffusion-only baselines on their test set and that the components are each necessary in ablation.

Load-bearing premise

The whole pipeline leans on the reconstructed and auto-rigged low-poly 3D character being accurate enough that its rendered coarse frames are trustworthy guides for identity and pose; if single-image reconstruction or auto-rigging goes wrong, the diffusion refinement inherits the error, and the paper itself notes back-view long-hair regions still misbehave.

Editorial extensions

If this is right

Given a single frontal drawing and a target 3D motion, the pipeline outputs a stylized 3D animation that preserves the drawing's identity, including long-hair characters that skeletal methods deform.
User-supplied masks can steer which regions receive secondary dynamics, and edits made to the reference frame propagate across the whole animation without re-rigging.
Because refinement is framed as inpainting, the system needs only a small stylized training set; fine-tuning spatial layers while freezing temporal layers preserves the motion prior and avoids catastrophic forgetting.
The method does not require physics simulation or manual multi-layered rigging, so it lowers the barrier for novices to animate single drawings.

Reading between the lines

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

Inference: the denoising-phase observation generalizes beyond hand-drawn characters, so any pose-conditioned video diffusion model could adopt the blend-and-redenoise trick to trade fidelity against injected dynamics in other stylized domains, such as furry creatures or loose clothing, that lack large training datasets.
Inference: the quality ceiling is set by the reconstructed 3D proxy, so as single-image-to-3D reconstruction improves, the same pipeline should improve without retraining; a testable prediction is that animation fidelity correlates with proxy reconstruction accuracy.
Inference: the SDI masks could be turned into an interactive 'dynamics brush' where the user paints where motion should be amplified and the system re-runs only the affected denoising phases, making animation editing a local, near-real-time operation.
Inference: Hair Layering Modeling could be extended to layered clothing by segmenting multiple garment layers in the implicit field rather than just hair versus body.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Short version: this is a genuinely hybrid system with two new mechanisms that address real failure modes, and the qualitative story is mostly convincing. But the central claim — that it produces natural secondary dynamics better than existing methods — is not backed by the evaluation. Table 1 reports only static per-frame texture scores (LPIPS/FID/CLIP), with gaps like 0.1733 vs 0.1734, no error bars, no significance tests, and no motion-aware metric. The perceptual user study promised in the Introduction is absent from the manuscript.

What's actually new: the SDI strategy — blending latent estimates from the adapted and pretrained diffusion models only inside user-specified masks during a middle denoising window, then re-denoising with Poisson-blended coarse frames — is a real mechanism, and the ablation shows it does something visible. The HLM hair/body separation via segmentation maps is a practical fix for a real artifact of single-mesh low-poly reconstruction. Fine-tuning only spatial layers while freezing temporal layers is a sensible choice given the small dataset, and the inpainting formulation of coarse-render enhancement is reasonable.

The soft spots are mostly in the evaluation. The same-source dependency is real: the training video is synthesized from DrawingSpinUp, and DrawingSpinUp is also a comparison baseline. This doesn't make the work circular, because the final output is not derived from that ground truth alone, but it makes the comparison against DrawingSpinUp more like an ablation of their own prior pipeline than a neutral SOTA comparison. Excluding UniAnimate, AnimateAnyone, and MikuDance from Table 1 is defensible if they fail on pose, but then the quantitative comparison is only against two skeletal baselines plus a fine-tuned UniAnimate. The "observed for the first time" claim about denoising steps is overstated; the phenomenon of early structure and late detail is known. The code and data are promised in an anonymous repository but not verifiable from the submission.

To the authors' credit, the limitations are stated honestly: inverted poses and back-view long hair remain problematic. That is the right kind of candor.

Who this is for: graphics and animation researchers working on hand-drawn character animation, or on applying video diffusion to stylized content. It deserves a serious referee, but with a clear expectation of major revision: either add motion-aware evaluation (temporal consistency, pose accuracy, a real user study) or tone down the abstract. I would not desk-reject it.

Referee Report

6 major / 7 minor

Summary. The paper proposes a hybrid animation system for hand-drawn characters that combines skeletal animation with video diffusion priors. A 3D proxy is reconstructed from a single drawing via Wonder3D and Mixamo, a Hair Layering Modeling (HLM) step separates hair from the body for long-hair characters, and the rigged model is retargeted to target motions to render pose, mask, and coarse color guidance sequences. A domain-adapted diffusion model, built on UniAnimate and fine-tuned on a synthesized drawing-animation dataset, refines these coarse frames. Secondary Dynamics Injection (SDI) blends latent estimates from the domain-adapted and pre-trained models during early denoising steps, followed by a re-denoising phase with Poisson-blended coarse guidance. The paper claims superior performance over both skeletal and diffusion-based state-of-the-art methods in quantitative and qualitative evaluations, and it reports ablations for HLM, the coarse prior encoder, spatial-layer tuning, and SDI components and hyperparameters.

Significance. If the claims are substantiated, the hybrid skeletal-plus-diffusion pipeline would be a useful contribution to stylized character animation: it directly addresses the geometric-consistency versus expressive-dynamics trade-off, and the HLM and SDI ideas are plausible and likely to interest the computer graphics community. The manuscript is also commendable for providing an anonymous code/data repository, detailed pipeline diagrams, and a broad set of ablations of the main components. However, the central quantitative claim of outperforming state-of-the-art methods is not currently supported by the evidence in the paper, and the key contribution (SDI) is evaluated only through selected qualitative frames. The paper is therefore best viewed as presenting a promising system whose headline claims require substantially stronger empirical support before publication.

major comments (6)

[Section 4.4, Table 1] The quantitative evidence for the abstract's claim that the method 'outperforms state-of-the-art methods in both quantitative and qualitative evaluations' is insufficient. The differences in Table 1 are extremely small (LPIPS 0.1733 vs. 0.1734, FID 152.90 vs. 157.75, CLIP 0.9030 vs. 0.8964), no error bars or significance tests are reported, and the metrics are computed per-frame against the reference image, measuring texture/semantic fidelity rather than motion quality, temporal consistency, pose accuracy, or secondary dynamics. The paper should add error bars across evaluation clips, significance tests, and motion-aware or temporal-consistency metrics (e.g., warping error, flow consistency, or a well-specified user study) before claiming quantitative superiority.
[Section 4.4] The main diffusion baselines are excluded from the quantitative comparison. The paper states that UniAnimate, AnimateAnyone, and MikuDance may keep reference texture and are therefore omitted, but these are the state-of-the-art methods the abstract claims to outperform. The comparison is restricted to DrawingSpinUp and UniAnimate*, the latter being fine-tuned on the same synthesized dataset used to train the proposed method. This setup biases the comparison in favor of the proposed system, and it does not establish superiority over the pose-controllable diffusion baselines. The authors should either include those baselines with a protocol that controls for the texture-preservation issue, or clearly delimit the quantitative claim to the compared methods.
[Sections 4.5.3 and 4.5.4] The central contribution, Secondary Dynamics Injection, is validated only through selected visual frames in Figures 13 and 14. Since the paper's headline claim is about 'natural secondary dynamics,' the lack of any quantitative or statistically tested evaluation of dynamics is load-bearing. The claim in Section 6 that the authors 'observed for the first time that different denoising steps are closely associated with distinct types of motion' is likewise supported only by a qualitative inspection of decoded latents in Figure 7. The paper needs either a quantitative dynamics evaluation, a complete user-study report with protocol and statistics, or a substantial softening of these claims.
[Section 4.1 and Section 4.4] There is a same-source dependency that is not fully acknowledged or tested. The training set is synthesized by DrawingSpinUp, which is also one of the two quantitative baselines and a prior method by the same group; the evaluation set contains characters from external drawing datasets but the ground-truth videos are again generated by DrawingSpinUp. This makes the comparison against DrawingSpinUp partly a test of how well the model reproduces the outputs on which it was trained, and does not establish generalization to real hand-drawn animations or to other animation pipelines. The authors should clarify the exact overlap between training and evaluation sources and provide at least one evaluation on data or annotations independent of the training pipeline.
[Sections 3.2, 3.3, and 6] The system's robustness depends on the quality of the single-image 3D reconstruction and auto-rigging, but the paper provides no systematic analysis of failure rates or manual effort. The method requires manual hair-body segmentation maps for long-hair characters and relies on Mixamo auto-rigging, and Section 6 admits that back-view long-hair regions remain problematic (Fig. 18b). Given that HLM is claimed to address the main failure case of prior skeletal methods, the paper should quantify how often HLM succeeds, how much manual correction is needed, and how sensitive the final results are to reconstruction/rigging errors. Without this, the generality of the central claim is not established.
[Section 1 and Section 4.4] The Introduction states that 'the results of comprehensive experiments and a perceptual user study' demonstrate the system's performance, but no user study appears in the manuscript. There is no protocol, number of participants, stimuli, task description, or statistical analysis. Since this sentence is part of the paper's evidence for superiority, the authors must either include the full user study or remove the claim. A qualitative figure comparison alone is not a substitute for a reported user study.

minor comments (7)

[Section 3.2, Eq. (5)] The notation in Eq. (5) mixes set operations on segmentation maps (∪, ∩) with element-wise multiplication of implicit fields (⊙); please clarify how binary masks are converted to the spatial domain of the implicit fields and how the 'back' mask is derived.
[Section 3.2] The phrase 'manually-processed hair-body segmentation maps' should state explicitly that the user must provide these maps for long-hair characters and should indicate the typical effort involved, as this affects reproducibility.
[Section 3.5.3] The statement that 'when τ1 < 0.7, the impact on motion distribution remains minimal' appears to contradict the earlier observation that β affects runtime; the relationship between β, motion content, and runtime should be stated more precisely.
[Section 4.4] The metric descriptions should specify how per-frame LPIPS/FID/CLIP scores are aggregated across frames and clips, and why LPIPS is described as 'texture consistency' when it compares generated frames to the reference image.
[Figure 7 and Section 3.5.1] The axis labels in Figure 7 are unclear; please label the denoising timestep or percentage explicitly and state which model's latent estimates are being visualized.
[Section 4.2] The implementation details state that the model 'randomly decided whether to apply a mask' and also 'enhanced mask diversity'; please clarify the exact masking schedule and whether masks are applied to training videos consistently.
[References] The reference for Marching Cubes contains a typo ('Alghrithm'); please correct it.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity: the pipeline's output is not defined by its inputs, and the same-source DrawingSpinUp dependency is a benchmark caveat, not a derivation loop.

full rationale

The manuscript's chain is a conditioning pipeline rather than a formal derivation. Coarse guidance frames are rendered from the system's own rigged 3D proxy (Section 3.3); the domain-adapted diffusion model is initialized from UniAnimate and fine-tuned on DrawingSpinUp-generated stylized videos (Section 4.1: "We use the stylized animation videos generated by DrawingSpinUp [Zhou et al. 2024] as ground truth to fine-tune our diffusion models"); and the SDI step blends latents from the domain-adapted model and the pre-trained model using user-specified masks (Section 3.5, Eq. 6). None of these equations restates the final frames as an input by construction, and no fitted parameter is renamed as a prediction. The use of DrawingSpinUp as both training-ground-truth source and a quantitative baseline (Table 1) is a same-source benchmarking dependency that should be disclosed and may bias the comparison, but it does not make the central claim of natural secondary dynamics equivalent to the DrawingSpinUp output: the final video is additionally conditioned on the system's own coarse renders, the reference image, and the pre-trained UniAnimate prior, and the HLM and SDI ablations provide independent content. Therefore no circular step meets the required standard of Eq. X = Eq. Y by construction or a fitted input called a prediction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The system introduces no new physical entities. Its load-bearing assumptions are about transferability of motion priors, sufficiency of a small training set, and reliability of automatically generated 3D proxies. The two manually chosen hyperparameters alpha and beta are the only free parameters that materially affect the SDI behavior.

free parameters (2)

alpha (SDI start threshold) = 0.7-0.95
Chosen by hand based on ablations in Section 4.5.5; controls when latent blending with the pre-trained model begins.
beta (SDI end threshold) = 0.5-0.7
Chosen by hand based on ablations in Section 4.5.5; controls when latent blending ends and re-denoising begins.

assumptions (3)

domain assumption Pre-trained UniAnimate encodes human motion priors that can be injected into hand-drawn characters via latent blending.
Used in SDI (Section 3.5); the paper assumes these priors transfer across the domain gap.
domain assumption A small dataset of 428 clips from 124 characters is sufficient to adapt the diffusion model to hand-drawn styles without catastrophic forgetting.
Sections 3.4.2 and 4.1; the claim of domain adaptation rests on this scale.
domain assumption Automatically reconstructed low-poly 3D geometry with auto-rigging provides sufficiently accurate geometric guidance for the diffusion model.
Sections 3.2 and 3.3; if reconstruction or rigging fails, coarse guidance is wrong.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Rigging to Waving: 3D-Guided Diffusion for Natural Animation of Hand-Drawn Characters." pith.science (2026). https://pith.science/paper/JKGN3MYR

@misc{pith2026250906573,
  author       = {Pith},
  title        = {Pith review of: From Rigging to Waving: 3D-Guided Diffusion for Natural Animation of Hand-Drawn Characters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKGN3MYR}},
  note         = {Machine review of arXiv:2509.06573}
}

read the original abstract

Hand-drawn character animation is a vibrant field in computer graphics, presenting challenges in achieving geometric consistency while conveying expressive motion. Traditional skeletal animation methods maintain geometric consistency but struggle with complex non-rigid elements like flowing hair and skirts, leading to unnatural deformation. Conversely, video diffusion models synthesize realistic dynamics but often create geometric distortions in stylized drawings due to domain gaps. This work proposes a hybrid animation system that combines skeletal animation and video diffusion. Initially, coarse images are generated from characters retargeted with skeletal animations for geometric guidance. These images are then enhanced in texture and secondary dynamics using video diffusion priors, framing this enhancement as an inpainting task. A domain-adapted diffusion model refines user-masked regions needing improvement, especially for secondary dynamics. To enhance motion realism further, we introduce a Secondary Dynamics Injection (SDI) strategy in the denoising process, incorporating features from a pre-trained diffusion model enriched with human motion priors. Additionally, to tackle unnatural deformations from low-poly single-mesh character modeling, we present a Hair Layering Modeling (HLM) technique that uses segmentation maps to separate hair from the body, allowing for more natural animation of long-haired characters. Extensive experiments show that our system outperforms state-of-the-art methods in both quantitative and qualitative evaluations.

Figures

Figures reproduced from arXiv: 2509.06573 by the authors.

**Figure 1.** Our system produces natural character animations (Right) given single input drawings (Left) and target 3D motions (Middle). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Illustration of issues with existing methods: (a)-(b) unnatural defor [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** An illustration of our pipeline, which consists of three main parts. (a) Given a hand-drawn character image [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (13 more)

**Figure 4.** Figure 4: An illustration of hair layering modeling. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: An illustration of how we obtain the SDI mask sequence. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: An illustration of our domain-adapted diffusion model [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Gradually decreasing t-step denoised latent estimation results. The numbers below the images represent the percentage of the denosing process. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: An example illustrating Poisson Blending for inpainting the n-th [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Visual comparisons with two traditional skeletal animation methods. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

**Figure 10.** Figure 10: Visual comparisons with four diffusion-based animation methods. [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

**Figure 11.** Figure 11: Ablation on our hair layering modeling (HLM) method. [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]

**Figure 14.** Figure 14: Ablations on three key components of our Secondary Dynamics [PITH_FULL_IMAGE:figures/full_fig_p009_14.png]

**Figure 15.** Figure 15: The impact of different values of 𝛼 between 0.70 and 0.95 on the secondary dynamics of the animation results, with 𝛽 = 0.60. Reference Driving poses [PITH_FULL_IMAGE:figures/full_fig_p009_15.png]

**Figure 16.** Figure 16: The impact of different values of 𝛽 (0.50, 0.60, 0.70) on the secondary dynamics of the animation results, with 𝛼 = 0.95. , Vol. 1, No. 1, Article . Publication date: September 2025 [PITH_FULL_IMAGE:figures/full_fig_p009_16.png]

**Figure 18.** Figure 18: Two examples for the limitations of our method. [PITH_FULL_IMAGE:figures/full_fig_p010_18.png]

**Figure 17.** Figure 17: Edit propagation. The reference frame (left) is edited, and the change [PITH_FULL_IMAGE:figures/full_fig_p010_17.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 2 canonical work pages

[2017]

InProceedings of the 30th Annual ACM Symposium on User Interface Software and Technology

Secondary Motion for Performed 2D Animation. InProceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. 97–108. Tianyi Xie, Yiwei Zhao, Ying Jiang, and Chenfanfu Jiang. 2025. Physanimator: Physics- guided generative cartoon animation. InProceedings of the Computer Vision and Pattern Recognition Conference. 10793–10804. Jinbo...

arXiv 2024
[2018]

InProceedings of the Joint Symposium on Computational Aesthetics and Sketch-Based Interfaces and Modeling and Non-Photorealistic Animation and Rendering

ToonCap: A Layered Deformable Model for Capturing Poses from Cartoon Characters. InProceedings of the Joint Symposium on Computational Aesthetics and Sketch-Based Interfaces and Modeling and Non-Photorealistic Animation and Rendering. 1–12. Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a T...

work page 2017
[2022]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

High-Resolution Image Synthesis With Latent Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695. Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. 2025. Semantic Image Inversion and Editing using Rectified Stochas- tic Differential Equations. In...

work page 2025
[2023]

Jiaming Song, Chenlin Meng, and Stefano Ermon

A Method for Animating Children’s Drawings of the Human Figure.ACM Transactions on Graphics42, 3 (2023), 1–15. Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InInternational Conference on Learning Representations. Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, DanDan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong...

arXiv 2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [2017]

InProceedings of the 30th Annual ACM Symposium on User Interface Software and Technology

Secondary Motion for Performed 2D Animation. InProceedings of the 30th Annual ACM Symposium on User Interface Software and Technology. 97–108. Tianyi Xie, Yiwei Zhao, Ying Jiang, and Chenfanfu Jiang. 2025. Physanimator: Physics- guided generative cartoon animation. InProceedings of the Computer Vision and Pattern Recognition Conference. 10793–10804. Jinbo...

arXiv 2024

[2] [2018]

InProceedings of the Joint Symposium on Computational Aesthetics and Sketch-Based Interfaces and Modeling and Non-Photorealistic Animation and Rendering

ToonCap: A Layered Deformable Model for Capturing Poses from Cartoon Characters. InProceedings of the Joint Symposium on Computational Aesthetics and Sketch-Based Interfaces and Modeling and Non-Photorealistic Animation and Rendering. 1–12. Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a T...

work page 2017

[3] [2022]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

High-Resolution Image Synthesis With Latent Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695. Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. 2025. Semantic Image Inversion and Editing using Rectified Stochas- tic Differential Equations. In...

work page 2025

[4] [2023]

Jiaming Song, Chenlin Meng, and Stefano Ermon

A Method for Animating Children’s Drawings of the Human Figure.ACM Transactions on Graphics42, 3 (2023), 1–15. Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InInternational Conference on Learning Representations. Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, DanDan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong...

arXiv 2023