Pith. sign in

REVIEW 6 major objections 6 minor 45 references

AnimeColor: Reference-based Animation Colorization with Diffusion Transformers

T0 review · 6 major / 6 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 A diffusion-transformer framework with two color-control modules beats six existing methods on reference-based animation colorization.

desk verdict A competent, well-ablated DiT colorization system whose SOTA claim is plausible but needs data-curation and baseline-training details before it's bulletproof. read the letter →

arxiv 2507.20158 v1 pith:H7KM5GOK submitted 2025-07-27 cs.CV

Yuhong Zhang , Liyao Wang , Han Wang , Danni Wu , Zuzeng Lin , Feng Wang , Li Song This is my paper
classification cs.CV
keywords AnimationcolorizationDiffusionTransformersVideoReference-basedTemporalconsistencySketchguidanceColorcontrolMulti-stagetraining
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

AnimeColor is a reference-based animation colorization framework built on a Diffusion Transformer (DiT) video diffusion model. The paper's central claim is that combining sketch-sequence conditioning with two complementary color-extraction modules—a High-level Color Extractor for semantic color and a Low-level Color Guider for fine-grained detail—gives better color accuracy, sketch alignment, temporal consistency, and visual quality than existing U-Net-based diffusion colorizers and a plain DiT baseline. In the reported comparison over six methods, AnimeColor obtains the best PSNR (23.56), SSIM (0.8288), FID (4.87), and FVD (23.77), and a user study prefers it. If the gains survive on real production footage, the framework would be a practical way to automate a labor-intensive step of animation production.

What carries the argument

The framework uses three mechanisms working together. First, sketch latents are concatenated with the noised video latents as input to a pre-trained DiT (CogVideoX-2B), so the model treats sketches as part of its context. Second, a High-level Color Extractor (HCE) encodes the reference image with RADIO and condenses the features with a Q-Former into 32 learnable color tokens that guide the DiT through cross-attention. Third, a Low-level Color Guider (LCG), a trainable copy of the DiT, produces fine-grained reference-image features that are merged with the denoising stream's vision and text tokens via self-attention. A four-stage training procedure first learns sketch control, then HCE, then LCG, then jointly fine-tunes the denoiser to integrate both color signals.

What would settle it

Inspect the data-curation code or pipeline for the 100,000 training clips: if Open-Sora was used to generate rather than filter the footage, re-run AnimeColor and the strongest baselines on a test set of hand-drawn production frames; the central claim is refuted if the PSNR, FVD, and user-study margins shrink or reverse on that real animation test set.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that a DiT-based video diffusion model can be adapted for reference-based animation colorization by injecting sketch structure through latent concatenation and reference color through two attention-level control modules. The authors report that this design outperforms three U-Net-based reference colorization methods (AniDoc, ToonCrafter, LVCD), a ControlNet-plus-IP-Adapter workflow, and a DiT baseline that concatenates the reference image with the sketches, on color accuracy, sketch alignment, temporal consistency, and perceptual quality. Their ablation studies attribute the biggest color-accuracy gain to the Low-level Color Guider and show that the four-stage training schedule further improves temporal stability.

Load-bearing premise

The load-bearing premise is that the 100,000 training clips are real, representative anime footage; the paper says they were 'systematically curated using Open-Sora' but does not state whether Open-Sora generated or filtered them, so if they are synthetic, the reported gains over baselines may not transfer to hand-drawn production animation.

Editorial extensions

If this is right

  • Sketch conditioning by latent concatenation is better suited to DiT than ControlNet-style encoder-decoder adapters, needing fewer parameters and improving sketch alignment (the SA metric).
  • Separating geometric control (sketches) from color control (reference image) avoids the confusing artifacts produced when the reference image and sketches are concatenated together into the input noise.
  • The Low-level Color Guider contributes the largest measured gain: removing it drops PSNR from 23.56 to 16.35 and raises FVD from 23.77 to 61.96 in the ablation.
  • The fourth-stage joint fine-tuning improves temporal consistency and color stability over the same architecture without it, confirming that separate module training alone is not enough.

Reading between the lines

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

  • If the same HCE/LCG conditioning were applied to other DiT video models, it could generalize to sketch-guided video editing or in-betweening beyond anime colorization, since nothing in the design is anime-specific.
  • Because the LCG is a trainable copy of the entire DiT, the method roughly doubles the denoiser's parameters and inference cost; part of the reported quality gap over single-stream baselines may come from this extra capacity rather than from the guidance mechanism itself.
  • The fixed test set of 1,800 clips is drawn from the same curation pipeline as training, so the reported margins are in-distribution; a cross-dataset evaluation on independently sourced animation would test actual generalization.
  • The framework's ability to colorize natural-image references suggests the color-reference modules encode palette and lighting rather than object identity, which could make the approach useful for stylistic transfer across domains.
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

6 major / 6 minor

Summary. The paper proposes AnimeColor, a reference-based animation colorization framework built on a Diffusion Transformer (DiT) backbone (CogVideoX-2B). Sketch latents are concatenated with the noised video latent as the conditioning input, a High-level Color Extractor (HCE) using RADIO plus a Q-Former supplies semantic color tokens via cross-attention, and a Low-level Color Guider (LCG), implemented as a trainable copy of the DiT, contributes fine-grained color tokens through self-attention. A four-stage training strategy first trains sketch-guided generation, then HCE, then LCG, and finally fine-tunes the denoising DiT jointly. The authors report quantitative comparisons (PSNR, SSIM, LPIPS, SA, FID, FVD) against AniDoc, ToonCrafter, LVCD, a ControlNet+IP-Adapter workflow, and a self-trained Concat baseline, along with ablations and a user study, and claim state-of-the-art performance and industrial applicability.

Significance. If the empirical claims are sustained, the paper makes a useful systems contribution: it demonstrates that DiT-based video diffusion with in-context sketch conditioning can be adapted to reference-based animation colorization, and that separating high-level semantic color control from low-level fine-grained color control, together with stage-wise training, improves color accuracy and temporal consistency. The paper contains multiple ablations, a user study, and consistent gains on most metrics over several baselines, which are genuine strengths. However, the load-bearing evaluation protocol is not fully specified: the role of Open-Sora in dataset curation is ambiguous, baseline training conditions are undocumented, the test protocol uses the first frame of each clip as the reference (making the task a self-reconstruction), and no significance testing or error bars are provided. These issues make the headline state-of-the-art claim provisional.

major comments (6)
  1. [Section 4.1] The description 'systematically curate them using Open-Sora' is ambiguous, and since Open-Sora is a video generation model rather than a curation tool, it likely means the 100,000 training clips were either generated or heavily processed by a generative model. If the training data are synthetic, the reported gains over baselines trained on real animation may not transfer to production anime, directly undermining the abstract's 'practical solution for industrial applications' claim. Please state explicitly whether Open-Sora generated or filtered the clips, describe the data source and any manual filtering, and discuss the implications for real-world deployment.
  2. [Section 4.2, Table 1] Only the Concat baseline is described as trained by the authors; for AniDoc, ToonCrafter, LVCD, and CN+IPA, the paper does not report whether they were fine-tuned on the same training distribution, which checkpoints were used, or the exact inference settings. Without this information, the numeric gaps in Table 1 (e.g., FVD 23.77 vs. 71.63 for LVCD) could reflect base-model capacity or domain shift rather than the proposed HCE/LCG/sketch-concatenation design. For a fair state-of-the-art comparison, the authors should document each baseline's weights, adaptation procedure, and evaluation conditions, and ideally retrain or fine-tune the baselines under matched conditions.
  3. [Section 4.1, test protocol] The test set uses the first frame of each video clip as the reference image, so the generated frames are compared against the remaining original frames of the same clip. This makes the quantitative evaluation a self-reconstruction task and does not measure the more practical setting of colorizing a new sketch with an independent reference image. This protocol may inflate similarity metrics and limits the external validity of the claims. Please add an evaluation where the reference image is not drawn from the target clip, or clearly state and justify the reconstruction-only setup.
  4. [Section 3.3, Eq. (4)] The text states that LCG concatenates low-level reference tokens with vision and text tokens for self-attention, but Eq. (4) shows Self-Attention(Concat(F_text, F_vision, F_H)), where F_H is the high-level reference token from HCE, not low-level tokens. The mechanism by which the trainable DiT copy extracts fine-grained color representations from the reference image is not specified, and it is unclear how Stage 3 can train LCG while the denoising DiT is frozen. Please provide a precise forward-pass description of LCG, including the role of the reference image latent and the exact token sequence entering self-attention.
  5. [Table 3, ablation on sketch conditional model] In Table 3, the proposed sketch conditional model ('Ours') reports FVD 163.78, which is substantially worse than both ControlNet Variation 1 (106.93) and ControlNet Variation 2 (136.11), yet the text claims the proposed method 'surpasses other methods on nearly all metrics, particularly on the SA metric.' The FVD regression is not discussed. Either correct the claim or provide an explanation for why the proposed design is preferable despite this metric.
  6. [Tables 1-4] No error bars, confidence intervals, or significance tests are reported anywhere. Several headline differences are small relative to the scale of the metrics—for instance, LPIPS is 0.2018 vs. 0.2004 for LVCD in Table 1, and in Table 4 the RADIO-based HCE is worse on PSNR, SSIM, and SA while better on LPIPS, FID, and FVD. Without multiple runs or statistical testing, the paper cannot support fine-grained superiority claims such as 'slightly higher' or 'significantly outperforms.' Please report variance across seeds or runs and, where possible, appropriate significance measures.
minor comments (6)
  1. [Section 4.2.2] The sketch-alignment metric SA is described as 'RMSE between the sketches of the generated frames and the original input sketches,' but the paper does not specify how sketches are obtained from generated frames or whether the comparison is performed in the sketch-extraction space. Please define the extraction method and the exact computation.
  2. [Section 4.2.3] The user study says participants compare 'five different methods,' but Table 1 lists six methods including Concat. Please clarify which methods were included in the user study and why Concat was excluded, if it was.
  3. [Table 1] In the abstract and Section 4.2.2, the paper claims superiority on 'most metrics,' but LPIPS for the proposed method (0.2018) is slightly worse than LVCD (0.2004). Please acknowledge this explicitly and discuss whether the difference is meaningful.
  4. [Section 3.2, Eqs. (1)-(2)] The notation is inconsistent: Eq. (1) defines F_sum and F_spa from RADIO, but Eq. (2) uses F_RADIO without defining how it is obtained from F_sum and F_spa. Please clarify the aggregation through the MLP layers and define all symbols used in the equations.
  5. [Section 4.1] The implementation details state that 'CogvideoX-2B' is used as the baseline, but the model name is typically written as CogVideoX-2B; please standardize the capitalization. Also, the details of sketch-latent concatenation (channel-wise concatenation, dimension matching, and handling of varying resolutions) are not given and would help reproducibility.
  6. [Throughout] There are several minor typos and inconsistencies in method names, such as 'Tooncrafter' vs. 'ToonCrafter' and 'Anidoc' vs. 'AniDoc,' and the phrase 'Finetune' appears in Table 2 as 'FT' without a formal definition in the table caption. Please proofread and harmonize the terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper derives its SOTA claim from held-out quantitative comparisons, internal ablations, and a user study, none of which reduce to the method's own definitions.

full rationale

AnimeColor is an empirical systems paper. The claimed contributions (HCE, LCG, sketch concatenation, multi-stage training) are evaluated against external baselines on a held-out set of 1,800 clips, and the ablations remove modules and retrain, producing independent internal comparisons. The reference image being the first frame of each test clip is a standard task formulation for reference-based colorization and does not make the prediction a fitted quantity, since the model is not optimized on the test clips. The description of the 100,000-clip training set as 'systematically curate[d] using Open-Sora' (Section 4.1) is ambiguous about whether Open-Sora generated or filtered the data, and the paper does not state whether AniDoc, ToonCrafter, and LVCD were fine-tuned on the same distribution; however, these are concerns about benchmarking fairness and generalizability, not circularity. No equation in the paper is defined in terms of the quantities it is used to predict, no fitted parameter is renamed as a prediction, and no load-bearing claim rests on a self-citation. The single citation of the authors' own prior work ([34] LayerAnimate) appears only in related work and is not used to justify the central result. The derivation chain is therefore self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard diffusion training assumptions, domain assumptions about the dataset and reference image sufficiency, and hand-chosen hyperparameters. No new physical entities are introduced; HCE and LCG are trainable modules whose contributions are evaluated via ablations.

free parameters (4)
  • Number of learnable color tokens N in Q-Former = 32
    Set by hand in Sec 4.1; controls the capacity of high-level color extraction.
  • Training iterations per stage = 50K, 25K, 30K, 50K
    Hand-chosen schedule in Sec 4.1; balances convergence across the four stages.
  • Learning rate = 2e-5
    AdamW learning rate chosen ad hoc in Sec 4.1.
  • Batch size = 16
    Fixed across all stages in Sec 4.1.
assumptions (6)
  • standard math Standard diffusion model epsilon-prediction objective, as in Eq. 5-8
    Assumed background; the loss functions in Sec 3.4 follow the standard DDPM formulation.
  • domain assumption The CogVideoX VAE latent space preserves enough structural and color information for colorization
    The pipeline encodes sketches and reference images into latents; if the VAE discards color details, the method's ceiling is limited.
  • domain assumption XDoG-extracted sketches are a faithful proxy for animation line art
    Used for both training conditioning and the SA metric in Sec 4.1.
  • domain assumption The 100,000-clip training set is representative of real animation, with Open-Sora used for curation rather than generation
    Sec 4.1 cites Open-Sora without explaining its role; if the clips are synthetic, generalization to production anime is unverified.
  • domain assumption A single reference frame provides sufficient color information for the entire video
    Training and testing use one reference image per clip (Sec 4.1); the paper acknowledges new objects can appear, requiring extrapolation.
  • ad hoc to paper The four-stage training strategy effectively decouples geometric and color controls
    Supported only by the ablations in Sec 4.3.1; no theoretical justification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AnimeColor: Reference-based Animation Colorization with Diffusion Transformers." pith.science (2026). https://pith.science/paper/H7KM5GOK

@misc{pith2026250720158,
  author       = {Pith},
  title        = {Pith review of: AnimeColor: Reference-based Animation Colorization with Diffusion Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7KM5GOK}},
  note         = {Machine review of arXiv:2507.20158}
}
read the original abstract

Animation colorization plays a vital role in animation production, yet existing methods struggle to achieve color accuracy and temporal consistency. To address these challenges, we propose \textbf{AnimeColor}, a novel reference-based animation colorization framework leveraging Diffusion Transformers (DiT). Our approach integrates sketch sequences into a DiT-based video diffusion model, enabling sketch-controlled animation generation. We introduce two key components: a High-level Color Extractor (HCE) to capture semantic color information and a Low-level Color Guider (LCG) to extract fine-grained color details from reference images. These components work synergistically to guide the video diffusion process. Additionally, we employ a multi-stage training strategy to maximize the utilization of reference image color information. Extensive experiments demonstrate that AnimeColor outperforms existing methods in color accuracy, sketch alignment, temporal consistency, and visual quality. Our framework not only advances the state of the art in animation colorization but also provides a practical solution for industrial applications. The code will be made publicly available at \href{https://github.com/IamCreateAI/AnimeColor}{https://github.com/IamCreateAI/AnimeColor}.

Figures

Figures reproduced from arXiv: 2507.20158 by the authors.

Figure 1
Figure 1. Given a reference image and a sequence of sketches, our method generates high-quality colorized animation videos. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the difference between our proposed [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The framework of our proposed AnimeColor. We first concatenate sketch latents with noise as the input for the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of reference-based coloriza [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Temporal profile comparison of different animation [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: The results of user study. Our AnimeColor is pre [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison about ablation study on [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison about ablation study on [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison about ablation study on [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 11
Figure 11. Figure 11: Impact of different lineart extraction methods. g [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Colorization performance with natural images as [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 23 canonical work pages

  1. [1]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. 2023. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)

    arXiv 2023

  2. [2]

    Yu Cao, Xiangqiao Meng, PY Mok, Tong-Yee Lee, Xueting Liu, and Ping Li. 2024. AnimeDiffusion: anime diffusion colorization. IEEE Transactions on Visualization and Computer Graphics (2024)

    work page 2024

  3. [3]

    Caroline Chan, Fredo Durand, and Phillip Isola. 2022. Learning to generate line drawings that convey geometry and semantics. (2022)

    work page 2022

  4. [4]

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. 2023. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023)

    arXiv 2023

  5. [5]

    Shu-Yu Chen, Jia-Qi Zhang, Lin Gao, Yue He, Shihong Xia, Min Shi, and Fang-Lue Zhang. 2020. Active colorization for cartoon line drawings. IEEE Transactions on Visualization and Computer Graphics 28, 2 (2020), 1198–1208

    work page 2020

  6. [6]

    Yuekun Dai, Shangchen Zhou, Qinyue Li, Chongyi Li, and Chen Change Loy

    work page

  7. [7]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

    2021

  8. [8]

    Zhi Dou, Ning Wang, Baopu Li, Zhihui Wang, Haojie Li, and Bin Liu. 2021. Dual color space guided sketch colorization. IEEE Transactions on Image Processing 30 (2021), 7292–7304

    work page 2021

Show all 45 references
  1. [9]

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. 2023. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)

    2023 arXiv

  2. [10]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  3. [11]

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. 2024. In-context lora for diffusion transformers. arXiv preprint arXiv:2410.23775 (2024)

    2024 arXiv

  4. [12]

    Zhitong Huang, Mohan Zhang, and Jing Liao. 2024. LVCD: reference-based lineart video colorization with diffusion models. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–11

    2024

  5. [13]

    Hyunsu Kim, Ho Young Jhoo, Eunhyeok Park, and Sungjoo Yoo. 2019. Tag2pix: Line art colorization using text tag with secat and changing loss. In Proceedings of the IEEE/CVF international conference on computer vision . 9056–9065

    2019

  6. [14]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  7. [15]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

    2023

  8. [16]

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. 2024. Hunyuan- DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:240...

    2024 arXiv

  9. [17]

    Yihao Meng, Hao Ouyang, Hanlin Wang, Qiuyu Wang, Wen Wang, Ka Leong Cheng, Zhiheng Liu, Yujun Shen, and Huamin Qu. 2024. AniDoc: Animation Creation Made Easier. arXiv preprint arXiv:2412.14173 (2024)

    2024 arXiv

  10. [18]

    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

  11. [19]

    Alibaba pai. 2024. CogVideoX-Fun. https://github.com/aigc-apps/CogVideoX-Fun

    2024

  12. [20]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  13. [21]

    Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. 2024. AM- RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 12490–12500

    2024

  14. [22]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  15. [23]

    Min Shi, Jia-Qi Zhang, Shu-Yu Chen, Lin Gao, Yu-Kun Lai, and Fang-Lue Zhang

  16. [24]

    Harrish Thasarathan, Kamyar Nazeri, and Mehran Ebrahimi. 2019. Automatic temporally coherent video colorization. In 2019 16th conference on computer and robot vision (CRV). IEEE, 189–194

    2019

  17. [25]

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. 2019. FVD: A new metric for video genera- tion. ICLR 2019 Workshop DeepGenStruct (2019)

    2019

  18. [26]

    Ashish Vaswani. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762 (2017)

    2017 arXiv

  19. [27]

    Ning Wang, Muyao Niu, Zhi Dou, Zhihui Wang, Zhiyong Wang, Zhaoyan Ming, Bin Liu, and Haojie Li. 2023. Coloring anime line art videos with transformation region enhancement network. Pattern Recognition 141 (2023), 109562

    2023

  20. [28]

    Holger Winnemöller, Jan Eric Kyprianidis, and Sven C Olsen. 2012. XDoG: An eXtended difference-of-Gaussians compendium including advanced image stylization. Computers & Graphics 36, 6 (2012), 740–753

    2012

  21. [29]

    Xiaoyu Xiang, Ding Liu, Xiao Yang, Yiheng Zhu, and Xiaohui Shen. 2021. Anime2Sketch: A Sketch Extractor for Anime Arts with Deep Networks. https: //github.com/Mukosame/Anime2Sketch

    2021

  22. [30]

    Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. 2024. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–11

    2024

  23. [31]

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. 2024. EasyAnimate: A High-Performance Long Video Gener- ation Method based on Transformer Architecture.arXiv preprint arXiv:2405.18991 (2024)

    2024

  24. [32]

    Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. 2024. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994 (2024)

    2024 arXiv

  25. [33]

    Dingkun Yan, Liang Yuan, Erwin Wu, Yuma Nishioka, Issei Fujishiro, and Suguru Saito. 2024. Colorizediffusion: Adjustable sketch colorization with reference image and text. arXiv preprint arXiv:2401.01456 (2024)

    2024 arXiv

  26. [34]

    Yuxue Yang, Lue Fan, Zuzen Lin, Feng Wang, and Zhaoxiang Zhang. 2025. Lay- erAnimate: Layer-specific Control for Animation. arXiv preprint arXiv:2501.08295 (2025)

    2025 arXiv

  27. [35]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072 (2024)

    2024 arXiv

  28. [36]

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. 2023. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)

    2023 arXiv

  29. [37]

    Yifeng Yu, Jiangbo Qian, Chong Wang, Yihong Dong, and Baisong Liu. 2024. Animation line art colorization based on the optical flow method. Computer Animation and Virtual Worlds 35, 1 (2024), e2229

    2024

  30. [38]

    Lvmin Zhang, Chengze Li, Tien-Tsin Wong, Yi Ji, and Chunping Liu. 2018. Two- stage sketch colorization. ACM Transactions on Graphics (TOG) 37, 6 (2018), 1–14

    2018

  31. [39]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision . 3836–3847

    2023

  32. [40]

    Qian Zhang, Bo Wang, Wei Wen, Hai Li, and Junhui Liu. 2021. Line art correlation matching feature transfer network for automatic animation colorization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 3872–3881

    2021

  33. [41]

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. 2024. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404 (2024)

    2024 arXiv

  34. [42]

    Changqing Zou, Haoran Mo, Chengying Gao, Ruofei Du, and Hongbo Fu. 2019. Language-based colorization of scene sketches. ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–16

    2019

  35. [2022]

    IEEE Transactions on Visualization and Computer Graphics 29, 6 (2022), 2965–2979

    Reference-based deep line art video colorization. IEEE Transactions on Visualization and Computer Graphics 29, 6 (2022), 2965–2979

    2022

  36. [2023]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015–4026

  37. [2024]

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

    Learning Inclusion Matching for Animation Paint Bucket Colorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 25544–25553

Pith tools

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