Pith. sign in

REVIEW 4 major objections 5 minor 50 references

LongAnimation: Long Animation Generation with Dynamic Global-Local Memory

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 LongAnimation claims that dynamic global-local memory keeps one keyframe's colors consistent across 500-frame animations.

desk verdict A genuinely new way to keep color stable over ~500 frames, with a mechanism that is plausible but not yet fully separated from generic feature matching. read the letter →

arxiv 2507.01945 v2 pith:EVJH6BHS submitted 2025-07-02 cs.CV

Nan Chen , Mengqi Huang , Yihao Meng , Zhendong Mao This is my paper
classification cs.CV
keywords animationcolorizationlongvideogenerationcolorconsistencydiffusiontransformerdynamicglobal-localmemoryKVcachereference-basedcoloringtext-guided
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

Long animation colorization in industry means filling colors into hundreds of sketches from a single keyframe, and prior systems drift because they only glue adjacent segments together. LongAnimation argues instead for a dynamic global-local paradigm: as each new segment is generated, all previously generated frames are compressed into a global memory and queried for color information relevant to the current sketch. The framework combines a DiT video generator, a SketchDiT that fuses reference image, sketch, and text conditions, a Dynamic Global-Local Memory module that retrieves historical color context, and a Color Consistency Reward that trains the model to match low-frequency color features of reference animations. On open-domain animations averaging 500 frames, it reports better LPIPS, SSIM, PSNR, FVD, and FID than prior methods, with the largest gains on the long-video metrics. If the claim holds, automated coloring could cover roughly five times more frames per keyframe before colors start to wander.

What carries the argument

The carrying mechanism is the Dynamic Global-Local Memory (DGLM). Historical frames are compressed by a long-video-understanding model into visual summary tokens, and the keys and values of those tokens from several middle layers are stored in a KV cache; a cross-attention layer uses the current hybrid reference features as the query to retrieve global and local color context. Middle layers are chosen because the paper cites evidence that they encode global visual properties better than the final layer. Two supporting mechanisms complete the argument: the Color Consistency Reward, which treats the same KV-cache features as a low-frequency color descriptor and trains the generator to match reference features without backpropagating through the 3D VAE, and color consistency fusion, which blends overlapping latents only in the late denoising stage to avoid brightness artifacts. Together they implement the dynamic global-local paradigm that the paper argues is necessary for long-term consistency.

What would settle it

Take one fixed sketch and reference image and generate the same segment twice, once with a global history whose colors match the reference and once with the same motion but different colors. If the DGLM output does not track the history's colors, or if the Color Consistency Reward cannot rank the matching history above the mismatched one, the central color-memory claim would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that long-term color consistency can be enforced by dynamically retrieving global color features from the generation history, rather than by fixing one reference image or fusing adjacent-segment overlaps. LongAnimation feeds all past frames and the most recent segment into a long-video-understanding model, reads the KV-cache keys and values of its visual summary tokens from several middle layers, and cross-attends those historical keys and values with the current sketch-and-reference query features. The Color Consistency Reward then aligns the same KV-cache features of the generated video with those of the reference video using a non-gradient reward objective, and a late-stage latent fusion smooths segment boundaries without perturbing early layout features. The reported result is stable color over roughly 500 frames, about five times longer than the previous methods it compares against.

Load-bearing premise

The whole method rests on the assumption that the KV-cache features taken from a few middle layers of the video-understanding model describe color as a low-frequency property that can be compared across generated, reference, and historical videos without being tangled up with motion or object identity.

Editorial extensions

If this is right

  • The effective colorization length rises from the roughly 100-frame ceiling of prior methods to about 500 frames from a single keyframe, with the largest measured gains on long-term metrics.
  • Because latent fusion is confined to the late denoising stage, segment transitions no longer corrupt the layout or brightness of non-overlapping frames.
  • Text and reference image jointly guide generation, so regions not determined by the sketch, such as backgrounds, can be steered by language while the character color stays anchored.
  • The frequency analysis implies that low-frequency color information is preserved over time, while sketch-controlled high-frequency detail is maintained by all compared methods.
  • The memory is re-compressed and re-queried for each new segment, so the approach should extend beyond 500 frames without requiring the full history to be re-processed.

Reading between the lines

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

  • Beyond the paper: the same KV-cache feature alignment could serve as a general style-consistency reward for other long-horizon generation tasks, such as keeping a character's outfit, lighting, or palette stable across shots.
  • A length-ablation curve would test an implicit prediction of the paper's logic: the DGLM advantage should grow with the number of segments and shrink toward zero for very short videos.
  • Comparing the Color Consistency Reward against a cheap histogram- or statistics-based color regularizer would reveal whether the LVU features themselves, or merely any color-matching signal, are what drive the measured gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper addresses long animation colorization (video colorization conditioned on a reference keyframe and sketches) for sequences of roughly 500 frames. The authors propose LongAnimation, built on CogVideoX-1.5-5B, with three main components: SketchDiT, a lightweight DiT branch that extracts hybrid reference features from reference image, sketches, and text; a Dynamic Global-Local Memory (DGLM) that uses Video-XL's KV caches to compress historical and local video segments and cross-attends these features with the current reference features; and a Color Consistency Reward (CCR) that aligns the KV-cache features of generated and reference videos via a non-gradient reward. At inference, a color consistency fusion (CCF) blends overlapping latents only in the late denoising stage. The method is trained on a filtered subset of Sakuga-42M and evaluated on short-term (14 frames) and long-term (~500 frames) tests, reporting large gains over UNet-based baselines on LPIPS, SSIM, PSNR, FVD, and FID.

Significance. The paper addresses a practically important and under-explored problem. The dynamic global-local memory design, with LVU-based compression of history, is conceptually novel for colorization, and the inference-time late-fusion idea is well motivated by the observed degradation in Table 3. The ablations (Table 2) isolate the contribution of each component, and the authors provide a code link. If the color-consistency claim is confirmed by direct color metrics, the reported improvements would be substantial and useful to the community. However, the current evidence relies on holistic video-quality metrics and an unverified feature-interpretability assumption, and the reward objective as written has a sign inconsistency; these need to be resolved before the central claim can be accepted.

major comments (4)
  1. [Sec. 3.2.3, Eqs. (2) and (5)] The reward function in Eq. (5) is defined as a squared L2 distance between generated and reference KV caches. Since Eq. (2) is written as the gradient of an expected reward to be maximized, maximizing r as defined would increase the KV-cache mismatch rather than reduce it. If the implementation actually uses a negative distance or a baseline-corrected surrogate, that needs to be stated explicitly. This is load-bearing because the CCR contribution (Table 2, ID-4) depends on this objective.
  2. [Sec. 3.2.2/3.2.3, Tables 2 and Fig. 8] The central claim is long-term color consistency, but the supporting evidence does not directly measure color consistency. The metrics in Tables 1-3 (FVD, LPIPS, PSNR, SSIM) are holistic, and the low-frequency PSNR decay analysis in Fig. 8 is only a proxy; low-frequency image content is not the same as color appearance. The premise that Video-XL middle-layer KV caches encode color rather than layout, motion, or object identity is not validated. Please add direct color-consistency evaluations, for example per-object or per-region color histograms over time, and/or an analysis showing that the M-layer features are sensitive to color perturbations while being robust to motion and structure changes. Without this, the gains from DGLM and CCR could be attributed to generic feature alignment rather than to color consistency.
  3. [Sec. 4.3, Table 3] The fusion start timestep tst is an inference hyperparameter selected on the same long-term test set used to produce the final reported numbers in Table 1. The choice tst=20 is justified by comparing FVD/LPIPS on that test set against tst=0, 40, and 50, which makes the headline improvements partially the result of test-set selection. Please use a validation split for hyperparameter selection, or report results across multiple seeds and random test subsets. In addition, no error bars or significance tests are provided for any of the quantitative comparisons in Tables 1-3, so the reliability of the reported margins is unclear.
  4. [Sec. 4.1, Tables 1 and 2] The comparison against prior state of the art is confounded by backbone capacity and training data. LongAnimation is a 5B DiT model finetuned on Sakuga-42M, while the baselines are UNet-based models not trained on this dataset, and LVCD* additionally uses grayscale sketches. Notably, the SketchDiT-only ablation (ID-0 in Table 2) already achieves a long-term FVD of 321.62, which is much better than LVCD*'s 473.02 in Table 1. Thus the headline 49.1% long-term FVD improvement is not solely attributable to the proposed DGLM and CCR components. Please add a DiT-based baseline trained on the same data, or clearly separate the gains from the backbone and the proposed modules.
minor comments (5)
  1. [Throughout] There are several typos: 'donates' should be 'denotes' in Eq. (5), 'gril' should be 'girl' in the Fig. 4 caption, 'hybird' should be 'hybrid' in Sec. 3.2.2, 'Artifcial' should be 'Artificial' in the acknowledgements, 'Binaryzation' should be 'Binarization', and 'Lon-gAnimation' in Sec. 4.2 should be 'LongAnimation'.
  2. [Sec. 3.2.2] The sentence 'After estimating the visual feature changes between adjacent frames using CLIP [29, 39]' cites reference [39] (VideoLLaMB), which is not a CLIP work; please correct the citation.
  3. [Sec. 3.3, Eq. (6)] The notation in Eq. (6) is hard to follow: the subscript on z_{F-C+k}^{(i-1)t} is not defined clearly, and the range of k and the role of the '+1' in C+1 should be spelled out so that the overlap weighting is unambiguous.
  4. [Sec. 4.4, Fig. 8] The definition of the 'PSNR relative decay ratio' and the low-frequency/high-frequency filter cutoff are not given; please specify how the frequency split is computed and how the 'sub-optimal method' is selected for the 8.2% comparison.
  5. [Sec. 4.1] In the implementation paragraph, 'GLM is trained for 10k steps' appears to refer to the Dynamic Global-Local Memory module, but the abbreviation GLM is used nowhere else; please unify the terminology (DGLM or GLM).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method's components are trained and evaluated against external metrics; the feature-space assumption is empirical, not definitional.

full rationale

The paper's central claims are validated with external metrics (FVD, FID, LPIPS, SSIM, PSNR) computed against ground-truth frames, not with quantities that appear in the training objective or inference conditioning. The Color Consistency Reward in Eq. (5) uses a fixed, pretrained Video-XL KV-cache feature space as a learning signal, and the DGLM cross-attention in Eq. (4) conditions on the same feature space for memory retrieval; however, this alignment is an architectural and training choice, not a prediction that reduces to its inputs. No fitted parameter is relabeled as a result, and no equation is shown to be equivalent to another by construction. The paper's self-citations ([13], [19], [36]) appear only in related-work discussions of other video-generation tasks and are not load-bearing for the long-animation colorization claim. The assumption that Video-XL middle-layer KV features encode low-frequency color appearance is an empirical, falsifiable premise, and Section 4.4 provides an external frequency-domain analysis as indirect evidence. Even if that premise were weak, weakness of an assumption is a correctness or robustness concern, not circularity. Therefore, no circular step is present.

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

The framework introduces no new physical or ontological entities. The dynamic global-local memory and color consistency reward are architectural and objective components, not newly postulated entities with independent falsifiable handles. The free parameters listed are inference or preprocessing choices, none of which are part of a derivation that assumes the target result.

free parameters (5)
  • tst (fusion start timestep) = 20
    Inference hyperparameter selected via ablation on the long-term test set (Section 4.3, Table 3); the final reported numbers use this value.
  • gamma (weight factor in Eq. 3) = not specified
    Weight factor for injecting SketchDiT features; value not reported in the main text, chosen by hand or learned.
  • M (LVU KV-cache layer count) = not specified
    Number of middle layers from which global and local memory keys and values are extracted; not specified in the main text, affects memory capacity.
  • CLIP similarity threshold for dynamic segment selection = not specified
    Threshold used to dynamically select 2, 4, or 8 frame segments in DGLM; not quantified in the paper.
  • Binarization threshold = 200
    Pixel values above 200 are set to 0 and all others to 1; a preprocessing choice that influences all methods using binary sketches.
assumptions (5)
  • domain assumption Middle layers of multimodal LLMs capture global visual features more reliably than final layers.
    Cited [4,33] and used to decide which LVU layers contribute to DGLM memory (Section 3.2.2).
  • domain assumption Self-attention in transformers acts as a low-pass filter, so low-frequency features such as color are better captured in KV cache features.
    Cited [32,35] and used to justify the Color Consistency Reward in Section 3.2.3.
  • domain assumption CogVideoX-1.5-5B provides a strong pretrained DiT video generation backbone that can be frozen and conditioned on external features.
    The entire framework builds on CogVideoX, freezing it during training (Sections 3.1 and 3.2.1).
  • domain assumption Video-XL is a suitable pretrained long-video-understanding model whose KV cache features are useful for both memory extraction and reward computation.
    DGLM and CCR rely on Video-XL features (Sections 3.2.2 and 3.2.3).
  • domain assumption The filtered Sakuga-42M subset (about 80k clips longer than 91 frames) is representative of open-domain animation for training.
    Training data is a filtered subset of Sakuga-42M (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of LongAnimation: Long Animation Generation with Dynamic Global-Local Memory." pith.science (2026). https://pith.science/paper/EVJH6BHS

@misc{pith2026250701945,
  author       = {Pith},
  title        = {Pith review of: LongAnimation: Long Animation Generation with Dynamic Global-Local Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EVJH6BHS}},
  note         = {Machine review of arXiv:2507.01945}
}
read the original abstract

Animation colorization is a crucial part of real animation industry production. Long animation colorization has high labor costs. Therefore, automated long animation colorization based on the video generation model has significant research value. Existing studies are limited to short-term colorization. These studies adopt a local paradigm, fusing overlapping features to achieve smooth transitions between local segments. However, the local paradigm neglects global information, failing to maintain long-term color consistency. In this study, we argue that ideal long-term color consistency can be achieved through a dynamic global-local paradigm, i.e., dynamically extracting global color-consistent features relevant to the current generation. Specifically, we propose LongAnimation, a novel framework, which mainly includes a SketchDiT, a Dynamic Global-Local Memory (DGLM), and a Color Consistency Reward. The SketchDiT captures hybrid reference features to support the DGLM module. The DGLM module employs a long video understanding model to dynamically compress global historical features and adaptively fuse them with the current generation features. To refine the color consistency, we introduce a Color Consistency Reward. During inference, we propose a color consistency fusion to smooth the video segment transition. Extensive experiments on both short-term (14 frames) and long-term (average 500 frames) animations show the effectiveness of LongAnimation in maintaining short-term and long-term color consistency for open-domain animation colorization task. The code can be found at https://cn-makers.github.io/long_animation_web/.

Figures

Figures reproduced from arXiv: 2507.01945 by the authors.

Figure 1
Figure 1. Comparison with existing paradigm. (a) Existing stud [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the LongAnimation. (a) During training, the reference information is fed into the CogvideoX [46] and SketchDiT, respectively, for efficient extraction of hybrid reference features. These reference features are then fused with the historical information in Dynamic Global-Local Memory (DGLM) for consistency generation. (b) For the first segment generation, the reference features are fed into SketchDiT and … view at source ↗
Figure 3
Figure 3. Overview of color consistency fusion during inference. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with existing methods. LongAnimation achieves long-term color consistency through Dynamic Global￾Local Memory (e.g., the gril’s dress and leaves). In contrast, previous methods exhibit unstable color changes. We highly recommend watching the vide…
Figure 5
Figure 5. Figure 5: Text and reference image jointly control background [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: (a) PSNR with existing methods. LongAnimation out￾performs previous methods in PSNR metric. (b) PSNR relative decay ratio in the low-frequency domain. Our method exhibits the least attenuation in low-frequency information (e.g., color), indi￾cating that our proposed Dy…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 14 canonical work pages

  1. [1]

    Training diffusion models with reinforce- ment learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforce- ment learning. arXiv preprint arXiv:2305.13301, 2023. 3, 5

    arXiv 2023

  2. [2]

    Ditctrl: Exploring attention control in multi-modal dif- fusion transformer for tuning-free multi-prompt longer video generation

    Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. Ditctrl: Exploring attention control in multi-modal dif- fusion transformer for tuning-free multi-prompt longer video generation. arXiv preprint arXiv:2412.18597, 2024. 2, 5

    arXiv 2024

  3. [3]

    Learning to generate line drawings that convey geometry and semantics

    Caroline Chan, Fr ´edo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915–7925, 2022. 7

    work page 2022

  4. [4]

    Explainable and interpretable multimodal large language models: A comprehensive survey

    Yunkai Dang, Kaichen Huang, Jiahao Huo, Yibo Yan, Sirui Huang, Dongrui Liu, Mengxi Gao, Jie Zhang, Chen Qian, Kun Wang, et al. Explainable and interpretable multimodal large language models: A comprehensive survey. arXiv preprint arXiv:2412.02104, 2024. 5

    arXiv 2024

  5. [5]

    Re- inforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Re- inforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Sys- tems, 36, 2024. 3, 5

    work page 2024

  6. [6]

    Video-ccam: Enhancing video-language un- derstanding with causal cross-attention masks for short and long videos

    Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language un- derstanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 3

    arXiv 2024

  7. [7]

    Ltx-video: Realtime video latent diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103 ,

    arXiv

  8. [8]

    Streamingt2v: Con- sistent, dynamic, and extendable long video generation from text

    Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Con- sistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 2

    arXiv 2024

Show all 50 references
  1. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 7

    2017

  2. [10]

    Cogvlm2: Visual language mod- els for image and video understanding

    Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Jun- hui Ji, Zhao Xue, et al. Cogvlm2: Visual language mod- els for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 3

    2024 arXiv

  3. [11]

    Slowfast-vgen: Slow- fast learning for action-driven long video generation

    Yining Hong, Beide Liu, Maxine Wu, Yuanhao Zhai, Kai- Wei Chang, Lingjie Li, Kevin Lin, Chung-Ching Lin, Jian- feng Wang, Zhengyuan Yang, et al. Slowfast-vgen: Slow- fast learning for action-driven long video generation. arXiv preprint arXiv:2410.23277, 2024. 2

    arXiv

  4. [12]

    Lvcd: reference-based lineart video colorization with diffusion models

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

  5. [13]

    Dˆ 2it: Dynamic diffusion transformer for accurate image generation

    Weinan Jia, Mengqi Huang, Nan Chen, Lei Zhang, and Zhendong Mao. Dˆ 2it: Dynamic diffusion transformer for accurate image generation. In Proceedings of the Computer Vision and Pattern Recognition Conference , pages 12860– 12870, 2025. 2

    2025

  6. [14]

    Hunyuanvideo: A systematic framework for large video generative models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 3

    2024 arXiv

  7. [15]

    Parrot: Pareto-optimal multi-reward reinforce- ment learning framework for text-to-image generation

    Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Jun- feng He, et al. Parrot: Pareto-optimal multi-reward reinforce- ment learning framework for text-to-image generation. In European Conference on Computer Vision, pag...

    2024

  8. [16]

    Mvbench: A comprehensive multi-modal video understand- ing benchmark

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understand- ing benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195–...

    2024

  9. [17]

    Llama-vid: An image is worth 2 tokens in large language models

    Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 3

    2024

  10. [18]

    Video-llava: Learning united visual rep- resentation by alignment before projection

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual rep- resentation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 3

    2023 arXiv

  11. [19]

    Realgeneral: Unifying visual generation via tempo- ral in-context learning with video models

    Yijing Lin, Mengqi Huang, Shuhan Zhuang, and Zhendong Mao. Realgeneral: Unifying visual generation via tempo- ral in-context learning with video models. arXiv preprint arXiv:2503.10406, 2025. 2

    2025 arXiv

  12. [20]

    Free- long: Training-free long video generation with spectralblend temporal attention

    Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Free- long: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918, 2024. 2

    2024 arXiv

  13. [21]

    Follow your pose: Pose- guided text-to-video generation using pose-free videos

    Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose- guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 4117–4125, 2024. 2

    2024

  14. [22]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024

    2024

  15. [23]

    Follow-your-click: Open-domain regional image animation via motion prompts

    Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Leqi Shen, Chenyang Qi, Jixuan Ying, Chengfei Cai, Zhifeng Li, Heung-Yeung Shum, et al. Follow-your-click: Open-domain regional image animation via motion prompts. In Proceed- ings of the AAAI Conference on Artificial Intelligence...

    2025

  16. [24]

    Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning

    Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207, 2025. 2

    2025

  17. [25]

    Anidoc: Animation creation made easier

    Yihao Meng, Hao Ouyang, Hanlin Wang, Qiuyu Wang, Wen Wang, Ka Leong Cheng, Zhiheng Liu, Yujun Shen, and Huamin Qu. Anidoc: Animation creation made easier. arXiv preprint arXiv:2412.14173, 2024. 2, 3, 6, 7

    2024 arXiv

  18. [26]

    Sakuga-42m dataset: Scaling up cartoon research

    Zhenglin Pan, Yu Zhu, and Yuxuan Mu. Sakuga-42m dataset: Scaling up cartoon research. arXiv preprint arXiv:2405.07425, 2024. 6

    2024 arXiv

  19. [27]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4195–4205,

  20. [28]

    Freenoise: Tuning-free longer video diffusion via noise rescheduling.arXiv preprint arXiv:2310.15169, 2023

    Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xin- tao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling.arXiv preprint arXiv:2310.15169, 2023. 2

    2023 arXiv

  21. [29]

    Learning transferable visual models from natural language supervi- sion

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

    2021

  22. [30]

    Consisti2v: Enhanc- ing visual consistency for image-to-video generation

    Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhanc- ing visual consistency for image-to-video generation. arXiv preprint arXiv:2402.04324, 2024. 2

    2024 arXiv

  23. [31]

    Video-xl: Extra-long vision language model for hour-scale video understanding

    Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 3, 5

    2024 arXiv

  24. [32]

    Inception transformer

    Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495–23509,

  25. [33]

    Probing multimodal large lan- guage models for global and local semantic representations

    Mingxu Tao, Quzhe Huang, Kun Xu, Liwei Chen, Yansong Feng, and Dongyan Zhao. Probing multimodal large lan- guage models for global and local semantic representations. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and...

    2024

  26. [34]

    To- wards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To- wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7

    2018 arXiv

  27. [35]

    Anti-oversmoothing in deep vision trans- formers via the fourier domain analysis: From theory to practice

    Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision trans- formers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 5

    2022 arXiv

  28. [36]

    Dualreal: Adaptive joint training for loss- less identity-motion fusion in video customization

    Wenchuan Wang, Mengqi Huang, Yijing Tu, and Zhen- dong Mao. Dualreal: Adaptive joint training for loss- less identity-motion fusion in video customization. arXiv preprint arXiv:2505.02192, 2025. 2

    2025 arXiv

  29. [37]

    Retake: Reducing temporal and knowledge redundancy for long video understanding

    Xiao Wang, Qingyi Si, Jianlong Wu, Shiyu Zhu, Li Cao, and Liqiang Nie. Retake: Reducing temporal and knowledge redundancy for long video understanding. arXiv preprint arXiv:2412.20504, 2024. 3

    2024 arXiv

  30. [38]

    Longllava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture

    Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longllava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 3

    2024

  31. [39]

    Videollamb: Long-context video understanding with recur- rent memory bridges

    Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recur- rent memory bridges. arXiv preprint arXiv:2409.01071 ,

    arXiv

  32. [40]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 7

    2004

  33. [41]

    Longvlm: Efficient long video understand- ing via large language models

    Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understand- ing via large language models. In European Conference on Computer Vision, pages 453–470. Springer, 2024. 3

    2024

  34. [42]

    Deep reward supervisions for tuning text-to-image diffusion models

    Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In European Conference on Computer Vision, pages 108–124. Springer, 2024. 3

    2024

  35. [43]

    Tooncrafter: Generative cartoon interpolation

    Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xin- tao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG), 43(6):1–11, 2024. 2, 3, 6, 7

    2024

  36. [44]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation. Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023. 3

    2023

  37. [45]

    Layeranimate: Layer-specific control for anima- tion

    Yuxue Yang, Lue Fan, Zuzen Lin, Feng Wang, and Zhaoxi- ang Zhang. Layeranimate: Layer-specific control for anima- tion. arXiv preprint arXiv:2501.08295, 2025. 2, 3

    2025 arXiv

  38. [46]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 3, 4, 6

    2024 arXiv

  39. [47]

    Animation line art colorization based on the op- tical flow method

    Yifeng Yu, Jiangbo Qian, Chong Wang, Yihong Dong, and Baisong Liu. Animation line art colorization based on the op- tical flow method. Computer Animation and Virtual Worlds, 35(1):e2229, 2024. 2, 3

    2024

  40. [48]

    Make pixels dance: High- dynamic video generation

    Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High- dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8850–8860, 2024. 2

    2024

  41. [49]

    Mim- icmotion: High-quality human motion video generation with confidence-aware pose guidance

    Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mim- icmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 3, 5

    2024 arXiv

  42. [50]

    Magiccolor: Multi-instance sketch colorization

    Yinhan Zhang, Yue Ma, Bingyuan Wang, Qifeng Chen, and Zeyu Wang. Magiccolor: Multi-instance sketch colorization. arXiv preprint arXiv:2503.16948, 2025. 2

    2025 arXiv

Pith tools

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