REVIEW 4 major objections 6 minor 1 cited by
LMP: Leveraging Motion Prior in Zero-Shot Video Generation with Diffusion Transformer
T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that a frozen diffusion transformer video generator can be made to follow the motion of a user-supplied reference video at inference time, with no training, in both text-to-video and image-to-video generation.
desk verdict A plausible training-free DiT motion transfer pipeline whose real-video use case rests on an unvalidated noise-injection shortcut; the merits are real but the evidence needs work. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the extended self-attention in each MM-DiT block: the target video's queries attend to both target tokens and reweighted reference foreground keys and values, letting the reference video steer motion without training. Around this sit three modules: Fore-Background Disentangle Module (FBDM), which uses text-video and video-text attention maps to locate the moving subject tokens; Reweighted Motion Transfer Module (RMTM), which scales reference keys by $\lambda=0.98$ and injects them; and Appearance Separation Module (ASM), which optimizes target hidden states to suppress attention to the reference subject prompt when target and reference subjects differ. Real videos are handled by direct noise addition, $z'_t = \lambda_t z'_0 + (1-\lambda_t)\epsilon$, synchronizing the reference denoising with the target.
What would settle it
Take a real reference video, add noise using the paper's Eq. 10, denoise it with the frozen generator, and track the subject; if the reconstructed trajectory differs from the original, the injected tokens do not carry the real motion.
Extended reading notes
Core claim
Under the DiT architecture, text, image, and video information are all handled as tokens in one self-attention mechanism, so tokens from a reference video can be concatenated into the target video's attention. LMP separates the reference moving subject from its background by reading text-video attention maps; passes only foreground reference keys and values into the target attention, reweighting the reference keys to keep appearance from overshadowing the prompt; and applies a gradient-descent step that lowers attention between the reference subject prompt and target tokens when the two subjects differ. For real reference videos, it skips inversion and instead noisifies the video directly along the same schedule used for generation. The paper reports that on the DAVIS-derived evaluation, LMP outperforms the prior zero-shot motion-transfer method and the base model on motion fidelity, consistency, and prompt alignment.
Load-bearing premise
The fragile step is the assumption that adding noise to a real clip and letting the frozen generator denoise it honestly preserves the clip's motion; if that fails, the extracted tokens do not describe the real movement.
Editorial extensions
If this is right
- Users could steer generated video motion by example clip, including motions hard to describe in prompts.
- The same framework works for image-to-video, so a still image of a new subject can be animated with the reference clip's motion.
- No control-signal preprocessing such as depth, pose, or trajectories is needed, lowering the barrier compared to training-based controllers.
- Because the base model stays frozen, the method can be applied to newer DiT checkpoints without retraining.
- Motion transfer stays responsive to prompt edits, so subject and background can be changed independently of the copied motion.
Reading between the lines
- The paper does not test this, but a reconstruction experiment on real videos would tell whether the direct-noise tokens carry true motion or only low-frequency cues.
- The same token-injection mechanism could presumably transfer camera motion or object trajectories, not just subject motion, since the foreground-background separation defines what gets injected.
- If the direct-noise assumption holds, LMP could be combined with other token-level edits to get simultaneous motion and content control in one DiT pipeline.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes LMP, a training-free framework for controlling motion in DiT-based video generation (CogVideoX-5B) by injecting motion tokens extracted from a reference video during inference. Three modules are introduced: a foreground-background disentangle module (FBDM) that uses attention maps to identify foreground tokens, a reweighted motion transfer module (RMTM) that concatenates reference keys/values into the target self-attention with a reweighting factor, and an appearance separation module (ASM) that suppresses reference-subject appearance via gradient descent on hidden states. For real reference videos, the paper proposes directly noising the video with Eq. (10) rather than DDIM inversion. Experiments on a prompt-annotated DAVIS subset compare LMP against CogVideoX and DMT, reporting improved alignment, consistency, PickScore, and a self-defined Motion Fidelity metric, plus qualitative results for text-to-video and image-to-video settings.
Significance. If the central claims hold, LMP would be a useful plug-and-play method for zero-shot motion control on frozen DiT video generators, extending motion transfer to image-to-video without training. The paper explicitly ships a frozen-base-model inference-time algorithm, which is reproducible in principle, and it identifies a real limitation of prior UNet-based motion-transfer methods. However, the significance is tempered by the fact that the quantitative gains over the baseline are small (e.g., Mean Align 25.68 vs 24.84; Mean Pick 20.69 vs 20.46), the comparison set is limited to one prior method (DMT) plus baseline, and the real-video procedure is not validated by a reconstruction experiment. The Motion Fidelity metric is also not fully specified, making the headline motion-control gains difficult to interpret.
major comments (4)
- [Sec. 4.5, Eq. (10)] The real-video motion injection procedure is load-bearing and unvalidated. Directly setting z'_t = λ_t·z'_0 + (1-λ_t)·ε and then denoising with the frozen DiT is not an inversion; the paper asserts that this yields a 'controllable' trajectory because at t=0 it recovers z_0, but no experiment shows that the denoised reference video preserves the original subject motion, position, speed, or background-foreground separation. Since the quantitative evaluations on DAVIS and all qualitative real-video examples rely on Eq. (10), the extracted motion tokens may not correspond to the user-provided motion. I request a reconstruction experiment: noise a real video with Eq. (10) at the operating timesteps, denoise with the frozen model, and report how well the denoised output matches the original video's motion (e.g., tracker overlap, trajectory error, or qualitative side-by-side). Without this, the real-video use case and the reported MF gains are not established.
- [Sec. 5.1 and Table 1] The quantitative evaluation is underspecified. The paper reports Align, Cons., Pick., and MF but does not state how many DAVIS videos/clips are used, how the test set is split, whether the numbers are averages over multiple seeds, or whether any statistical significance is computed. The Motion Fidelity metric is only described as 'leveraging a pre-trained tracker to monitor subject movement'; the exact formula, normalization, and what constitutes a 'correct' motion match are missing. Given that the headline claim is state-of-the-art motion control, the MF definition and the size/selection of the evaluation set need to be provided, along with error bars or per-video variance. Without this, the small improvements in Table 1 cannot be distinguished from noise.
- [Sec. 4.3, 4.4, and 5.1] The core mechanism depends on several hand-set hyperparameters with no sensitivity analysis: the reweighting coefficient λ=0.98 in Eq. (6), the ASM learning rate β=100 in Eq. (9), the stage thresholds T1=40, T2=45, T3=35, and the top-k=1/5 selection in Eq. (8). These values are presented without justification or ablation. Since the method's behavior likely changes sharply with λ and β (e.g., λ too close to 1 may over-inject appearance; β too large may distort the target), a sweep or at least a small ablation table for λ and β is necessary to argue that the reported performance is not a narrow peak. I also note that ASM's gradient descent is an optimization procedure, so the claim in Sec. 4.2 that FBDM incurs 'no additional computational or time overhead' should be qualified; ASM clearly adds per-step gradient computations.
- [Sec. 5.3 and Abstract] The claim of 'state-of-the-art performance' is not supported by the comparison set. Table 1 compares only to the baseline CogVideoX and the UNet-based DMT, both of which are not current DiT motion-transfer methods. The related work mentions a DiT-based motion transfer method [25], but it is not compared. To substantiate SOTA status, the paper should compare against at least one additional recent DiT-based or training-free motion-transfer approach, or temper the claim to 'superior to DMT and baseline on our evaluation set.' The current framing overstates what the experiments demonstrate.
minor comments (6)
- [General] The manuscript uses an ACM conference template with placeholder text ('Conference acronym XX', 'June 03–05, 2018', 'Received 20 February 2007') and a broken citation for [17]; these formatting issues should be fixed before resubmission.
- [Sec. 4.3, Eq. (5)-(6)] The notation is inconsistent: the text introduces \(\hat{K}^v_{ref}\) and \(\hat{V}^v_{ref}\) as 'reference motion tokens,' but Eq. (6) uses the same symbols with a reweighting factor λ applied only to K. It would be clearer to define whether the hats mean the foreground-extracted tokens after FBDM and to state explicitly whether the reweighting applies to the keys only, the values only, or both.
- [Algorithm 1] Line 4 calls 'Mapping(z_T, c)' which is not defined in the main text; the mapping from the latent video to the hidden states of the MM-DiT block should be described. Also, the notation REST(h_tar) is used without explanation.
- [Sec. 5.1] The sentence 'Following [25, 44], we use the DAVIS dataset' is ambiguous: [25] is a DiT motion transfer paper, but the dataset construction (three parts with modified subject/background prompts) is not described in enough detail to replicate. Please specify how many videos are in each part and what the 'modified subject annotations' look like.
- [Sec. 4.2] The FBDM relies on an LLM (qwen2.5-VL-72B-instruct) to generate the reference prompt and on another LLM step to extract subject words. The paper provides no test of the reliability of these LLM-generated prompts; a wrong subject word would break the attention-based foreground extraction. A simple validation (e.g., manually inspecting a sample of extracted subject words) would increase confidence.
- [Sec. 5.2, Fig. 5] The image-to-video qualitative examples are limited to two cases. Since image-to-video is a stated novelty, more examples (including failure cases) would strengthen the paper. The current figure does not show whether the reference image's background is preserved or whether the subject's initial pose is respected.
Circularity Check
No significant circularity: LMP is an inference-time token-injection method whose design is not derived from its evaluation metrics, and its self-citation is background only.
full rationale
The paper's derivation chain is self-contained: the three modules (FBDM, RMTM, ASM) directly manipulate attention maps and hidden states of a frozen DiT model rather than fitting to the evaluation metrics. The motion-transfer claim is tested against an external baseline (CogVideoX) and a prior method (DMT), with standard CLIP-based alignment and consistency scores plus PickScore; the self-defined Motion Fidelity metric measures the intended behavior but does not appear in the method's construction. The hyperparameters (lambda=0.98, beta=100, T1/T2/T3) are hand-set, which is a robustness concern rather than circularity. The only questionable passage is Section 4.5, Eq. (10), where direct noise addition is asserted to make the denoising trajectory 'controllable'; this is an unvalidated assumption about DDIM-inversion-free conditioning, but it is not circular because it defines the reference-input encoding rather than deriving the output from the metric or from a fitted parameter. The self-citation to Motion-Zero [3] appears only as related-work background and is not load-bearing. Overall, no claimed result reduces by construction to its inputs.
Assumptions & free parameters
free parameters (5)
- RMTM reweighting coefficient λ =
0.98
- ASM learning rate β =
100
- Denoising stage thresholds T1, T2, T3 =
T1=40, T2=45, T3=35
- ASM top-k fraction =
1/5 (k=ceil(n/5))
- Real-video noise schedule λ_t in Eq. (10) =
unspecified
assumptions (4)
- domain assumption DiT attention maps between prompt tokens and video tokens reliably localize the reference video's moving subject.
- domain assumption The LLM (qwen2.5-VL-72B-instruct) produces accurate dense prompts and correct subject-word extraction for arbitrary reference videos.
- ad hoc to paper Adding noise to a real reference video by Eq. (10) and denoising with the frozen DiT yields a trajectory whose tokens represent the original video's motion.
- domain assumption CogVideoX-5B, with keys and values from reference foreground tokens injected, can generate a new subject that follows the reference motion while respecting the text prompt.
Cite this review
Pith. "Pith review of LMP: Leveraging Motion Prior in Zero-Shot Video Generation with Diffusion Transformer." pith.science (2026). https://pith.science/paper/TY67CKWD
@misc{pith2026250514167,
author = {Pith},
title = {Pith review of: LMP: Leveraging Motion Prior in Zero-Shot Video Generation with Diffusion Transformer},
year = {2026},
howpublished = {\url{https://pith.science/paper/TY67CKWD}},
note = {Machine review of arXiv:2505.14167}
}
read the original abstract
In recent years, large-scale pre-trained diffusion transformer models have made significant progress in video generation. While current DiT models can produce high-definition, high-frame-rate, and highly diverse videos, there is a lack of fine-grained control over the video content. Controlling the motion of subjects in videos using only prompts is challenging, especially when it comes to describing complex movements. Further, existing methods fail to control the motion in image-to-video generation, as the subject in the reference image often differs from the subject in the reference video in terms of initial position, size, and shape. To address this, we propose the Leveraging Motion Prior (LMP) framework for zero-shot video generation. Our framework harnesses the powerful generative capabilities of pre-trained diffusion transformers to enable motion in the generated videos to reference user-provided motion videos in both text-to-video and image-to-video generation. To this end, we first introduce a foreground-background disentangle module to distinguish between moving subjects and backgrounds in the reference video, preventing interference in the target video generation. A reweighted motion transfer module is designed to allow the target video to reference the motion from the reference video. To avoid interference from the subject in the reference video, we propose an appearance separation module to suppress the appearance of the reference subject in the target video. We annotate the DAVIS dataset with detailed prompts for our experiments and design evaluation metrics to validate the effectiveness of our method. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in generation quality, prompt-video consistency, and control capability. Our homepage is available at https://vpx-ecnu.github.io/LMP-Website/
Figures
Figures from the paper (6 more)
Forward citations
Cited by 1 Pith paper
-
O-DisCo-Edit: Object Distortion Control for Unified Realistic Video Editing
A video editor trained on randomly distorted objects, then steered by adaptive noise at inference, is claimed to surpass dedicated and unified editors across eight tasks with far less training.
Reference graph
Works this paper leans on
-
[25]
Alexander Pondaven, Aliaksandr Siarohin, Sergey Tulyakov, Philip H. S. Torr, and Fabio Pizzati. 2024. Video Motion Transfer with Diffusion Transformers. ArXiv abs/2412.07776 (2024). https://api.semanticscholar.org/CorpusID:274610723
work page Pith review arXiv 2024
-
[1]
Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...
arXiv 2023
-
[2]
Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. 2024. DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation. ArXiv abs/2412.18597 (2024). https://api.semanticscholar.org/ CorpusID:274992610
arXiv 2024
-
[3]
Changgu Chen, Junwei Shu, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. 2024. Motion-Zero: Zero-Shot Moving Object Control Framework for Diffusion-Based Video Generation. arXiv preprint arXiv:2401.10150 (2024)
arXiv 2024
-
[4]
Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. 2023. VideoCrafter1: Open Diffusion Models for High-Quality Video Generation. arXiv:2310.19512 [cs.CV]
arXiv 2023
-
[5]
Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. 2023. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv e-prints (2023), arXiv–2305
work page 2023
-
[6]
Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. 2023. AnimateAnything: Fine-Grained Open Domain Image Animation with Motion Guidance. arXiv preprint arXiv:2311.12886 (2023)
arXiv 2023
-
[7]
Yingying Deng, Xiangyu He, Fan Tang, and Weiming Dong. 2024. Z*: Zero- shot Style Transfer via Attention Reweighting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024), 6934–6944. https://api. semanticscholar.org/CorpusID:272717799
work page 2024
Show all 54 references
-
[8]
Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. 2023. Structure and Content-Guided Video Synthesis with Diffusion Models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (2023), 7312–7322. https://api.semantics...
2023
-
[9]
Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. 2023. TokenFlow: Consistent Diffusion Features for Consistent Video Editing.ArXiv abs/2307.10373 (2023). https://api.semanticscholar.org/CorpusID:259991741
2023 arXiv
-
[10]
Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, 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
-
[11]
Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. 2022. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221 (2022)
2022 arXiv
-
[12]
Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. 2022. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)
2022 arXiv
-
[13]
Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851
2020
-
[14]
Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. 2024. Peekaboo: Interactive video generation via masked-diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8079–8088
2024
-
[15]
Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. 2024. Pick-a-pic: An open dataset of user preferences for text-to- image generation. Advances in Neural Information Processing Systems 36 (2024)
2024
-
[16]
Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)
2024 arXiv
-
[17]
Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux
2024
-
[18]
Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. 2023. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 9298– 9309
2023
-
[19]
Wan-Duo Kurt Ma, J. P. Lewis, and W. Kleijn. 2023. TrailBlazer: Trajectory Control for Diffusion-Based Video Generation. In ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia . https: //api.semanticscholar.org/CorpusID:266725649
2023
-
[20]
Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. 2024. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 4117–4125
2024
-
[21]
Tuna Han Salih Meral, Hidir Yesiltepe, Connor Dunlop, and Pinar Yanardag. 2024. MotionFlow: Attention-Driven Motion Transfer in Video Diffusion Models.arXiv preprint arXiv:2412.05275 (2024)
2024 arXiv
-
[22]
Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. 2023. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)
2023 arXiv
-
[23]
John Mullan, Duncan Crawbuck, and Aakash Sastry. 2023. Hotshot-XL. https: //github.com/hotshotco/hotshot-xl
2023
-
[24]
William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. In Proceedings of the IEEE/CVF international conference on computer vision . 4195–4205
2023
-
[26]
Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. 2017. The 2017 DAVIS Challenge on Video Object Segmentation. arXiv:1704.00675 (2017)
2017 arXiv
-
[27]
Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. 2023. Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023)
2023 arXiv
-
[28]
Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen
-
[29]
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
-
[30]
Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep Conference acronym ’XX, June 03–05, 2018, Woodst...
2022
-
[31]
Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. 2022. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)
2022 arXiv
-
[32]
Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)
2020 arXiv
-
[33]
Spencer Sterling. 2023. ZeroScope. https://huggingface.co/cerspense/zeroscope_ v2_576w
2023
-
[34]
Yoad Tewel, Rinon Gal, Dvir Samuel Yuval Atzmon, Lior Wolf, and Gal Chechik
-
[35]
Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. 2024. SV3D: Novel Multi-view Synthesis and 3D Generation from a Single Image using Latent Video Diffusion. arXiv preprint arXiv:2403.12008 (2024)
2024 arXiv
-
[36]
Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, et al. 2025. Wan: Open and Advanced Large-Scale Video Generative Models. arXiv preprint arXiv:2503.20314 (2025)
2025 arXiv
-
[37]
Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. 2023. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)
2023 arXiv
-
[38]
Luozhou Wang, Ziyang Mai, Guibao Shen, Yixun Liang, Xin Tao, Pengfei Wan, Di Zhang, Yijun Li, and Yingcong Chen. 2024. Motion Inversion for Video Customization. arXiv:2403.20193 [cs.CV] https://arxiv.org/abs/2403.20193
2024 arXiv
-
[39]
Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. 2023. Videocomposer: Composi- tional video synthesis with motion controllability.Advances in Neural Information Processing Systems 36 (2023), 7594–7611
2023
-
[40]
Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. 2024. Motionctrl: A unified and flexible motion controller for video generation. InACM SIGGRAPH 2024 Conference Papers. 1–11
2024
-
[41]
Zeqi Xiao, Yifan Zhou, Shuai Yang, and Xingang Pan. 2024. Video Diffusion Models are Training-free Motion Interpreter and Controller.ArXiv abs/2405.14864 (2024). https://api.semanticscholar.org/CorpusID:269983459
2024 arXiv
-
[42]
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
-
[43]
Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. 2023. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 14246–14255
2023
-
[44]
Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. 2023. Space-Time Diffusion Features for Zero-Shot Text-Driven Motion Transfer. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023), 8466–8476. https://api.semanticscholar.org/...
2023
-
[45]
Sheng-Siang Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. 2023. DragNUWA: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. ArXiv abs/2308.08089 (2023). https: //api.semanticscholar.org/CorpusID:260925229
2023 arXiv
-
[46]
Yongsheng Yu, Ziyun Zeng, Haitian Zheng, and Jiebo Luo. 2025. OmniPaint: Mas- tering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting. https://api.semanticscholar.org/CorpusID:276929167
2025
-
[47]
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
-
[48]
Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. 2024. Motiondirector: Motion customiza- tion of text-to-video diffusion models. InEuropean Conference on Computer Vision. Springer, 273–290
2024
-
[49]
Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. 2024. Uni-controlnet: All-in-one control to text-to- image diffusion models. Advances in Neural Information Processing Systems 36 (2024)
2024
-
[50]
Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. 2022. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018 (2022)
2022 arXiv
-
[51]
Yufan Zhou, Haoyu Shen, and Huan Wang. 2025. FreeBlend: Advancing Con- cept Blending with Staged Feedback-Driven Interpolation Diffusion. ArXiv abs/2502.05606 (2025). https://api.semanticscholar.org/CorpusID:276250497
2025
-
[52]
Tianrui Zhu, Shiyi Zhang, Jiawei Shao, and Yansong Tang. 2025. KV-Edit: Training-Free Image Editing for Precise Background Preservation. ArXiv abs/2502.17363 (2025). https://api.semanticscholar.org/CorpusID:276574734 Received 20 February 2007; revised 12 March 2009; accepted 5...
2025 arXiv
-
[2022]
arXiv preprint arXiv:2204.06125 1, 2 (2022), 3
Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1, 2 (2022), 3
2022 arXiv
-
[2024]
ArXiv abs/2411.07232 (2024)
Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models. ArXiv abs/2411.07232 (2024). https://api.semanticscholar.org/CorpusID: 273962996
2024 arXiv
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.