Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Unsupervised Cardiac Video Translation Via Motion Feature Guided Diffusion Model

T0 review · 3 major / 4 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 MFD-V2V synthesizes high-contrast cine cardiac videos from low-contrast DENSE scans by conditioning a diffusion model on learned motion, and reports a 70% relative gain in downstream segmentation Dice (0.31 to 0.81).

desk verdict A genuinely new DENSE-to-cine translation idea, undermined by a training/inference conditioning mismatch that leaves the central claim unvalidated. read the letter →

arxiv 2507.02003 v2 pith:JFNA4VOP submitted 2025-07-01 eess.IV

Swakshar Deb , Nian Wu , Frederick H. Epstein , Miaomiao Zhang This is my paper · ORCID
classification eess.IV
keywords cardiacMRIDENSECMRcinevideo-to-videotranslationdiffusionmodelmotionestimationunsupervisedlearningmyocardiumsegmentation
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

This paper tries to establish that cardiac motion, not just appearance, can drive unpaired video-to-video translation between two cardiac MRI sequences. Its model, MFD-V2V, learns motion from cine videos with a temporal-attention registration network, then guides a video diffusion model with those motion features so that, at inference, it can turn low-contrast DENSE CMR videos into high-contrast, temporally coherent cine videos. The authors report that the synthesized videos improve a downstream myocardium segmentation task, raising Dice from 0.31 to 0.81, and that the motion-conditioned model beats five baseline generative models on FID, KID, FVD, and FID-VID. If this holds, DENSE scans, which already carry precise motion information, would also yield usable anatomical images without an extra acquisition.

What carries the argument

The load-bearing machinery is the motion-conditioned video diffusion model whose conditioning signal is a multi-level spatiotemporal motion feature map produced by STME from displacement fields. During training, LTMA estimates those displacement fields from cine CMR by applying multi-head self-attention across time in a latent velocity space, capturing long-range temporal dependencies in cardiac motion. At inference, STME encodes the displacement field that comes with DENSE CMR, and cross-attention in the decoder of a 3D U-Net denoiser treats that feature map as key and value, so the generated video is steered by the supplied motion.

What would settle it

For the same hearts, compute the per-voxel displacement fields packaged with DENSE CMR and the displacement fields LTMA estimates from spatially aligned cine CMR, then measure the distribution shift between them (for example, with Wasserstein distance or overlap of displacement histograms); a large shift would mean the inference-time conditioning is out of distribution, so the generated cine would not faithfully reproduce true DENSE motion.

Watch

Extended reading notes

Core claim

The central claim is that a motion-conditioned video diffusion model can bridge unpaired cardiac MRI domains: when the denoiser is told where the myocardium is moving, it produces a cine-looking video that follows that motion. The paper introduces LTMA, a registration network with temporal multi-head self-attention, to estimate consistent displacement fields from cine CMR, and STME, a spatiotemporal encoder, to turn those displacement fields into multi-level motion features. Those features condition a 3D U-Net diffusion denoiser through cross-attention. At inference, DENSE magnitude frames are incorporated alongside random noise, and the displacement field packaged with DENSE CMR is encoded by STME to condition the denoiser. The paper reports best-in-class scores on FID, KID, FVD, and FID-VID against five baselines, and a segmentation Dice improvement from 0.31 on DENSE to 0.81 on the synthesized cine.

Load-bearing premise

The method assumes that the motion measurements that come with DENSE CMR and the motion the network learns from cine CMR line up in meaning and scale, so one motion guide can drive both without calibration.

Editorial extensions

If this is right

  • DENSE CMR, which encodes precise motion but suffers from low contrast, could be converted into cine-like videos without a second scan, making it easier for clinicians to see anatomy while keeping the motion information.
  • Motion conditioning itself is what drives the gain: the motion-conditioned ControlNet baseline beats unconditional VDM on temporal metrics, and adding STME on top of LTMA improves FID from 47.4 to 43.4 and FVD from 56.0 to 51.0.
  • Synthesized cine CMR can serve as input or training data for downstream analytical tasks, as shown by myocardium segmentation Dice rising from 0.31 on DENSE to 0.81 on generated cine.
  • The same motion-guided, unpaired translation recipe may apply to other imaging pairs where one modality carries reliable motion or deformation information but poor contrast.

Reading between the lines

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

  • The paper does not demonstrate that DENSE displacement fields and LTMA-estimated cine displacement fields are statistically interchangeable; a reader testing this would likely find a domain gap, and adding a small calibration step on the DENSE fields could make inference more robust.
  • Because the evaluation uses distributional video metrics rather than paired ground truth, the reported gains measure how cine-like the outputs are, not whether the generated videos reproduce the exact motion encoded in the DENSE scan; a direct strain comparison would be a stronger test.
  • The improvement in downstream segmentation could be driven partly by the higher contrast of generated cine rather than by motion fidelity; ablating the motion condition while keeping the same diffusion backbone would separate those contributions.
  • The 1000-step DDPM sampler makes the method computationally heavy; the same motion-conditioned framework could likely be distilled into fewer steps or moved to a latent space for clinical practicality.
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

3 major / 4 minor

Summary. The manuscript proposes MFD-V2V, an unsupervised video-to-video translation framework that synthesizes cine CMR videos from DENSE CMR videos. It introduces a Latent Temporal Multi-Attention (LTMA) registration network to estimate temporally consistent motion from cine CMR, and a Spatio-Temporal Motion Encoder (STME) that extracts multi-level motion features to condition a video diffusion model. The method is evaluated on a multi-site cardiac dataset; the authors report superior FID/KID/FVD/FID-VID over CycleGAN, NiceGAN, RecycleGAN, VDM, and ControlNet, and a downstream myocardium segmentation Dice improvement from 0.31 on DENSE to 0.81 on synthesized cine. The code is publicly available.

Significance. If the reported results are valid, the paper is among the first to tackle DENSE-to-cine CMR synthesis with an unpaired diffusion model, and the combination of an explicit motion-estimation network with motion-conditioned diffusion is a reasonable and potentially useful design direction. The public code release, the use of distribution-level video metrics, and the multi-site dataset are strengths. However, the paper's central inference scenario, conditioning on DENSE displacement fields, is not quantitatively validated as described, and the absence of uncertainty estimates and detailed experimental protocols makes the claimed gains difficult to assess.

major comments (3)
  1. [Sec. 3.2 (Inference Process) and Table 1] The training distribution and inference distribution of the motion condition are not matched. The diffusion model is trained with F extracted from displacement fields u_i produced by LTMA applied to cine CMR (Eqs. 2-3), while inference conditions on "the displacement field, directly provided with the DENSE CMR sequences." The manuscript gives no normalization, calibration, unit conversion, or distributional analysis connecting LTMA cine displacements (learned from image registration in pixel space) with DENSE displacement fields (encoded in phase, with different noise and artifact structure and possibly different units). Table 1 lists the condition for "Ours" as "Motion LTMA", which suggests the reported FID/KID/FVD/FID-VID numbers were computed under the cine-LTMA condition rather than the DENSE condition used at inference. The authors should either report the same metrics with DENSE displacement conditioning during evaluation, or explicitly state that the quantitative comparison does not test the method's stated inference procedure.
  2. [Sec. 4.2 (downstream segmentation)] The claim that synthesized cine CMR improves Dice from 0.31 to 0.81 ("a 70% relative increase") is reported without any experimental protocol. It is unclear which sequences were used, whether test subjects were held out from training the segmenter, what motion condition was used to generate the synthesized cine, whether the segmenter was trained on real cine, and whether the 0.81 Dice was obtained on real cine or synthesized cine. Because a segmenter trained on cine would be expected to perform better on cine-like inputs than on DENSE, this comparison alone does not validate translation quality. Please provide the full evaluation setup, including data splits and variability estimates.
  3. [Sec. 4.1 and Table 1] All quantitative comparisons are reported as single point estimates with no error bars, confidence intervals, or significance tests, and no train/validation/test split is described. FID/KID/FVD/FID-VID are stochastic and depend on the number of samples and the model checkpoint; without repeated runs or bootstrap intervals, the claimed superiority over VDM and ControlNet is not statistically supported. The authors should report results over multiple seeds or resamples and specify the exact partition of the 741 videos.
minor comments (4)
  1. [Eq. (2)] The loss is written as l(θ) while the text defines Θ for all network parameters; the notation should be unified, and the value of λ should be stated.
  2. [Sec. 4 (Dataset)] The text states that cine and DENSE sequences were "temporally and spatially aligned" while describing the setting as unpaired; please clarify whether alignment was used only for preprocessing or whether it permits paired evaluation, as this bears on the unpaired claim.
  3. [Sec. 4.1] Training is on 10 frames at 64×64 resolution, while the dataset is described as 40 frames at 128×128; please clarify how the reported video metrics were computed and at what resolution and sequence length.
  4. [Fig. 1 caption and Sec. 3] The phrase "height and weight" should be "height and width".

Circularity Check

1 steps flagged · score 6.0 of 10

Table 1's metrics use 'Motion LTMA' (motion estimated from real cine) as the condition, not the DENSE displacement fields named in the inference procedure, so the reported SOTA gains do not validate DENSE-to-cine translation.

  1. fitted input called prediction [Sec. 3.2 (Inference Process); Sec. 4.2, Table 1]
    "Table 1 reports 'Ours Motion LTMA 43.432 0.0179 50.962 20.124', while Sec. 3.2 states: 'The displacement field, directly provided with the DENSE CMR sequences [11], is used to extract motion conditions (see Fig. 1(b)).'"

    The conditioning motion u_i comes from LTMA applied to cine videos ('Given the learned motion fields, u_i = ... from our previously introduced LTMA registration network'), and Table 1 lists 'Motion LTMA' as the condition. Thus FID/KID/FVD/FID-VID measure cine-to-cine generation whose condition is computed from the very target videos being compared. They do not test the stated inference use of DENSE displacement fields, which have different physics, units, and noise. The only DENSE-side result (Dice 0.81 vs 0.31) omits the condition used, subject split, and any real-cine comparison. The headline SOTA claim therefore reduces to a target-derived-condition evaluation: fitted LTMA cine motion is used as the condition while the paper claims to predict DENSE-to-cine translation.

full rationale

The paper is not circular in its mathematical derivation: the diffusion objective L_epsilon = E[||epsilon - epsilon_theta(x_t, t, F)||^2] is a standard conditional denoising objective, and the STME features F are a learned function of the input motion field. There is no load-bearing self-citation or imported uniqueness theorem; citations such as [3,28] only supply architectural backbones. The circularity is confined to the experimental validation of the central DENSE-to-cine claim. Table 1's 'Motion LTMA' condition is an LTMA displacement estimated from the real cine videos, i.e., from the target distribution, so the quantitative SOTA comparisons are not independent evidence that the model translates DENSE displacement into cine. Because the inference procedure explicitly substitutes DENSE displacement fields, and no quantitative metric is reported for that condition, the paper's strongest quantitative claim partially reduces to a target-conditioned generation benchmark. The Dice experiment could have provided the missing DENSE-side validation but is under-specified. This is a partial circularity/evaluation gap, reflected in a score of 6; the underlying model equations are not themselves self-referential.

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

No new physical entities are proposed; LTMA and STME are network modules, not scientifically postulated entities. The load-bearing premise is the assumed interchangeability of DENSE displacement and LTMA-learned cine motion. Hyperparameters such as lambda and h are free design choices.

free parameters (3)
  • lambda_loss_weight
    Positive weighting parameter in LTMA registration loss (Eq. 2); set by hand, value not reported.
  • attention_heads_h
    Number of heads in LTMA multi-head attention (Eq. 1); chosen by hand, value not reported.
  • diffusion_steps_T = 1000
    Total diffusion steps, set following DDPM; a design choice.
assumptions (4)
  • standard math Diffusion model formulation (Ho et al. 2020) is assumed; the noisy forward process and denoising objective are taken as given.
    Used as backbone in Sections 2 and 3.2.
  • standard math Stationary velocity field parameterization for diffeomorphic registration (Vercauteren et al. 2008) is assumed for the transformation fields in Eq. (2).
    Used to define transformation between frames.
  • ad hoc to paper DENSE displacement fields are directly usable as the motion condition at inference, despite the diffusion being trained on LTMA motion from cine.
    Stated in Sec. 3.2 'Inference Process'; never validated. This is the weakest load-bearing premise.
  • domain assumption Temporal/spatial alignment preprocessing (resampling to 1 mm^2 and 40 frames) preserves cardiac motion information needed for synthesis.
    Described in Section 4 Dataset paragraph; no validation that the resampling retains clinically meaningful motion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Cardiac Video Translation Via Motion Feature Guided Diffusion Model." pith.science (2026). https://pith.science/paper/JFNA4VOP

@misc{pith2026250702003,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Cardiac Video Translation Via Motion Feature Guided Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFNA4VOP}},
  note         = {Machine review of arXiv:2507.02003}
}
read the original abstract

This paper presents a novel motion feature guided diffusion model for unpaired video-to-video translation (MFD-V2V), designed to synthesize dynamic, high-contrast cine cardiac magnetic resonance (CMR) from lower-contrast, artifact-prone displacement encoding with stimulated echoes (DENSE) CMR sequences. To achieve this, we first introduce a Latent Temporal Multi-Attention (LTMA) registration network that effectively learns more accurate and consistent cardiac motions from cine CMR image videos. A multi-level motion feature guided diffusion model, equipped with a specialized Spatio-Temporal Motion Encoder (STME) to extract fine-grained motion conditioning, is then developed to improve synthesis quality and fidelity. We evaluate our method, MFD-V2V, on a comprehensive cardiac dataset, demonstrating superior performance over the state-of-the-art in both quantitative metrics and qualitative assessments. Furthermore, we show the benefits of our synthesized cine CMRs improving downstream clinical and analytical tasks, underscoring the broader impact of our approach. Our code is publicly available at https://github.com/SwaksharDeb/MFD-V2V.

Figures

Figures reproduced from arXiv: 2507.02003 by the authors.

Figure 1
Figure 1. An overview of our proposed MFD-V2V model. Left to right: (a) the training process with LTMA registration and multli-level motion feature guided video diffusion. (b) the inference process, where encoded DENSE motion features are used as condition. say, for the ith training data, we have a sequence of image frames {I 0 i , · · · , IT i }. By setting the first frame {I 0 i } as a reference (source) image, there exists… view at source ↗
Figure 2
Figure 2. Unpaired Cine and DENSE cardiac sequences of the left ventricle from the Cine-DENSE dataset [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of Cine CMR synthesis. Our approach (bottom row) preserves both spatial and temporal coherence, while VDM struggles with anatomical consistency during motion, and ControlNet shows improved temporal stability but slightly reduced spatial details. causes degradation in the quality of the synthesize cardiac sequences due to lack of spatiotemporal continuity in the motion, while inclusion of LTA f… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages

  1. [1]

    Journal of magnetic resonance (San Diego, Calif.: 1997)137(1), 247 (1999)

    Aletras, A.H., Ding, S., Balaban, R.S., Wen, H.: Dense: displacement encoding with stimulated echoes in cardiac functional mri. Journal of magnetic resonance (San Diego, Calif.: 1997)137(1), 247 (1999)

    work page 1999

  2. [2]

    In: IJCAI

    Balaji, Y., Min, M.R., Bai, B., Chellappa, R., Graf, H.P.: Conditional gan with discriminative filter generation for text-to-video synthesis. In: IJCAI. vol. 1, p. 2 (2019)

    work page 2019

  3. [3]

    IEEE transactions on medical imaging38(8), 1788–1800 (2019)

    Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging38(8), 1788–1800 (2019)

    2019

  4. [4]

    In: Proceedings of the European conference on computer vision (ECCV)

    Bansal, A., Ma, S., Ramanan, D., Sheikh, Y.: Recycle-gan: Unsupervised video re- targeting. In: Proceedings of the European conference on computer vision (ECCV). pp. 119–135 (2018)

    work page 2018

  5. [5]

    arXiv preprint arXiv:1801.01401 (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)

    arXiv 2018

  6. [6]

    Journal of the American College of Cardiology84, 648–659 (2024)

    Chadalavada, S., Fung, K., Rauseo, E., Lee, A.M., Khanji, M.Y., Amir-Khalili, A., Paiva, J., Naderi, H., Banik, S., Chirvasa, M., et al.: Myocardial strain measured by cardiac magnetic resonance predicts cardiovascular morbidity and death. Journal of the American College of Cardiology84, 648–659 (2024)

    work page 2024

  7. [7]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Chen, J., Li, Y., Ma, K., Zheng, Y.: Generative adversarial networks for video-to- video domain adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3462–3469 (2020)

    work page 2020

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, R., Huang, W., Huang, B., Sun, F., Fang, B.: Reusing discriminators for encoding: Towards unsupervised image-to-image translation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8168–8177 (2020)

    work page 2020

Show all 31 references
  1. [9]

    In: Pro- ceedings of the AAAI Conference on Artificial Intelligence

    Chu, E., Huang, T., Lin, S.Y., Chen, J.C.: Medm: Mediating image diffusion mod- els for video-to-video translation with temporal correspondence guidance. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1353–1361 (2024)

    2024

  2. [10]

    In: Medical Image ComputingandComputer-AssistedIntervention–MICCAI2016:19thInternational Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19

    Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image ComputingandComputer-AssistedIntervention–MICCAI2016:19thInternational Conference, Athens, Greece, October 17-21, 20...

    2016

  3. [11]

    Journal of Cardiovascular Magnetic Resonance26 (2024)

    Ghadimi, S., Bivona, D., Bilchick, K., Epstein, F.: Deep learning-based prognostic model using cine dense mri for outcome prediction after cardiac resynchronization therapy. Journal of Cardiovascular Magnetic Resonance26 (2024)

    2024

  4. [12]

    IEEE transactions on medical imaging31(9), 1669–1681 (2012) 10 S

    Gilliam, A.D., Epstein, F.H.: Automated motion estimation for 2-d cine dense mri. IEEE transactions on medical imaging31(9), 1669–1681 (2012) 10 S. Deb et al

    2012

  5. [13]

    Advances in neural information processing systems30 (2017)

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30 (2017)

    2017

  6. [14]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  7. [15]

    Advances in Neural Information Processing Systems35, 8633– 8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in Neural Information Processing Systems35, 8633– 8646 (2022)

    2022

  8. [16]

    arXiv preprint arXiv:2307.14073 (2023)

    Hu,Z.,Xu,D.:Videocontrolnet:Amotion-guidedvideo-to-videotranslationframe- work by using diffusion model with controlnet. arXiv preprint arXiv:2307.14073 (2023)

    2023 arXiv

  9. [17]

    In: International Conference on Machine Learning

    Jabri, A., Fleet, D.J., Chen, T.: Scalable adaptive computation for iterative gener- ation. In: International Conference on Machine Learning. pp. 14569–14589. PMLR (2023)

    2023

  10. [18]

    Journal of Cardiovascular Magnetic Resonance 27 (2025)

    Lei, P., Xing, J., Wang, Y., Ghadimi, S., Epstein, F.H., Zhang, M.: Improved dense- guided deep networks to predict myocardial strain from routine cine magnetic resonance images by accounting for rotation. Journal of Cardiovascular Magnetic Resonance 27 (2025)

    2025

  11. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Liang, F., Wu, B., Wang, J., Yu, L., Li, K., Zhao, Y., Misra, I., Huang, J.B., Zhang, P., Vajda, P., et al.: Flowvid: Taming imperfect optical flows for consistent video- to-video synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  12. [20]

    arXiv preprint arXiv:1611.02163 (2016)

    Metz, L., Poole, B., Pfau, D., Sohl-Dickstein, J.: Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163 (2016)

    2016 arXiv

  13. [21]

    Medical image analysis12(6), 752–763 (2008)

    Reinhardt, J.M., Ding, K., Cao, K., Christensen, G.E., Hoffman, E.A., Bodas, S.V.: Registration-based estimates of local lung tissue expansion compared to xenon ct measures of specific ventilation. Medical image analysis12(6), 752–763 (2008)

    2008

  14. [22]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  15. [23]

    Reviews in Cardiovascular Medicine 24(9), 261 (2023)

    Sillanmäki, S., Vainio, H.L., Ylä-Herttuala, E., Husso, M., Hedman, M.: Mea- suring cardiac dyssynchrony with dense (displacement encoding with stimulated echoes)—a systematic review. Reviews in Cardiovascular Medicine 24(9), 261 (2023)

    2023

  16. [24]

    arXiv preprint arXiv:1812.01717 (2018)

    Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)

    2018 arXiv

  17. [25]

    Advances in neural information pro- cessing systems 30 (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems 30 (2017)

    2017

  18. [26]

    In: International conference on medical image computing and computer-assisted intervention

    Vercauteren, T., Pennec, X., Perchant, A., Ayache, N.: Symmetric log-domain dif- feomorphic registration: A demons-based approach. In: International conference on medical image computing and computer-assisted intervention. Springer (2008)

    2008

  19. [27]

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

    Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text- to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623–7633 (2023)

    2023

  20. [28]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Wu, N., Xing, J., Zhang, M.: Tlrn: Temporal latent residual networks for large deformation image registration. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 728–738. Springer (2024) Title Suppressed Due to Excessive Length 11

    2024

  21. [29]

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

    Xu, Z., Zhang, J., Liew, J.H., Yan, H., Liu, J.W., Zhang, C., Feng, J., Shou, M.Z.: Magicanimate: Temporally consistent human image animation using diffu- sion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1481–1490 (2024)

    2024

  22. [30]

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

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836–3847 (2023)

    2023

  23. [31]

    In: Proceedings of the IEEE interna- tional conference on computer vision

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017)

    2017

Pith tools

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