Pith. sign in

REVIEW 4 major objections 5 minor 35 references

SG2VID: Scene Graphs Enable Fine-Grained Control for Video Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read SG2VID claims that scene graphs, not text prompts, are the right conditioning signal for surgical video synthesis, because editing a graph node gives direct control over tool size, movement, and scene layout.

desk verdict A real extension with good generation numbers, but the fine-grained control headline is not yet backed by evidence that graph edits translate to output changes. read the letter →

arxiv 2506.03082 v2 pith:HE63WP2M submitted 2025-06-03 cs.CV

classification cs.CV
keywords scenegraphconditioningvideosynthesissurgicalsimulationdiffusionmodelsfine-grainedcontrolgenerativeaugmentationcontrollablegeneration
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

SG2VID aims to show that scene graphs—structured graphs whose nodes are surgical tools and anatomical structures and whose edges record spatial adjacency—can serve as the conditioning signal for video synthesis, giving users fine-grained control that text prompts, masks, or trajectories do not easily provide. The paper claims this is the first diffusion-based scene-graph-to-video model and demonstrates it on three public surgical video datasets spanning cataract and cholecystectomy procedures. If the claim is right, a surgeon or trainer could edit a graph to resize or move a tool, make a new tool enter the scene, or simulate a rare complication, and the model would render a corresponding video. The paper also reports that synthetic videos generated this way improve a downstream surgical phase-recognition model when added to its training set. The trade-off it rejects is that precise synthesis must come at the cost of human controllability.

What carries the argument

The load-bearing object is the enriched scene graph G=(V,E), built from per-frame segmentation masks. A node vj is a vector in $R^{{d+7}}$ containing d class dimensions, a 2D centroid, a 2D spatial spread (height and width), a 2D average optical flow, and a 1D average depth, with edges connecting spatially adjacent components. This representation is what makes the control fine-grained: editing one node property, such as pupil diameter, is a human-understandable operation, yet the node also carries enough motion and layout information for synthesis. Two graph attention encoders, one local-reconstruction encoder and one global-contrastive encoder, turn the graph into embeddings that are concatenated and injected into the temporal layers of a 3D U-Net latent video diffusion model; optional first-frame conditioning and first-frame feature injection provide appearance grounding.

What would settle it

Take a held-out video with manual segmentations, build the scene graph from the automatic masks, edit one node property such as pupil diameter or tool position, generate many videos, and check with an independently trained detector whether the edited property changes in lockstep with the graph; if it does not, the fine-grained-control claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the same scene graph representation that describes a surgical frame can be turned into a video-generation controller. Each node encodes a component's class, centroid, spatial spread, average optical flow, and average depth, and edges connect spatially adjacent components. Two graph encoders—one trained to reconstruct masked components locally, one trained contrastively to align graph embeddings with segmentation-mask embeddings globally—produce a conditioning embedding for a latent video diffusion model, with the first frame as an optional additional condition. The resulting system, SG2VID, is claimed to track ground-truth tool and anatomy size and movement, to synthesize the entry of new tools, and to follow the overall layout, while allowing the user to edit graph nodes—for example, shrinking the pupil across frames to simulate a rare irregularity—and generate the corresponding video. On the reported metrics it outperforms unconditional, text-conditioned, and trajectory-conditioned baselines, and extending a phase-recognition training set with its outputs improves that model.

Load-bearing premise

The entire pipeline depends on the automatically generated segmentation masks being accurate enough to serve as scene graph ground truth; if those masks are noisy or systematically wrong, the scene graphs, encoder training, and controllability evaluation all inherit that error.

Editorial extensions

If this is right

  • Surgical training simulators could be driven by editable graphs: changing a node's size, centroid, or appearance schedule changes the generated video accordingly.
  • Rare intra-operative events, such as sudden pupil contraction, can be synthesized by interpolating node attributes between two graphs, giving irregularity-detection systems examples they otherwise rarely see.
  • The model can mix a first frame from one video with scene graphs from another, which the paper uses to generate new training videos; if correct, this is a recipe for generative augmentation across patients.
  • Because the graph-only variant still tracks tool and anatomy movement, the graph alone carries most of the control, while appearance details absent from the graph are intentionally left free.
  • The reported phase-recognition improvement, with macro F1 rising from 0.794 to 0.805 and accuracy from 0.793 to 0.816 when synthetic videos are added to the training set, is a concrete downstream benefit if the central claim holds.

Reading between the lines

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

  • A natural extension not developed in the paper is to script long procedures by chaining graph edits over time, scheduling tool entry, movement, and exit as a sequence of graph states rather than interpolating a single property.
  • The paper's controllability check uses an object detector on real versus generated frames; a sharper test would regress each edited node property directly from the generated video and verify that the edit produces a proportional change, separating graph control from coincidental layout.
  • The two-encoder design—local reconstruction plus global contrastive alignment in segmentation space—could transfer to any video domain with object-level masks, making the claimed synthesis/control trade-off a testable general hypothesis.
  • Because the pipeline depends on automatically generated segmentation masks, an immediate experiment is to add controlled mask noise and measure how detection-based controllability scores degrade; the paper does not report this sensitivity.
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. The paper introduces SG2VID, a diffusion-based video synthesis model conditioned on scene graphs constructed from segmentation masks, with node features for class, centroid, spread, optical flow, and depth. It pre-trains two graph encoders (one local reconstruction, one global contrastive) and trains a latent video diffusion model with optional first-frame conditioning. Experiments on Cataract-1k, CATARACTS, and Cholec80 report FVD/FID/LPIPS improvements over baselines, a Mask R-CNN-based detection consistency check, qualitative edits for pupil contraction, cross-dataset generation on OphNet, and a downstream phase-recognition augmentation experiment showing improved macro F1 and accuracy.

Significance. If the controllability claim holds, SG2VID offers a useful new conditioning modality for surgical video generation, with practical potential for simulation and generative augmentation. The paper includes external quantitative anchors (FVD/FID, downstream MS-TCN++ phase recognition, cross-domain generation) and an ablation, which mitigate circularity concerns. The main weaknesses are the reliance on unvalidated SASVi segmentation masks and the lack of a direct quantitative controllability metric; these are substantial but fixable within the manuscript's scope.

major comments (4)
  1. [Section 2, paragraph 1] The scene graph construction rests entirely on SASVi-generated masks, because manual annotations cover only a small subset of frames. The manuscript reports no validation of SASVi on Cataract-1k, CATARACTS, or Cholec80. If SASVi systematically merges, splits, or mislabels components, the nodes users edit do not correspond to real tools or anatomy, and the claim of 'accurate control' would be about segmentation artifacts rather than the actual scene. Please add a quantitative evaluation of SASVi against the available manual annotations (e.g., the annotated subset used to train the Mask R-CNN in Section 3) and report per-class errors.
  2. [Section 3, Table 1] All quantitative results in Table 1 are reported as single numbers without error bars, confidence intervals, or significance tests. This matters because on CATARACTS the test split contains only 22 videos, and some pairwise differences are modest (e.g., SG2VID vs SG2VID-XIMG: FVD 523.8 vs 535.7, FID 40.9 vs 39.8). To support the claim that SG2VID 'outperforms previous methods' quantitatively, please report multiple seeds or bootstrap intervals and, where appropriate, paired significance tests.
  3. [Section 3, Table 1 (bottom)] The BB IoU/F1 evaluation compares Mask R-CNN outputs on generated sequences against those on real sequences. This measures whether the generated frames resemble real surgical frames in a detector's eyes, not whether each generated video adheres to its conditioning scene graph. It therefore cannot support the claim that graph edits transfer to the output. Please add a direct controllability metric, for example measuring the size or position of a specified component in generated frames against the edited graph parameters, or a quantitative comparison of edited versus unedited conditions.
  4. [Section 3, Controllability and Generalisation] The pupil-contraction demonstration is qualitative only. No measurement is provided that the pupil size actually decreases in the generated frames according to the edited graph, nor how accurately the interpolation of intermediate graphs is reflected in the video. Because fine-grained control is the paper's central contribution, this needs a quantitative evaluation of the edited attribute.
minor comments (5)
  1. [Section 2, Eq. (2)] The contrastive loss notation writes products between graphs and masks, but these are clearly meant to be embedding-space operations; please clarify the notation and define how positive and negative masks are constructed for each graph.
  2. [Section 3, Ablation] The ablation reports FVD/FID/LPIPS for local-only and global-only variants but does not define what these variants are precisely; please state which encoder is removed in each case and whether first-frame conditioning is used.
  3. [Section 3, Generative Augmentation] The mix-and-match procedure uses first-frames from one video and SGs from another, while phase annotations are taken from the graph-source video; please clarify whether the first-frame source could introduce a label bias and how this was controlled.
  4. [Section 3, Generalisation] The OphNet generalization result is shown qualitatively; please provide a quantitative comparison or a clearer protocol so that cross-dataset generalization can be assessed objectively.
  5. [Section 1, Contributions] The 'first diffusion-based SG-to-Video model' claim depends on the completeness of the cited related work; please ensure recent preprints on graph-conditioned video diffusion are cited and briefly compared.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: central claims rest on external benchmarks and independent annotations; same-group SASVi/SurGrID are dependencies, not reductions.

full rationale

The paper's central claims—SG-conditioned video synthesis, fine-grained control over component size/movement, and generative augmentation—are evaluated against external or held-out benchmarks: FVD/FID/LPIPS against public baselines, Mask R-CNN trained on manually annotated bounding-box subsets, video-wise held-out splits, a downstream CATARACTS phase-recognition task, and OphNet cross-domain first-frame generation. The scene graph is built from SASVi-generated masks, and the node encoding follows SurGrID; both are same-group citations. However, these are implementation dependencies rather than circular reductions: no equation or evaluation defines the predicted quantity in terms of the fitted quantity. The graph encoders are trained with reconstruction and contrastive objectives (Eqs. 1 and 2) that map graph embeddings to frames and masks, and the diffusion model is trained with a standard denoising objective; none of these losses presuppose the success claims. The controllability demonstrations are largely qualitative, which is an evaluation-strength limitation rather than a circularity. The Mask R-CNN alignment check compares generated videos with real videos using independently annotated bounding boxes, providing an external yardstick. Therefore, no load-bearing step reduces to its own input by construction.

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

No explicit fitted constants or newly invented physical entities appear in the paper. The learned network weights are not counted as free parameters. The load-bearing assumptions are all domain assumptions about the reliability of automatic masks, pretrained depth and flow models, and the detector-based evaluation proxy.

assumptions (4)
  • domain assumption SASVi-generated segmentation masks are sufficiently accurate to act as ground-truth scene structure for training the graph encoders and constructing user-facing scene graphs.
    Public datasets provide annotations for only a small subset of frames; the pipeline relies on SASVi [25] for all masks. Mask errors propagate into the SG conditioning and into the controllability evaluation.
  • domain assumption The contrastive objective over segmentation masks forces the global graph encoder to capture layout and ignore texture, so the graph embedding contains enough information for precise video synthesis.
    The paper states this property without a controlled test; it is central to why the graph embeddings can drive visual detail in the generated frames.
  • domain assumption Pretrained RAFT and MiDaS provide reliable per-component optical flow and depth estimates for temporal node features.
    Flow and depth are averaged over component regions and fed to the graph encoders; errors would distort the motion conditioning and undermine the claimed control over tool movement.
  • domain assumption Mask R-CNN trained on the small annotated subset is a valid proxy for measuring whether generated sequences adhere to the scene graph.
    The paper uses detector F1 and BB IoU on real versus generated frames to claim conditioning fidelity, but it does not validate the detector on the annotated subset or show that its errors are unbiased.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SG2VID: Scene Graphs Enable Fine-Grained Control for Video Synthesis." pith.science (2026). https://pith.science/paper/HE63WP2M

@misc{pith2026250603082,
  author       = {Pith},
  title        = {Pith review of: SG2VID: Scene Graphs Enable Fine-Grained Control for Video Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HE63WP2M}},
  note         = {Machine review of arXiv:2506.03082}
}
read the original abstract

Surgical simulation plays a pivotal role in training novice surgeons, accelerating their learning curve and reducing intra-operative errors. However, conventional simulation tools fall short in providing the necessary photorealism and the variability of human anatomy. In response, current methods are shifting towards generative model-based simulators. Yet, these approaches primarily focus on using increasingly complex conditioning for precise synthesis while neglecting the fine-grained human control aspect. To address this gap, we introduce SG2VID, the first diffusion-based video model that leverages Scene Graphs for both precise video synthesis and fine-grained human control. We demonstrate SG2VID's capabilities across three public datasets featuring cataract and cholecystectomy surgery. While SG2VID outperforms previous methods both qualitatively and quantitatively, it also enables precise synthesis, providing accurate control over tool and anatomy's size and movement, entrance of new tools, as well as the overall scene layout. We qualitatively motivate how SG2VID can be used for generative augmentation and present an experiment demonstrating its ability to improve a downstream phase detection task when the training set is extended with our synthetic videos. Finally, to showcase SG2VID's ability to retain human control, we interact with the Scene Graphs to generate new video samples depicting major yet rare intra-operative irregularities.

Figures

Figures reproduced from arXiv: 2506.03082 by the authors.

Figure 1
Figure 1. Surgical Video Synthesis using Scene Graphs (SGs) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Training Workflow of the Graph Encoders and Video Diffusion Model. first-frame x1. For computational efficiency, Rombach et al. [24] propose to per￾form the forward and backward process, p(zt−1|zt, c), in the latent space. So to apply the first-frame conditioning, we encode the first-frame z1 = Ex (x1) and replace it with the first-frame noise ϵ 1 . The model input is then constructed as ˆϵ = {z 1 , ϵ2 , ϵ3 , . . . … view at source ↗
Figure 3
Figure 3. SG-to-Video Generation. More samples are in the supplementary material. Qualitative Assessment [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Synthesisation of Pupil Contraction Irregularities and Generalisation on Unseen Images from the OphNet Dataset. More samples are in the supplementary material. Generative Augmentation on Surgical Phase Recognition: The qual￾itative results in [PITH_FULL_IMAGE:figures/…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 16 canonical work pages

  1. [1]

    MedIA52, 24–41 (2019)

    Al Hajj, H., Lamard, M., Conze, P.H., Roychowdhury, S., Hu, X., Maršalkait˙ e, G., Zisimopoulos, O., Dedmari, M.A., Zhao, F., Prellberg, J., et al.: Cataracts: Challenge on automatic tool annotation for cataract surgery. MedIA52, 24–41 (2019)

  2. [2]

    Brody, S., Alon, U., Yahav, E.: How attentive are graph attention networks? arXiv:2105.14491 (2021)

  3. [3]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650– 9660 (2021)

  4. [4]

    arXiv:2412.14018 (2024)

    Chen, T., Yang, S., Wang, J., Bai, L., Ren, H., Zhou, L.: Surgsora: Decoupled rgbd- flow diffusion model for controllable surgical video generation. arXiv:2412.14018 (2024)

  5. [5]

    arXiv:2408.14028 (2024)

    Cho, J., Schmidgall, S., Zakka, C., Mathur, M., Kaur, D., Shad, R., Hiesinger, W.: Surgen: Text-guided diffusion model for surgical video generation. arXiv:2408.14028 (2024)

  6. [6]

    SurGrID: Controllable Surgical Simulation via Scene Graph to Image Diffusion

    Frisch, Y., Sivakumar, S.K., Köksal, Ç., Böhm, E., Wagner, F., Gericke, A., Ghaz- aei, G., Mukhopadhyay, A.: Surgrid: Controllable surgical simulation via scene graph to image diffusion. arXiv:2502.07945 (2025)

  7. [7]

    arXiv:2312.06295 (2023)

    Ghamsarian, N., El-Shabrawi, Y., Nasirihaghighi, S., Putzgruber-Adamitsch, D., Zinkernagel, M., Wolf, S., Schoeffmann, K., Sznitman, R.: Cataract-1k: Cataract surgery dataset for scene segmentation, phase recognition, and irregularity detec- tion. arXiv:2312.06295 (2023)

  8. [8]

    In: ICCV

    He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV. pp. 2961–2969 (2017) 10 Authors Suppressed Due to Excessive Length

Show all 35 references
  1. [9]

    arXiv:2211.13221 (2022)

    He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation. arXiv:2211.13221 (2022)

  2. [10]

    NeurIPS33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS33, 6840–6851 (2020)

  3. [11]

    NeurIPS35, 8633–8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. NeurIPS35, 8633–8646 (2022)

  4. [12]

    In: ICCV Workshop

    Holm, F., Ghazaei, G., Czempiel, T., Özsoy, E., Saur, S., Navab, N.: Dynamic scene graph representation for surgical video. In: ICCV Workshop. pp. 81–87 (2023)

  5. [13]

    In: ECCV

    Hu, M., Xia, P., Wang, L., Yan, S., Tang, F., Xu, Z., Luo, Y., Song, K., Leitner, J., Cheng, X., et al.: Ophnet: A large-scale video benchmark for ophthalmic surgical workflow understanding. In: ECCV. pp. 481–500. Springer (2024)

  6. [14]

    In: MICCAI Workshop on Deep Generative Models

    Iliash, I., Allmendinger, S., Meissen, F., Kühl, N., Rückert, D.: Interactive genera- tion of laparoscopic videos with diffusion models. In: MICCAI Workshop on Deep Generative Models. pp. 109–118. Springer (2024)

  7. [15]

    American journal of ophthalmology157(6), 1190–1201 (2014)

    Kwong, A., Law, S.K., Kule, R.R., Nouri-Mahdavi, K., Coleman, A.L., Caprioli, J., Giaconi, J.A.: Long-term outcomes of resident-versus attending-performed primary trabeculectomy with mitomycin c in a united states residency program. American journal of ophthalmology157(6), 119...

  8. [16]

    In: MICCAI

    Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. In: MICCAI. pp. 230–240. Springer (2024)

  9. [17]

    IEEE PAMI45(6), 6647–6658 (2020)

    Li, S., Farha, Y.A., Liu, Y., Cheng, M.M., Gall, J.: Ms-tcn++: Multi-stage tempo- ral convolutional network for action segmentation. IEEE PAMI45(6), 6647–6658 (2020)

  10. [18]

    In: WACV

    Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: WACV. pp. 5404–5411 (2024)

  11. [19]

    IEEE TMI (2023)

    Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., Garcia, A., Okamoto, N., Mutter, D., Padoy, N.: Latent graph representations for critical view of safety assessment. IEEE TMI (2023)

  12. [20]

    Scientific reports11(1), 10945 (2021)

    Nair, A.G., Ahiwalay, C., Bacchav, A.E., Sheth, T., Lansingh, V.C., Vedula, S.S., Bhatt, V., Reddy, J.C., Vadavalli, P.K., Praveen, S., et al.: Effectiveness of simulation-based training for manual small incision cataract surgery among novice surgeons: a randomized controlled ...

  13. [21]

    In: ECCV

    Niu, M., Cun, X., Wang, X., Zhang, Y., Shan, Y., Zheng, Y.: Mofa-video: Con- trollable image animation via generative motion field adaptions in frozen image- to-video diffusion model. In: ECCV. pp. 111–128. Springer (2024)

  14. [22]

    IEEE PAMI44(3), 1623–1637 (2020)

    Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE PAMI44(3), 1623–1637 (2020)

  15. [23]

    arXiv:2402.04324 (2024)

    Ren, W., Yang, H., Zhang, G., Wei, C., Du, X., Huang, W., Chen, W.: Consisti2v: Enhancing visual consistency for image-to-video generation. arXiv:2402.04324 (2024)

  16. [24]

    In: CVPR

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022)

  17. [25]

    arXiv:2502.09653 (2025)

    Sivakumar, S.K., Frisch, Y., Ranem, A., Mukhopadhyay, A.: Sasvi-segment any surgical video. arXiv:2502.09653 (2025)

  18. [26]

    In: CVPR

    Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video gener- ator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626–3636 (2022)

  19. [27]

    arXiv:2407.08944 (2024) SG2VID: Scene Graphs Enable Fine-Grained Control for Video Synthesis 11

    Sun, W., You, X., Zheng, R., Yuan, Z., Li, X., He, L., Li, Q., Sun, L.: Bora: Biomedical generalist video generation model. arXiv:2407.08944 (2024) SG2VID: Scene Graphs Enable Fine-Grained Control for Video Synthesis 11

  20. [28]

    In: ECCV

    Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: ECCV. pp. 402–419. Springer (2020)

  21. [29]

    IEEE TMI36(1), 86–97 (2016)

    Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., De Mathelin, M., Padoy, N.: Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE TMI36(1), 86–97 (2016)

  22. [30]

    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:1812.01717 (2018)

  23. [31]

    arXiv:2411.01647 (2024)

    Wang, Z., Zhang, L., Wang, L., Zhu, M., Zhang, Z.: Optical flow representation alignment mamba diffusion model for medical video generation. arXiv:2411.01647 (2024)

  24. [32]

    Eye36(8), 1529–1531 (2022)

    Wood, T.C., Maqsood, S., Sancha, W., Saunders, A., Lockington, D., Nanavaty, M.A., Rajak, S.: Principles of simulation and their role in enhancing cataract surgery training. Eye36(8), 1529–1531 (2022)

  25. [33]

    In: MICCAI

    Yeganeh, Y., Lazuardi, R., Shamseddin, A., Dari, E., Thirani, Y., Navab, N., Far- shad, A.: Visage: Video synthesis using action graphs for surgery. In: MICCAI. pp. 146–156. Springer (2024)

  26. [34]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

  27. [35]

    In: MICCAI

    Zhou, X., Huang, Y., Xue, W., Dou, H., Cheng, J., Zhou, H., Ni, D.: Heartbeat: To- wards controllable echocardiography video synthesis with multimodal conditions- guided diffusion models. In: MICCAI. pp. 361–371. Springer (2024)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.