HieraSurg: Hierarchy-Aware Diffusion Model for Surgical Video Generation

Azade Farshad; Diego Biagini; Nassir Navab

REVIEW 4 major objections 5 minor 25 references

HieraSurg: Hierarchy-Aware Diffusion Model for Surgical Video Generation

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 HieraSurg generates realistic surgical videos by first predicting future segmentation maps and then rendering them into video, beating prior models on cholecystectomy clips.

desk verdict A genuinely novel two-stage surgical video generator whose plausible architecture is undercut by an evaluation built on two test videos and imported baselines. read the letter →

arxiv 2506.21287 v1 pith:JA5TPFCC submitted 2025-06-26 cs.CV

Diego Biagini , Nassir Navab , Azade Farshad This is my paper

classification cs.CV

keywords surgicalvideogenerationdiffusionmodelspanopticsegmentationactiontripletscholecystectomysemanticconditioningsynthesislatent

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

HieraSurg claims that surgical video generation can be made controllable and anatomically plausible by splitting the problem into two learned stages that mirror the hierarchy of a surgical scene. The first stage, S2M, takes a surgical phase, action triplets (instrument–action–target), and an initial frame, and predicts the evolution of a panoptic segmentation map of the operating field. The second stage, M2V, renders those segmentation maps into photorealistic video frames, so the final video is forced to respect the predicted semantic layout. On cholecystectomy clips from Cholec80 and CholecT45, the authors report better fidelity (FVD, FID) and much stronger adherence to tool positions than prior surgical video generators. They also show the pipeline can run at 8 FPS and can be used either end-to-end or as a purely segmentation-conditioned renderer when ground-truth maps are available.

What carries the argument

The load-bearing object is the pair of hierarchical diffusion models themselves: HieraSurg-S2M, a latent DiT that generates $F$ future panoptic segmentation maps from an initial frame $\mathbf{y}_1$, its segmentation $\mathbf{y}^{\mathrm{seg}}_1$, phase, and triplet text encodings (via PeskaVLP), using a temporally dense latent encoding that preserves per-frame detail; and HieraSurg-M2V, a latent DiT that takes a segmentation map sequence $\mathbf{c}$ and the initial frame and injects map tokens via cross-attention with sinusoidal temporal positional embeddings. The two models share the CogVideoX-2B prior and are trained sequentially, with S2M's discrete outputs serving as M2V's conditioning input. The framework also includes an automated SAM2/RADIO labeling pipeline that produces panoptic maps from unlabeled surgical videos, which is what makes training data abundant enough for the second stage.

What would settle it

Train or fine-tune HieraSurg-M2V and HieraSurg-S2M on a subset of hand-annotated segmentation maps from a dataset such as CholecSeg8k or Endoscapes and evaluate on the same held-out videos. If the auto-labeled pseudo-labels are the bottleneck, the hand-annotated model should yield measurably lower FVD and higher detector agreement (Hit Rate Real, MIoU) than the model trained on the paper's pipeline; if the gap is negligible, the pseudo-label noise is not limiting.

Watch

Extended reading notes

Core claim

The central claim is that a pair of coupled latent diffusion models, built on the CogVideoX-2B architecture, can generate surgical videos whose short-term scene evolution is semantically consistent with the actual procedure. HieraSurg-S2M predicts a sequence of future panoptic segmentation maps conditioned on the current frame and on phase and triplet textual embeddings, and its outputs are discretized with K-Means clustering to yield clean maps. HieraSurg-M2V then converts any such map sequence, together with the initial frame, into high-quality video by injecting the map tokens into every transformer block through cross-attention. The experiments show that GAN- and diffusion-based baselines, including a ControlNet variant conditioned on Canny edges extracted from the maps, produce worse fidelity metrics and much lower detector agreement; the full pipeline also outperforms the unconditional CogVideoX baseline. The paper argues this validates the pyramid-of-abstraction design: generation should be split vertically, with one model handling coarse semantic structure and a second handling texture and fine visual detail.

Load-bearing premise

The assumption that the automatically generated SAM2/RADIO panoptic annotations are reliable enough to serve as ground-truth supervision for the first-stage model; if these pseudo-labels carry systematic errors, both the segmentation prediction and the final video quality degrade despite the reported metrics.

Editorial extensions

If this is right

The full HieraSurg pipeline can generate a completely novel video from only an initial frame plus phase and triplet information, while the M2V component alone can render a known scene from its segmentation maps.
Because the model is conditioned on explicit semantic maps, the generated video's layout can be prescribed in advance, enabling tool-aware or structure-aware synthetic data generation.
The 8 FPS variant shows that the approach scales to higher frame rates without a drop in quality, which is needed for realistic surgical simulation.
The reported detector-agreement metric indicates that generated videos contain recognizable surgical tools at plausible locations, which is what a simulator for training perception systems would require.

Reading between the lines

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

A natural extension is to feed the model's predicted segmentation maps back into S2M as the new initial frame, creating a recursive video generator that could in principle synthesize arbitrarily long procedures; the paper only evaluates fixed 16- or 6-second clips, so long-horizon drift remains an open question.
The same two-stage hierarchy could translate to other constrained environments with known semantic structure, such as cardiac catheterization or endoscopy, where phase annotations and instrument-action labels exist.
The dependence on pseudo-labels suggests a promising test: measure how per-entity tracking accuracy of the SAM2/RADIO pipeline correlates with final video quality; the paper does not report this intermediate error.
If combined with an LLM or a surgical phase-prediction model, the S2M stage could be used for conditional forecasting of 'what happens next' under alternative actions, turning the generator into a queryable simulator.

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.

Desk Editor's Note

read the letter

The thing to know about HieraSurg is that the core idea is actually new: splitting surgical video generation into a semantic-to-map stage and a map-to-video stage, with panoptic segmentation as the intermediate and phase/triplet text conditioning injected into both. That is a clean, sensible decomposition, and it is not in Endora or VISAGE. The automatic SAM2/RADIO labeling pipeline is also a practical contribution, and the ablation showing that PeskaVLP text embeddings beat learned label embeddings is honest and informative. The paper reads like the authors understand their domain.

The problem is that the central claim—"significantly outperforms prior work"—is not supported by the evidence. The quantitative test set is two videos. FVD and FID on two surgeries are point estimates with huge variance, and no error bars or repeated-seed runs are reported. On top of that, the VISAGE numbers are imported from its original paper, likely on a different split, so the comparison is not apples-to-apples. The detector-agreement metric is custom and not validated. With those issues, the reported margins, like FVD 312 vs. 443, could easily be noise. The reader's concern about pseudo-label quality is real but secondary: even imperfect segmentation maps can produce good FVD because the M2V stage can learn texture rendering, and the paper's own detector metric partially addresses semantic adherence. The evaluation validity problem undermines every quantitative claim and has no mitigating evidence.

What the paper does well beyond the architecture: the ablations are thoughtful, the limitation paragraph in the conclusion is candid about reliance on segmentation quality, and the qualitative results look believable. But no code or data is released, which makes the evaluation even harder to assess.

Who is this for? Researchers in surgical data science and medical video generation who want controllable generation for augmentation or simulation. They will get a solid framework and a useful intermediate representation, but they should not trust the quantitative superiority claims as they stand. This deserves a serious referee, not a desk reject—the idea is strong enough to warrant a major revision with a proper test set, matched baselines, and uncertainty quantification.

Referee Report

4 major / 5 minor

Summary. The paper proposes HieraSurg, a two-stage hierarchy-aware diffusion framework for surgical video generation. The first stage, HieraSurg-S2M, takes an initial frame, its segmentation map, a surgical phase, and action triplets, and predicts a sequence of future panoptic segmentation maps. The second stage, HieraSurg-M2V, conditions a CogVideoX-2B latent diffusion transformer on those segmentation maps and the initial frame to synthesize the final video. Because manual segmentation annotations are scarce, the authors introduce an automated labeling pipeline based on SAM2 and RADIO features. Experiments on Cholec80 and CholecT45 report FVD, FID, SSIM, and a YOLOv8-based detector-agreement metric, and the paper claims that HieraSurg significantly outperforms prior work such as Endora, VISAGE, and a ControlNet variant.

Significance. The architectural idea is attractive: decomposing surgical video generation along a semantic hierarchy, with a dedicated model for coarse future evolution and a second model for texture rendering, is a sensible and potentially useful direction. The automated SAM2/RADIO labeling pipeline addresses a real annotation bottleneck, and the paper includes a useful ablation study of latent encoding and conditioning choices. The detector-based agreement metric is a reasonable attempt to measure semantic adherence. However, the main quantitative claim of 'significantly outperforms prior work' is not supported by the current evaluation: the test set is only two videos, no uncertainty estimates are given, and one baseline is imported from another paper under a different protocol. The pseudo-label reliability concern is secondary but also needs attention. As presented, the empirical contribution is not yet established, though the method itself is plausible and worth revising.

major comments (4)

[Section 3, Datasets and preprocessing; Table 1] The quantitative evaluation rests on a test set of only two videos. FVD and FID are computed as point estimates over 1024 generated samples against the same number of real data points drawn from those two videos. With only two underlying surgical videos, the real-data distribution is extremely poorly estimated, and FVD/FID have high variance. No confidence intervals, bootstrap estimates, repeated-seed runs, or significance tests are provided. Consequently, the margins in Table 1 (e.g., FVD 312.4 vs. 443.1 for S2M+M2V vs. CogVideoX) cannot be judged statistically significant. This is the central claim of the paper, so the evaluation must be strengthened: either a larger held-out set, or explicit uncertainty quantification and a matched sampling protocol.
[Table 1, footnote] The comparison is not apples-to-apples. The VISAGE row is imported from its original paper ('* Values reported on the paper'), likely using a different test split and protocol, while Endora is evaluated on this split but with no documented conditioning or sample-size matching. Because the baselines are not rerun under identical conditions, the reported improvements over VISAGE and Endora are not reliable. The authors should rerun all baselines under the same test split, conditioning inputs, and sample size, or explicitly state that cross-paper comparisons are indicative only.
[Section 3, Evaluation Setting; Table 1] The YOLOv8-based detector agreement metric is computed on only two videos and depends on several undocumented choices: the detector training set, the tool classes, the bounding-box matching threshold, and the handling of unmatched detections. As a custom metric, it needs a precise description of the matching procedure and some form of validation (e.g., comparison with manual annotations or a standard detection benchmark) before the reported Hit Rates and MIoU values can be interpreted. Without that, the claim of 'fine-grained adherence' is only weakly supported.
[Section 2, Video Segmentation] The automated SAM2/RADIO labeling pipeline generates all training segmentation maps, but its output is never quantitatively validated against manual annotations (e.g., CholecSeg8k or Endoscapes). The RADIO feature distance threshold and the K-Means cluster count are free parameters, and no sensitivity analysis is reported. If the pseudo-labels contain systematic errors in tool or organ tracking, both S2M training and the final S2M+M2V metrics would be affected. The paper should report agreement against existing segmentation benchmarks or a manually annotated sample to establish the reliability of this intermediate supervision.

minor comments (5)

[Table 1 caption] The caption contains a typo: 'Quantitive' should be 'Quantitative'.
[Figure 2 caption] The caption says 'first 6 seconds' while Table 1 reports 16-second videos at 1 FPS; please clarify whether the figure shows a shortened clip or a different setting.
[Table 2] The ablation design is not fully factorial: row 1 removes conditioning while retaining the dense latent, row 2 removes the dense latent while keeping PeskaVLP conditioning, and row 3 uses label embedding. The individual contributions of dense latent encoding and conditioning method are therefore confounded in some comparisons. A complete factorial ablation or a clearer presentation of the isolated factors would strengthen the conclusions.
[Section 2.1, HieraSurg-M2V] The attention equation is not fully defined: the meaning of the split operation, the dimensions of H_i and H_i^seg, and the role of T in 'split(WoZ, T)' should be stated explicitly for reproducibility.
[Section 3, Evaluation Setting] The PeskaVLP feature extractor used for FID is a model co-developed by one of the current authors; please state explicitly whether it is frozen and was not trained on the test split, since otherwise the metric could be seen as favoring the proposed pipeline.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the two-stage pipeline is an empirical system evaluated on held-out videos, and the only self-citation is a fixed pretrained encoder that is not fitted to the target result.

full rationale

HieraSurg is an empirical engineering contribution with no analytic derivation chain that could reduce to its inputs. HieraSurg-S2M is trained to predict future segmentation maps from an initial frame, phase, and triplet information, and HieraSurg-M2V is trained to render videos from segmentation maps and an initial frame. The full pipeline is evaluated on two held-out test videos that were not used for training, so the S2M-to-M2V chain is not fitted to the test outputs. The paper's central claim, quantitative superiority on FVD/FID, is thus a measured empirical outcome rather than a tautology. The only self-citation is PeskaVLP [21], co-authored by N. Navab, which is used in two places: as the best-performing phase/triplet text encoder in S2M (Table 2) and as one feature extractor for a reported FID variant. PeskaVLP is a fixed pretrained surgical video-language encoder, not a parameter fitted in this paper, and the best S2M configuration is also validated with standard FVD/FID and compared against a Label-Embedding ablation. Consequently, the PeskaVLP usage is not load-bearing circularity. The Conclusion openly states that predictive capacity is "heavily reliant on the quality of the segmentation maps," and Section 3 notes that only "two videos are held out as a test set"; these are genuine evaluation-validity concerns about statistical support for the claimed margins, but they are not instances of self-definitional or construction-level circularity. No quoted equation or fitted value is shown to equal another by construction, and no external benchmark is silently replaced by a self-derived quantity. Therefore, no specific circular step meets the required evidence bar, and the appropriate finding is a low score with no flagged circular steps.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the domain assumption that automatically generated pseudo-labels from SAM2/RADIO are accurate enough for training, and on the correctness of phase and triplet annotations. Free parameters include heuristic thresholds in the labeling pipeline and the K-Means cluster count.

free parameters (2)

RADIO feature distance threshold for object matching
Used in the automatic segmentation pipeline to decide whether two detections across frames are the same object; threshold is chosen by hand without sensitivity analysis.
K-Means cluster count for segmentation map discretization = determined by elbow method
Used to quantize the continuous output of S2M into discrete segmentation maps; the elbow method is a heuristic and the resulting count is data-dependent.

assumptions (3)

domain assumption SAM2 with RADIO features produces panoptic segmentation maps accurate enough to serve as training supervision
The entire training of S2M depends on these automatically generated labels; no manual verification or quality analysis is provided (Section 2, Video Segmentation).
domain assumption Phase and action triplet annotations from Cholec80 and CholecT45 are correct and sufficient to predict future scene evolution
S2M conditions on these labels to predict future segmentation; errors or incompleteness would directly reduce prediction accuracy (Section 2.1).
standard math Latent diffusion and VAE assumptions hold as in CogVideoX-2B
The paper relies on standard diffusion model theory and the pretrained CogVideoX-2B architecture without modification (Section 2, Diffusion Models).

how reviews work

0 comments

Cite this review

Pith. "Pith review of HieraSurg: Hierarchy-Aware Diffusion Model for Surgical Video Generation." pith.science (2026). https://pith.science/paper/JA5TPFCC

@misc{pith2026250621287,
  author       = {Pith},
  title        = {Pith review of: HieraSurg: Hierarchy-Aware Diffusion Model for Surgical Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JA5TPFCC}},
  note         = {Machine review of arXiv:2506.21287}
}

read the original abstract

Surgical Video Synthesis has emerged as a promising research direction following the success of diffusion models in general-domain video generation. Although existing approaches achieve high-quality video generation, most are unconditional and fail to maintain consistency with surgical actions and phases, lacking the surgical understanding and fine-grained guidance necessary for factual simulation. We address these challenges by proposing HieraSurg, a hierarchy-aware surgical video generation framework consisting of two specialized diffusion models. Given a surgical phase and an initial frame, HieraSurg first predicts future coarse-grained semantic changes through a segmentation prediction model. The final video is then generated by a second-stage model that augments these temporal segmentation maps with fine-grained visual features, leading to effective texture rendering and integration of semantic information in the video space. Our approach leverages surgical information at multiple levels of abstraction, including surgical phase, action triplets, and panoptic segmentation maps. The experimental results on Cholecystectomy Surgical Video Generation demonstrate that the model significantly outperforms prior work both quantitatively and qualitatively, showing strong generalization capabilities and the ability to generate higher frame-rate videos. The model exhibits particularly fine-grained adherence when provided with existing segmentation maps, suggesting its potential for practical surgical applications.

Figures

Figures reproduced from arXiv: 2506.21287 by the authors.

**Figure 1.** HieraSurg Pipeline. Left: we take the hierarchical representation of a surgical scene, Right: the inference pipeline and components of HieraSurg. S2M takes an initial segmentation map and predicts the future evolution of the surgical scene by considering phase and triplet information. Finally, the output of S2M is fed into M2V to generate a video according to the predicted segmentation. models as predictive tools f… view at source ↗

**Figure 2.** Visual comparison of the first 6 seconds generated by different models. For the conditional models, the input segmentation maps are given. Evaluation Setting We evaluate the visual quality of the generated videos using Frechet Video Distance (FVD) [22], Frechet Inception Distance (FID) [23] and FID using PeskaVLP [21] as a feature extractor. The reconstruction ability is evaluated using the Structural Similarity In… view at source ↗

**Figure 3.** Example output of the full HieraSurg pipeline, comparing the usage of M2V with the ground truth segmentation to when providing the S2M output to it [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

[1]

Stable video diffusion: Scal- inglatentvideodiffusionmodelstolargedatasets,

A. Blattmann, T. Dockhorn, S. Kulal,et al., “Stable video diffusion: Scal- inglatentvideodiffusionmodelstolargedatasets,” arXiv preprint arXiv:2311.15127, 2023

arXiv 2023
[2]

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

Z. Yang, J. Teng, W. Zheng, et al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” inThe Thirteenth International Con- ference on Learning Representations, 2025

work page 2025
[3]

W. Kong, Q. Tian, Z. Zhang,et al., HunyuanVideo: A systematic frame- work for large video generative models, Jan. 17, 2025

work page 2025
[4]

How far is video generation from world model:Aphysicallawperspective,

B. Kang, Y. Yue, R. Lu, et al., “How far is video generation from world model:Aphysicallawperspective,” arXiv preprint arXiv:2411.02385,2024

arXiv 2024
[5]

Using diffusion models to gener- ate synthetic labeled data for medical image segmentation,

D. G. Saragih, A. Hibi, and P. N. Tyrrell, “Using diffusion models to gener- ate synthetic labeled data for medical image segmentation,”International journal of computer assisted radiology and surgery, vol. 19, no. 8, pp. 1615– 1625, 2024

work page 2024
[6]

Endora: Video generation models as en- doscopy simulators,

C. Li, H. Liu, Y. Liu, et al., “Endora: Video generation models as en- doscopy simulators,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024, M. G. Linguraru, Q. Dou, A. Feragen,et al., Eds., Cham: Springer Nature Switzerland, 2024, pp. 230–240

work page 2024
[7]

Emerging properties in self- supervised vision transformers,

M. Caron, H. Touvron, I. Misra, et al., “Emerging properties in self- supervised vision transformers,” in Proceedings of the IEEE/CVF inter- national conference on computer vision, 2021, pp. 9650–9660

work page 2021
[8]

VISAGE: Video syn- thesis using action graphs for surgery,

Y. Yeganeh, R. Lazuardi, A. Shamseddin, et al., “VISAGE: Video syn- thesis using action graphs for surgery,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024 Workshops, M. E. Celebi, M. Reyes, Z. Chen, and X. Li, Eds., Cham: Springer Nature Switzerland, 2025, pp. 146–156

work page 2024

Show all 25 references

[9]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, et al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning, PmLR, 2021, pp. 8748–8763

2021
[10]

N. Ravi, V. Gabeur, Y.-T. Hu,et al., SAM 2: Segment anything in images and videos, Oct. 28, 2024

2024
[11]

EndoNet: A deep architecture for recognition tasks on la- paroscopic videos,

A. P. Twinanda, S. Shehata, D. Mutter, J. Marescaux, M. de Mathelin, and N. Padoy, “EndoNet: A deep architecture for recognition tasks on la- paroscopic videos,”IEEE Transactions on Medical Imaging, vol. 36, no. 1, pp. 86–97, Jan. 2017, Conference Name: IEEE Transactions on Me...

2017
[12]

Rendezvous: Attention mecha- nisms for the recognition of surgical action triplets in endoscopic videos,

C. I. Nwoye, T. Yu, C. Gonzalez,et al., “Rendezvous: Attention mecha- nisms for the recognition of surgical action triplets in endoscopic videos,” Medical Image Analysis, vol. 78, p. 102433, May 1, 2022

2022
[13]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2022, pp. 10684–10695. 10 Diego Biagini, Nassir N...

2022
[14]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, vol. 33, Curran Associates, Inc., 2020, pp. 6840–6851

2020
[15]

Hong, C.-L

W.-Y. Hong, C.-L. Kao, Y.-H. Kuo, J.-R. Wang, W.-L. Chang, and C.-S. Shih,CholecSeg8k: A semantic segmentation dataset for laparoscopic chole- cystectomy based on cholec80, Dec. 23, 2020

2020
[16]

The endoscapes dataset for surgical scene segmentation, object detection, and critical view of safety as- sessment:Officialsplitsandbenchmark,

A. Murali, D. Alapatt, P. Mascagni, et al., “The endoscapes dataset for surgical scene segmentation, object detection, and critical view of safety as- sessment:Officialsplitsandbenchmark,” arXiv preprint arXiv:2312.12429, 2023

2023 arXiv
[17]

Critical view of safety assessment in laparoscopic cholecystectomy via segment anything model,

Y. Li, H. Ling, I. V. Ramakrishnan, P. Prasanna, A. Sasson, and H. Gupta, “Critical view of safety assessment in laparoscopic cholecystectomy via segment anything model,” in2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)...

2024
[18]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi,et al., “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

2023
[19]

AM-RADIO: Agglomerative vision foundation model reduce all domains into one,

M. Ranzinger, G. Heinrich, J. Kautz, and P. Molchanov, “AM-RADIO: Agglomerative vision foundation model reduce all domains into one,” pre- sented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 12490–12500

2024
[20]

Scalable diffusion models with transformers,

W. Peebles and S. Xie, “Scalable diffusion models with transformers,” pre- sented at the Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

2023
[21]

Procedure-aware surgi- cal video-language pretraining with hierarchical knowledge augmentation,

K. Yuan, V. Srivastav, N. Navab, and N. Padoy, “Procedure-aware surgi- cal video-language pretraining with hierarchical knowledge augmentation,” presented at the The Thirty-eighth Annual Conference on Neural Informa- tion Processing Systems, Nov. 6, 2024

2024
[22]

FVD: A new metric for video generation,

T. Unterthiner, S. v. Steenkiste, K. Kurach, R. Marinier, M. Michalski, and S. Gelly, “FVD: A new metric for video generation,” Apr. 19, 2019

2019
[23]

Gans trained by a two time-scale update rule converge to a local nash equilibrium,

M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, vol. 30, 2017

2017
[24]

YOLOv8: A novel object detection algorithm withenhancedperformanceandrobustness,

R. Varghese and S. M., “YOLOv8: A novel object detection algorithm withenhancedperformanceandrobustness,” in 2024 International Confer- ence on Advances in Data Engineering and Intelligent Computing Systems (ADICS), Apr. 2024, pp. 1–6

2024
[25]

Adding conditional control to text- to-imagediffusionmodels,

L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text- to-imagediffusionmodels,” presentedattheProceedingsoftheIEEE/CVF International Conference on Computer Vision, 2023, pp. 3836–3847

2023

Pith tools

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

[1] [1]

Stable video diffusion: Scal- inglatentvideodiffusionmodelstolargedatasets,

A. Blattmann, T. Dockhorn, S. Kulal,et al., “Stable video diffusion: Scal- inglatentvideodiffusionmodelstolargedatasets,” arXiv preprint arXiv:2311.15127, 2023

arXiv 2023

[2] [2]

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

Z. Yang, J. Teng, W. Zheng, et al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” inThe Thirteenth International Con- ference on Learning Representations, 2025

work page 2025

[3] [3]

W. Kong, Q. Tian, Z. Zhang,et al., HunyuanVideo: A systematic frame- work for large video generative models, Jan. 17, 2025

work page 2025

[4] [4]

How far is video generation from world model:Aphysicallawperspective,

B. Kang, Y. Yue, R. Lu, et al., “How far is video generation from world model:Aphysicallawperspective,” arXiv preprint arXiv:2411.02385,2024

arXiv 2024

[5] [5]

Using diffusion models to gener- ate synthetic labeled data for medical image segmentation,

D. G. Saragih, A. Hibi, and P. N. Tyrrell, “Using diffusion models to gener- ate synthetic labeled data for medical image segmentation,”International journal of computer assisted radiology and surgery, vol. 19, no. 8, pp. 1615– 1625, 2024

work page 2024

[6] [6]

Endora: Video generation models as en- doscopy simulators,

C. Li, H. Liu, Y. Liu, et al., “Endora: Video generation models as en- doscopy simulators,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024, M. G. Linguraru, Q. Dou, A. Feragen,et al., Eds., Cham: Springer Nature Switzerland, 2024, pp. 230–240

work page 2024

[7] [7]

Emerging properties in self- supervised vision transformers,

M. Caron, H. Touvron, I. Misra, et al., “Emerging properties in self- supervised vision transformers,” in Proceedings of the IEEE/CVF inter- national conference on computer vision, 2021, pp. 9650–9660

work page 2021

[8] [8]

VISAGE: Video syn- thesis using action graphs for surgery,

Y. Yeganeh, R. Lazuardi, A. Shamseddin, et al., “VISAGE: Video syn- thesis using action graphs for surgery,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024 Workshops, M. E. Celebi, M. Reyes, Z. Chen, and X. Li, Eds., Cham: Springer Nature Switzerland, 2025, pp. 146–156

work page 2024