Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos

T0 review · 4 major / 6 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 With 55,803 captioned medical clips, MedGen rivals Sora on medical accuracy

desk verdict A genuinely useful medical video dataset, but the "leading open-source model" claim needs a stronger evaluation before it can be trusted. read the letter →

arxiv 2507.05675 v1 pith:O2ARR3XD submitted 2025-07-08 cs.CV cs.AI

Rongsheng Wang , Junying Chen , Ke Ji , Zhenyang Cai , Shunian Chen , Yunjin Yang , Benyou Wang This is my paper
classification cs.CVcs.AI
keywords medicalvideogenerationMedCap-55Ktext-to-videolatentdiffusionLoRAfine-tuningcaptioningdataaugmentationMed-VBench
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 argues that the reason general video generators produce anatomically wrong or implausible medical clips is missing domain data, not missing model capacity. It introduces MedVideoCap-55K, a dataset of 55,803 curated medical video clips paired with brief and detailed text captions, and uses it to fine-tune the open HunyuanVideo model with LoRA, producing MedGen. The authors report that MedGen leads all open-source baselines and approaches commercial systems on their Med-VBench benchmark, which adapts VBench to 200 medical prompts, and that doctor evaluators rate it favorably on text alignment, medical accuracy, and visual quality. If correct, the work establishes a scalable recipe for domain-specialized video generation and provides a reusable resource for medical training, simulation, and data augmentation.

What carries the argument

The load-bearing mechanism is the pairing of granular text captions with medical video clips. The construction pipeline filters 25 million YouTube videos down to 55,803 clips through a medical keyword and embedding classifier, per-frame medical classification, temporal consistency checks, and quality filters (black borders, OCR subtitles, aesthetic scores, and Dover technical scores), then uses a multimodal LLM to write both brief and detailed captions from eight sampled frames plus title, description, and transcript. Fine-tuning a latent video diffusion model on these caption-clip pairs is what converts general visual priors into medically grounded generation, and Med-VBench reuses VBench's hierarchical metrics on 200 medical prompts to measure that grounding.

What would settle it

Have independent clinicians write a held-out set of prompts from textbooks and clinical guidelines, never expose those prompts or their captions to MedVideoCap-55K, and compare MedGen against HunyuanVideo and Sora on medical accuracy; if the advantage shrinks to the base model's level, the claim that caption-rich medical pretraining transfers generally is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that domain-specific data, not model architecture, is the bottleneck for medical video generation. It constructs MedVideoCap-55K, 55,803 curated clips from public medical videos, each paired with brief and detailed captions produced by a multimodal LLM from sampled frames plus the video's title, description, and transcript, then fine-tunes the open HunyuanVideo model with LoRA on these caption-clip pairs to produce MedGen. The authors report that MedGen outperforms 15 open-source video generators on Med-VBench and approaches Sora, Kling, Pika, and Hailuo, while three doctor evaluators rate it best or tied on text alignment, medical accuracy, and visual quality for most tested prompts. A transfer experiment with Mochi-1-preview shows that fine-tuning on MedVideoCap-55K also improves another base model on warping error, image quality, and subject consistency.

Load-bearing premise

The 200 Med-VBench test prompts were authored by the same group that built the training captions, and the paper does not show that these prompts are independent of the MedVideoCap-55K caption distribution; if they overlap with training, MedGen's measured lead may reflect benchmark familiarity rather than general medical video ability.

Editorial extensions

If this is right

  • Generalist text-to-video models can be specialized to medicine with modest compute (LoRA fine-tuning on eight GPUs), making domain adaptation more a data problem than a model-size problem.
  • MedVideoCap-55K transfers across base models: the same fine-tuning setup improves Mochi-1-preview on warping error, image quality, and subject consistency.
  • MedGen-generated clips can augment small labeled medical video datasets, lifting F1 by up to 15.3 points on HyperKvasir and 11.7 on SurgVisDom relative to training on the original data alone.
  • Open-source medical video generation can reach commercial-level scores on Med-VBench, a VBench-derived benchmark that omits aesthetic quality because it is less relevant to medical content.

Reading between the lines

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

  • If the 200 Med-VBench prompts are genuinely independent of the training captions, medical video generation may be closer to a data-scaling problem than a modeling problem, and a similar captioning pipeline could transfer to other high-stakes domains such as radiology reporting or surgical rehearsal.
  • The paper's own transfer experiment hints at a diversity ceiling: fine-tuning on 55K clips slightly lowered background consistency and motion smoothness, suggesting that mixing general video data into the training set would preserve broad visual priors while adding medical grounding.
  • A stricter evaluation would separate prompt-caption overlap from genuine domain knowledge by testing on procedure names and anatomical terms that are rare in YouTube titles and descriptions, which would clarify whether the measured gains reflect learned medical semantics or benchmark familiarity.
  • For clinical use, synthetic medical footage raises unresolved ethical and safety questions; the authors restrict the dataset to research purposes, but the same technology could generate convincing misinformation if released without controls.
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 / 6 minor

Summary. The paper introduces MedVideoCap-55K, a large-scale (55,803 clips) medical video dataset with detailed captions, and MedGen, a medical video generation model obtained by fine-tuning HunyuanVideo with LoRA on this dataset. The authors claim leading performance among open-source models and competitive performance with commercial systems, evaluated on a medical-adapted VBench (Med-VBench), a human expert study, VideoScore, and a downstream data-augmentation experiment. The paper also presents a transferability experiment fine-tuning Mochi-1-preview on the dataset.

Significance. If the claims hold, this is a valuable contribution: MedVideoCap-55K is, to the best of my knowledge, the first large-scale, captioned medical video dataset for text-to-video generation, and the authors release code and data. The transferability experiment (Section 4.3) is a good check that the dataset benefits a second base model, and the use of an external benchmark (VideoScore) plus human evaluation with reported inter-rater agreement are strengths. However, the central performance claim depends on an author-constructed benchmark whose prompts may overlap with the training data, and the baseline set omits the domain-specific medical video generators cited in the Related Work. These issues need to be resolved before the headline claims can be accepted.

major comments (4)
  1. [Section 4.1 and Appendix A.3] The Med-VBench evaluation uses 200 author-designed prompts, but the paper provides no evidence that these prompts are disjoint from the MedVideoCap-55K captions used to train MedGen. Since both the dataset captions and the evaluation prompts are derived from the same broad YouTube medical video categories (clinical practice, teaching, imaging, animation), the benchmark may be evaluating MedGen on its training distribution while the baselines see those prompts for the first time. This would inflate the reported margin and undermine the 'leading among open-source models' claim. Please report a quantitative overlap analysis (e.g., embedding similarity between the 200 prompts and the training captions) and, ideally, evaluate on a held-out prompt set created independently of the dataset construction process.
  2. [Section 4.2, Table 2 and Section 6] The main comparison omits the domain-specific medical video generators cited in the Related Work (Bora, MedSora, Endora, Surgen, SurgSora). As a result, the claim of 'leading performance among open-source models' is only supported against general text-to-video baselines, not against existing medical-domain generators. Either include these models in the benchmark (where possible, e.g., Bora and SurgSora have open checkpoints) or explicitly limit the claim to general-purpose open-source models.
  3. [Section 4.1 and Table 2] The Med-VBench total score is recomputed after dropping the VBench aesthetic dimension, but the paper does not report the original VBench total or a sensitivity analysis showing how the ranking changes when the aesthetic dimension is included. Since the authors define the prompt set and the metric aggregation, the benchmark is at least partly self-adapted; reporting only the post-hoc total without the unmodified version makes it difficult for readers to assess how much of the improvement is driven by the choice of omitted dimension.
  4. [Section 5.1, Table 3] The downstream data-augmentation results are reported as single-run deltas without error bars, confidence intervals, or significance tests. Given the small base accuracies (e.g., HyperKvasir ACC 44.44, SurgVisDom ACC 14.29), the claimed improvements (up to +15.3 F1) may be within run-to-run variation. Please provide variance estimates over multiple augmentation seeds or repeated training runs, and report significance tests.
minor comments (6)
  1. [Appendix A.4] The sentence 'The results are shown in Figure 8' is incorrect: Figure 8 is the training-efficiency figure in Appendix A.1, while the VideoScore results are in Table 8. Please fix the cross-reference.
  2. [Section 3.2] The statement 'Videos with a Dover score greater than 0 are excluded' is confusing because Dover quality scores are typically higher-is-better. If a normalized or inverted score is being used, please define the range explicitly; otherwise the later 'Dover score >0.3' combined filter appears inconsistent.
  3. [Section 4.3, Table 4] The table caption says the comparison is 'on VBench', but the metrics (Total, WE, IQ, SC, BC, MS) are the Med-VBench metrics described in Section 4.1. Please use consistent terminology so readers know which prompt set was used.
  4. [Appendix A.4, Table 8] The 'Warping Error' column in the VideoScore table repeats the same values as in Table 2. If this is the same human annotation reused across both benchmarks, please state that; if it is new, please clarify why the numbers are identical.
  5. [Section 4.3] There is a typo in 'warping wrror' and a grammatical error in 'Experiments demonstrates' in the same section; please proofread.
  6. [Appendix A.3] Only one of the 200 evaluation prompts is shown. For reproducibility and transparency, please release the full prompt list (or a large representative sample) with the code/data release.

Circularity Check

1 steps flagged · score 4.0 of 10

Med-VBench prompt set lacks demonstrated independence from MedGen's training captions, partially circular for the headline open-source claim.

  1. fitted input called prediction [Section 4.1 (Evaluation Metrics) and Appendix A.3 (Evaluation Details)]
    "Building on this framework, we introduce medical video prompts and rename the benchmark Med-VBench ... To prompt all models to generate medical videos for evaluation, we designed 200 different prompts, covering a variety of medical video categories and medical scenarios."

    MedGen is fine-tuned on the GPT-4o-generated detailed captions of MedVideoCap-55K (Sec 3.1). The Med-VBench benchmark that underwrites the 'leading open-source' claim is constructed by the same authors from 200 prompts covering the same medical video categories, with no reported check that the prompts are disjoint from or distributionally independent of the training captions. The A.3 example prompt ('The short video depicts a surgical procedure taking place in an operating room...') is stylistically and topically nearly indistinguishable from the Fig. 2 dataset caption ('The video shows a surgical procedure in a sterile operating room...').

full rationale

The paper is an empirical systems contribution rather than a formal derivation, so the main circularity risk is in the evaluation protocol. MedGen is trained on MedVideoCap-55K, whose captions are generated by GPT-4o from medical videos. The authors then introduce Med-VBench by adding 200 author-designed medical prompts to VBench and omitting the aesthetic dimension; they do not establish that these prompts are held out from, or even distributionally different from, the training captions. The example prompt in Appendix A.3 is strikingly similar in style and content to the sample caption in Figure 2, supporting the concern that MedGen is being evaluated on its own training distribution while baselines are zero-shot. This makes the Med-VBench comparison partially circular for the headline claim. However, the paper also reports results on VideoScore, an external benchmark, and a human doctor evaluation with high inter-rater agreement; both provide independent evidence that MedGen improves medical video generation. No self-citation chains, imported uniqueness claims, or ansatz-by-citation patterns are present. The downstream augmentation experiments are not circular in the derivation sense, though they would benefit from error bars and contamination checks. Overall, the central claim has genuine independent support, but the self-constructed Med-VBench weakens the strength of the 'leading open-source' statement.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's central claims depend on the reliability of automated filtering and captioning tools (CLIP, OCR, DOVER, GPT-4o) and on the transferability of VBench metrics to medical content. No new theoretical entities or fitted laws are introduced; the free parameters are hand-chosen dataset thresholds and LoRA hyperparameters.

free parameters (6)
  • Temporal consistency threshold τ = not reported (described as threshold τ)
    Clip segmentation keeps adjacent frames only when CLIP similarity exceeds τ (Section 3.1). The numeric value is omitted, which affects dataset composition and is needed for replication.
  • Minimum clip length = 6 frames
    Clips must satisfy I[j-i >= 6] (Section 3.1), a hand-chosen duration floor.
  • OCR subtitle threshold = 20 words
    Videos with more than 20 detected OCR words on 5 sampled frames are discarded (Section 3.2).
  • Aesthetic score threshold = 3.0
    Average LAION aesthetic score below 3.0 removes blurry or cluttered videos (Section 3.2).
  • Dover score exclusion threshold = 0 (and >0.3 for combined filter)
    Videos with technical quality score >0 are removed; combined filter removes videos with Dover >0.3 and LAION <4.0 (Section 3.2).
  • LoRA rank = 32
    MedGen training uses LoRA rank 32; a hand-chosen hyperparameter (Section 4.1 and Table 5).
assumptions (4)
  • domain assumption The frame-level medical classifier C, trained on human-annotated data using CLIP embeddings, correctly distinguishes medical from non-medical frames.
    Section 3.1: all frames in a kept clip must satisfy C(x_k)=1; if the classifier is biased, the dataset inherits systematic errors.
  • domain assumption CLIP embedding similarity between adjacent frames is a valid proxy for temporal coherence.
    Section 3.1: clips are retained only when S(x_k, x_{k-1}) > τ; no validation of this proxy on medical content is provided.
  • domain assumption GPT-4o captions, generated from 8 sampled frames plus title, description, and transcript, are sufficiently accurate for training and evaluation.
    Section 3.1: captions are MLLM-generated; errors propagate into both training signal and any caption-based evaluation.
  • domain assumption VBench metrics, when applied to medical prompts, measure medical video quality without modification beyond dropping the aesthetic dimension.
    Section 4.1: the paper renames VBench to Med-VBench; no validation that the per-dimension metrics are calibrated for medical content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos." pith.science (2026). https://pith.science/paper/O2ARR3XD

@misc{pith2026250705675,
  author       = {Pith},
  title        = {Pith review of: MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2ARR3XD}},
  note         = {Machine review of arXiv:2507.05675}
}
read the original abstract

Recent advances in video generation have shown remarkable progress in open-domain settings, yet medical video generation remains largely underexplored. Medical videos are critical for applications such as clinical training, education, and simulation, requiring not only high visual fidelity but also strict medical accuracy. However, current models often produce unrealistic or erroneous content when applied to medical prompts, largely due to the lack of large-scale, high-quality datasets tailored to the medical domain. To address this gap, we introduce MedVideoCap-55K, the first large-scale, diverse, and caption-rich dataset for medical video generation. It comprises over 55,000 curated clips spanning real-world medical scenarios, providing a strong foundation for training generalist medical video generation models. Built upon this dataset, we develop MedGen, which achieves leading performance among open-source models and rivals commercial systems across multiple benchmarks in both visual quality and medical accuracy. We hope our dataset and model can serve as a valuable resource and help catalyze further research in medical video generation. Our code and data is available at https://github.com/FreedomIntelligence/MedGen

Figures

Figures reproduced from arXiv: 2507.05675 by the authors.

Figure 1
Figure 1. Failure cases of Sora, Pika, and Hailuo on medi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Sample from MedVideoCap-55K. Each data point consists of a medical video clip, a brief caption, and a detailed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the data construction process. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Overview of our data filtering pipeline. Each stage applies specific filters and shows the volume of data removed and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Data distribution in our MedVideoCap-55K. (a): category distribution of medical videos. (b): word count distribution [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Results of the doctor evaluators assessing the medical videos generated by MedGen and other models across three [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Examples of MedGen-generated videos across diverse real-world medical scenarios. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of training efficiency. (a): the change [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Example video clips of distortion at different levels, with the distorted areas highlighted in red boxes. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. From Articulated Kinematics to Routed Visual Control for Action-Conditioned Surgical Video Generation

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    A kinematic-to-visual lifting paradigm combined with hierarchically routed control generates action-conditioned surgical videos with better faithfulness, fidelity, and efficiency.

  2. UniReason-Med: A Shared Grounded Reasoning Interface for 2D-to-3D Transfer in Medical VQA

    cs.CV 2026-06 unverdicted novelty 4.0 of 10

    UniReason-Med introduces a unified framework for 2D and 3D medical VQA with shared grounded reasoning, trained on a 220K dataset, claiming that joint 2D+3D supervision improves 3D performance over 3D-only training.

Reference graph

Works this paper leans on

45 extracted references · 21 canonical work pages · cited by 2 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bain, M.; Nagrani, A.; Varol, G.; and Zisserman, A. 2022. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. arXiv:2104.00650

    work page Pith review arXiv 2022

  4. [4]

    The SARAS Endoscopic Surgeon Action Detection (ESAD) dataset: Challenges and methods

    Bawa, V. S.; Singh, G.; KapingA, F.; Skarga-Bandurova, I.; Oleari, E.; Leporini, A.; Landolfo, C.; Zhao, P.; Xiang, X.; Luo, G.; et al. 2021. The saras endoscopic surgeon action detection (esad) dataset: Challenges and methods. arXiv preprint arXiv:2104.03178

    work page Pith review arXiv 2021

  5. [5]

    Blattmann, A.; Dockhorn, T.; Kulal, S.; Mendelevitch, D.; Kilian, M.; Lorenz, D.; Levi, Y.; English, Z.; Voleti, V.; Letts, A.; et al. 2023. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127

    arXiv 2023

  6. [6]

    H.; Hicks, S.; Jha, D.; Eskeland, S

    Borgli, H.; Thambawita, V.; Smedsrud, P. H.; Hicks, S.; Jha, D.; Eskeland, S. L.; Randel, K. R.; Pogorelov, K.; Lux, M.; Nguyen, D. T. D.; et al. 2020. HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy. Scientific data, 7(1): 283

    work page 2020

  7. [7]

    Chen, H.; Zhang, Y.; Cun, X.; Xia, M.; Wang, X.; Weng, C.; and Shan, Y. 2024 a . VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. arXiv:2401.09047

    arXiv 2024

  8. [8]

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

    work page Pith review arXiv 2024

Show all 45 references
  1. [9]

    Cho, J.; Schmidgall, S.; Zakka, C.; Mathur, M.; Kaur, D.; Shad, R.; and Hiesinger, W. 2024. SurGen: Text-Guided Diffusion Model for Surgical Video Generation. arXiv:2408.14028

    2024 arXiv

  2. [10]

    Z.; Psychogyios, D.; Zhang, Y.; Hanrahan, J

    Das, A.; Khan, D. Z.; Psychogyios, D.; Zhang, Y.; Hanrahan, J. G.; Vasconcelos, F.; Pang, Y.; Chen, Z.; Wu, J.; Zou, X.; et al. 2024. Pitvis-2023 challenge: Workflow recognition in videos of endoscopic pituitary surgery. arXiv preprint arXiv:2409.01184

    2024 arXiv

  3. [11]

    Fan, W.; Si, C.; Song, J.; Yang, Z.; He, Y.; Zhuo, L.; Huang, Z.; Dong, Z.; He, J.; Pan, D.; et al. 2025. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models. arXiv preprint arXiv:2501.08453

    2025 arXiv

  4. [12]

    Ghamsarian, N.; Taschwer, M.; Putzgruber-Adamitsch, D.; Sarny, S.; and Schoeffmann, K. 2021. Relevance detection in cataract surgery videos by spatio-temporal action localization. In 2020 25th International conference on pattern recognition (ICPR), 10720--10727. IEEE

    2021

  5. [13]

    Gupta, D.; Attal, K.; and Demner-Fushman, D. 2023. A dataset for medical instructional video classification and question answering. Scientific Data, 10(1): 158

    2023

  6. [14]

    HaCohen, Y.; Chiprut, N.; Brazowski, B.; Shalem, D.; Moshe, D.; Richardson, E.; Levin, E.; Shiran, G.; Zabari, N.; Gordon, O.; et al. 2024. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103

    2024 arXiv

  7. [15]

    He, X.; Jiang, D.; Zhang, G.; Ku, M.; Soni, A.; Siu, S.; Chen, H.; Chandra, A.; Jiang, Z.; Arulraj, A.; et al. 2024. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252

    2024 arXiv

  8. [16]

    Huang, Z.; He, Y.; Yu, J.; Zhang, F.; Si, C.; Jiang, Y.; Zhang, Y.; Wu, T.; Jin, Q.; Chanpaisit, N.; et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21807--21818

    2024

  9. [17]

    Jin, Y.; Sun, Z.; Li, N.; Xu, K.; Xu, K.; Jiang, H.; Zhuang, N.; Huang, Q.; Song, Y.; Mu, Y.; and Lin, Z. 2024. Pyramidal Flow Matching for Efficient Video Generative Modeling. arXiv:2410.05954

    2024

  10. [18]

    Kong, W.; Tian, Q.; Zhang, Z.; Min, R.; Dai, Z.; Zhou, J.; Xiong, J.; Li, X.; Wu, B.; Zhang, J.; et al. 2024. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603

    2024 arXiv

  11. [19]

    Kuaishou. 2025. Kling. https://klingai.kuaishou.com/

    2025

  12. [20]

    Y.; Li, W.; Liu, X.; Chen, Z.; Shao, J.; and Yuan, Y

    Li, C.; Liu, H.; Liu, Y.; Feng, B. Y.; Li, W.; Liu, X.; Chen, Z.; Shao, J.; and Yuan, Y. 2024 a . Endora: Video generation models as endoscopy simulators. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 230--240. Springer

    2024

  13. [21]

    Li, L.; Qiu, J.; Saha, A.; Li, L.; Li, P.; He, M.; Guo, Z.; and Yuan, W. 2024 b . Artificial Intelligence for Biomedical Video Generation. arXiv:2411.07619

    2024 arXiv

  14. [22]

    Lin, B.; Ge, Y.; Cheng, X.; Li, Z.; Zhu, B.; Wang, S.; He, X.; Ye, Y.; Yuan, S.; Chen, L.; et al. 2024. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131

    2024 arXiv

  15. [23]

    Ma, X.; Wang, Y.; Jia, G.; Chen, X.; Liu, Z.; Li, Y.-F.; Chen, C.; and Qiao, Y. 2024. Latte: Latent Diffusion Transformer for Video Generation. arXiv preprint arXiv:2401.03048

    2024 arXiv

  16. [24]

    Merrick, L.; Xu, D.; Nuti, G.; and Campos, D. 2024. Arctic-embed: Scalable, efficient, and accurate text embedding models. arXiv preprint arXiv:2405.05374

    2024 arXiv

  17. [25]

    Mesejo, P.; Pizarro, D.; Abergel, A.; Rouquette, O.; Beorchia, S.; Poincloux, L.; and Bartoli, A. 2016. Computer-aided classification of gastrointestinal lesions in regular colonoscopy. IEEE transactions on medical imaging, 35(9): 2051--2063

    2016

  18. [26]

    MiniMax. 2025. Hailuo. https://hailuoai.com/video

    2025

  19. [27]

    Nan, K.; Xie, R.; Zhou, P.; Fan, T.; Yang, Z.; Chen, Z.; Li, X.; Yang, J.; and Tai, Y. 2024. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371

    2024 arXiv

  20. [28]

    I.; Yu, T.; Gonzalez, C.; Seeliger, B.; Mascagni, P.; Mutter, D.; Marescaux, J.; and Padoy, N

    Nwoye, C. I.; Yu, T.; Gonzalez, C.; Seeliger, B.; Mascagni, P.; Mutter, D.; Marescaux, J.; and Padoy, N. 2022. Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. Medical Image Analysis, 78: 102433

    2022

  21. [29]

    OpenAI. 2025. Video generation models as world simulators. https://openai.com/sora/

    2025

  22. [30]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  23. [31]

    Schoeffmann, K.; Husslein, H.; Kletz, S.; Petscharnig, S.; Muenzer, B.; and Beecks, C. 2018. Video retrieval in laparoscopic video recordings with dynamic content descriptors. Multimedia Tools and Applications, 77: 16813--16832

    2018

  24. [32]

    Sharma, A.; Yu, A.; Razavi, A.; Toor, A.; Pierson, A.; Gupta, A.; Waters, A.; van den Oord, A.; Tanis, D.; Erhan, D.; Lau, E.; Shaw, E.; Barth-Maron, G.; Shaw, G.; Zhang, H.; Nandwani, H.; Moraldo, H.; Kim, H.; Blok, I.; Bauer, J.; Donahue, J.; Chung, J.; Mathewson, K.; David,...

    2024

  25. [33]

    Sun, W.; You, X.; Zheng, R.; Yuan, Z.; Li, X.; He, L.; Li, Q.; and Sun, L. 2024. Bora: Biomedical generalist video generation model. arXiv preprint arXiv:2407.08944

    2024 arXiv

  26. [34]

    Team, G. 2024. Mochi 1. https://github.com/genmoai/models

    2024

  27. [35]

    Team, P. 2025 a . Pika. https://pika.art/

    2025

  28. [36]

    Team, W. 2025 b . Wan: Open and Advanced Large-Scale Video Generative Models. https://github.com/Wan-Video/Wan2.1

    2025

  29. [37]

    P.; Shehata, S.; Mutter, D.; Marescaux, J.; De Mathelin, M.; and Padoy, N

    Twinanda, A. P.; Shehata, S.; Mutter, D.; Marescaux, J.; De Mathelin, M.; and Padoy, N. 2016. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36(1): 86--97

    2016

  30. [38]

    Wang, J.; Yuan, H.; Chen, D.; Zhang, Y.; Wang, X.; and Zhang, S. 2023. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571

    2023 arXiv

  31. [39]

    J.; Dai, Q.; and Fang, L

    Wang, X.; Zhang, X.; Zhu, Y.; Guo, Y.; Yuan, X.; Xiang, L.; Wang, Z.; Ding, G.; Brady, D. J.; Dai, Q.; and Fang, L. 2020. PANDA: A Gigapixel-level Human-centric Video Dataset. arXiv:2003.04852

    2020 arXiv

  32. [40]

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

    2024 arXiv

  33. [41]

    H.; Wang, A.; Sun, W

    Wu, H.; Zhang, E.; Liao, L.; Chen, C.; Hou, J. H.; Wang, A.; Sun, W. S.; Yan, Q.; and Lin, W. 2023. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. In International Conference on Computer Vision (ICCV)

    2023

  34. [42]

    Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; et al. 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072

    2024 arXiv

  35. [43]

    Zheng, Z.; Peng, X.; Yang, T.; Shen, C.; Li, S.; Liu, H.; Zhou, Y.; Li, T.; and You, Y. 2024. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404

    2024 arXiv

  36. [44]

    Zhou, Y.; Wang, Q.; Cai, Y.; and Yang, H. 2024. Allegro: Open the Black Box of Commercial-Level Video Generation Model. arXiv preprint arXiv:2410.15458

    2024 arXiv

  37. [45]

    Zia, A.; Bhattacharyya, K.; Liu, X.; Berniker, M.; Wang, Z.; Nespolo, R.; Kondo, S.; Kasai, S.; Hirasawa, K.; Liu, B.; et al. 2023. Surgical tool classification and localization: results and methods from the miccai 2022 surgtoolloc challenge. arXiv preprint arXiv:2305.07152

    2023 arXiv

Pith tools

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