pith. sign in

arxiv: 2606.22416 · v1 · pith:BZMTZ3L3new · submitted 2026-06-21 · 💻 cs.CV

Gen2Balance: Generative Balancing for Long-Tailed Video Action Recognition

Prajwal Gatti , Simon Jenni , Fabian Caba Heilbron , Dima Damen This is my paper

Pith reviewed 2026-06-26 10:25 UTC · model grok-4.3

classification 💻 cs.CV
keywords long-tailed learningvideo action recognitiongenerative augmentationtext-to-video modelsimbalanced datasetssynthetic datadomain shift mitigation
0
0 comments X

The pith

Gen2Balance augments long-tailed video datasets with text-to-video generated clips and two-stage training to raise accuracy on rare actions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes using a text-to-video generative model conditioned on action profiles and training exemplars to create synthetic clips that balance an imbalanced video action dataset. It converts the training set into a mix of real and generated clips and applies a two-stage training process to reduce the effects of domain shift between the two sources. On long-tailed versions of UCF-101 and a Kinetics subset the method improves accuracy by 5.1% and 7.0% over strong baselines, with a 31.9% gain on rare actions from the RareAct dataset. Experiments also show that using only a fraction of the synthetic data captures most of the benefit at much lower compute cost.

Core claim

Gen2Balance converts an imbalanced training set into a balanced combination of real and generated video clips by conditioning a text-to-video generative model on diverse text prompts grounded in action profiles and training exemplars. A two-stage training strategy mitigates domain shift between real and synthetic data and produces higher recognition accuracy on long-tailed video benchmarks.

What carries the argument

Two-stage training on a balanced mixture of real videos and synthetic clips generated by a text-to-video model conditioned on action profiles and exemplars.

Load-bearing premise

The text-to-video generative model produces clips that are sufficiently realistic, diverse, and action-relevant that the two-stage training can mitigate any domain shift between real and synthetic data without introducing new biases or artifacts that degrade performance.

What would settle it

If adding the generated clips to the training set produces lower accuracy on the tail classes than training on the real data alone, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.22416 by Dima Damen, Fabian Caba Heilbron, Prajwal Gatti, Simon Jenni.

Figure 1
Figure 1. Figure 1: We introduce Gen2Balance, a generative balancing approach for long-tail video action recognition. (Left) For four actions (three from K100-LT, one from RareAct [48]), we show a real video from the training dataset alongside a generated sample with its conditioning prompt (bottom text), illustrating the visual diversity pro￾duced by our pipeline. (Right) Class-wise accuracy on K100-LT and RareAct classes, s… view at source ↗
Figure 2
Figure 2. Figure 2: We progressively illustrate our video generation pipeline on the semantically ambiguous Kinetics class Robot Dancing (a human dance style imitating a robot). Naive prompting and diversification alone incorrectly generate mechanical robots dancing (Cols. 1–2). Adding an Action Profile (Col. 3) improves the definition, but it remains ambiguous (shows both humans and robots dancing). By providing real exempla… view at source ↗
Figure 3
Figure 3. Figure 3: The Gen2Balance Training Strategy. Stage 1 trains on the filled dataset (Daug) with Balanced Softmax loss (LBS), using margins based on real-data frequencies (Nc). Stage 2 fine-tunes only on real data with the same loss to rectify domain shift. 3.3 Training Gen2Balance While generative filling balances the dataset, directly training on Daug (i.e., using standard cross-entropy without any long-tail adjustme… view at source ↗
Figure 4
Figure 4. Figure 4: Per-class accuracy improvement of Gen2Balance over the CE baseline (top) and BSCE [54] (bottom) on K100-LT. Classes are ordered by size, from largest to smallest (left to right). Background colour indicates head/tail/few-shot splits [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results from three datasets comparing CE, Logit Adj., and Gen2Balance (Ours). Colours indicate the category of the predicted class. For each test video, we also show the two nearest generated samples from Dgen in feature space [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of Scaling Generated Data. We analyse the impact of Filling Thresh￾old B (bottom axis) and corresponding generation cost in GPU hours (top axis) on model performance. (Left) On K100-LT, while full balancing (B=990) yields the high￾est accuracy, partial balancing (B=330) achieves competitive results at just 27% of the compute cost. (Right) On UCF-LT, performance continues to improve even when oversam… view at source ↗
Figure 7
Figure 7. Figure 7: User Study Interface. Users are provided with a target action class and 3 real reference videos to establish semantic grounding. They are then asked to select all candidate-generated videos (out of 5) that accurately depict the target action, rejecting the distractors, testing human recognisability of generated videos. In this example, the fourth video depicts parasailing rather than swimming butterfly str… view at source ↗
Figure 8
Figure 8. Figure 8: Failures in Gen2Balance generated videos. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Label Noise in real Kinetics-100 videos. [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
read the original abstract

We address the problem of training on long-tailed data for video action recognition. We propose to augment the training set using a text-to-video generative model, conditioned on diverse text prompts grounded in action profiles and training exemplars. Our approach, called Gen2Balance, converts an imbalanced training set into a balanced combination of real and generated video clips. To effectively learn from such data, we employ a two-stage training strategy that mitigates domain shift and yields significant improvements. We evaluate on long-tailed versions of standard benchmarks: UCF-101 (UCF-LT) and a 100-class subset of Kinetics (K100-LT) selected to prioritise temporally challenging actions. Gen2Balance improves accuracy over the strongest baselines for long-tailed learning by 5.1% and 7.0% on the respective datasets. On rare actions from the RareAct dataset (e.g., cut keyboard), Gen2Balance improves accuracy by 31.9%, demonstrating effectiveness for scarce actions. By varying the amount of synthetic data added, we show that partial balancing already achieves 79% of the performance gains at 27% of the compute cost on K100-LT, highlighting the practical scalability of Gen2Balance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes Gen2Balance to address long-tailed video action recognition by augmenting imbalanced training sets with synthetic video clips generated via a text-to-video model conditioned on diverse text prompts derived from action profiles and training exemplars. The method converts the data into a balanced real+synthetic mixture and applies a two-stage training strategy to mitigate domain shift. Evaluations on long-tailed versions of UCF-101 (UCF-LT) and a 100-class Kinetics subset (K100-LT) report accuracy gains of 5.1% and 7.0% over the strongest long-tailed learning baselines, with a 31.9% improvement on rare actions from the RareAct dataset. Partial balancing is shown to achieve 79% of the gains at 27% of the compute cost on K100-LT.

Significance. If the empirical claims hold after verification of generated data quality and ablations, the work would be significant for demonstrating a scalable use of generative models to balance video datasets, particularly benefiting rare and temporally complex actions. The efficiency result on partial balancing provides a practical contribution that could influence data augmentation strategies in imbalanced computer vision tasks.

major comments (3)
  1. [Abstract] Abstract: The headline gains (5.1% UCF-LT, 7.0% K100-LT, 31.9% rare actions) are attributed to balancing via generated clips and two-stage training, yet no quantitative evidence on generated clip quality (e.g., video FID, human ratings, or classification accuracy on synthetic data alone) is supplied; this is load-bearing because poor realism or semantic mismatch would invalidate attribution of improvements to balancing rather than artifacts or negative transfer.
  2. [Abstract] Abstract: No ablation is described that isolates the two-stage training component from simply adding synthetic samples; without this, it is impossible to confirm that the procedure mitigates domain shift without introducing new biases, which directly supports the central claim that the full pipeline is required for the reported gains.
  3. [Abstract] Abstract: Details on baseline implementations (e.g., exact long-tailed methods, backbones, and hyper-parameters), the precise generative conditioning procedure, and statistical significance of the improvements are absent, preventing verification that the gains are reproducible and fairly compared to prior work.
minor comments (2)
  1. [Abstract] Abstract: The term 'action profiles' is introduced without a brief definition or example, which would aid clarity for readers unfamiliar with the conditioning approach.
  2. [Abstract] Abstract: The efficiency claim ('79% of the performance gains at 27% of the compute cost') references a specific experiment but does not point to the corresponding figure or table, reducing traceability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that additional evidence and details are needed to strengthen the claims and will incorporate them in the revision. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline gains (5.1% UCF-LT, 7.0% K100-LT, 31.9% rare actions) are attributed to balancing via generated clips and two-stage training, yet no quantitative evidence on generated clip quality (e.g., video FID, human ratings, or classification accuracy on synthetic data alone) is supplied; this is load-bearing because poor realism or semantic mismatch would invalidate attribution of improvements to balancing rather than artifacts or negative transfer.

    Authors: We agree that quantitative validation of generated clip quality is essential to attribute gains to balancing. In the revised manuscript we will add video FID scores (comparing synthetic to real videos), human ratings on realism and semantic fidelity, and classification accuracy when training a model on synthetic data alone. These additions will directly address concerns about realism and semantic match. revision: yes

  2. Referee: [Abstract] Abstract: No ablation is described that isolates the two-stage training component from simply adding synthetic samples; without this, it is impossible to confirm that the procedure mitigates domain shift without introducing new biases, which directly supports the central claim that the full pipeline is required for the reported gains.

    Authors: We acknowledge that an explicit ablation separating the two-stage training from naive addition of synthetic samples is required. We will add this ablation in the revision, reporting performance when synthetic samples are added with and without the two-stage procedure, to isolate its contribution to mitigating domain shift. revision: yes

  3. Referee: [Abstract] Abstract: Details on baseline implementations (e.g., exact long-tailed methods, backbones, and hyper-parameters), the precise generative conditioning procedure, and statistical significance of the improvements are absent, preventing verification that the gains are reproducible and fairly compared to prior work.

    Authors: We will expand the experimental section to include full details on baseline implementations (specific long-tailed methods, backbones, and hyper-parameters), the exact text-prompt conditioning procedure used for generation, and statistical significance via mean and standard deviation over multiple random seeds. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claims on held-out sets with no derivations

full rationale

The paper presents Gen2Balance, an empirical method that augments long-tailed video datasets with synthetic clips from a text-to-video model and uses two-stage training. All reported results are accuracy improvements on held-out test sets (UCF-LT, K100-LT, RareAct). No equations, mathematical derivations, or self-referential definitions are present in the abstract or described method. The central claims are falsifiable empirical measurements rather than reductions to fitted inputs or self-citations. This is a standard empirical ML paper with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; the central claim rests on the unverified premise that generated videos add useful signal rather than noise once domain shift is addressed by two-stage training. No explicit free parameters, invented entities, or formal axioms are stated in the abstract.

axioms (1)
  • domain assumption A text-to-video generative model conditioned on action profiles and training exemplars can produce videos that improve recognition of rare classes when mixed with real data.
    This is the load-bearing premise that allows the augmentation step to be treated as beneficial rather than neutral or harmful.

pith-pipeline@v0.9.1-grok · 5756 in / 1466 out tokens · 35541 ms · 2026-06-26T10:25:54.760509+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

96 extracted references

  1. [1]

    In: ICLR (2023)

    Ahn, S., Ko, J., Yun, S.Y.: CUDA: Curriculum of data augmentation for long-tailed recognition. In: ICLR (2023)

    2023

  2. [2]

    In: CVPR (2022)

    Alshammari, S., Wang, Y.X., Ramanan, D., Kong, S.: Long-tailed recognition via weight balancing. In: CVPR (2022)

    2022

  3. [3]

    arXiv preprint (2025)

    Assran, M., Bardes, A., Fan, D., Garrido, Q., Howes, R., Muckley, M., Rizvi, A., Roberts, C., Sinha, K., Zholus, A., et al.: V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint (2025)

    2025

  4. [4]

    TMLR (2023)

    Azizi, S., Kornblith, S., Saharia, C., Norouzi, M., Fleet, D.J.: Synthetic data from diffusion models improves imagenet classification. TMLR (2023)

    2023

  5. [5]

    arXiv preprint (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-VL technical report. arXiv preprint (2025)

    2025

  6. [6]

    In: ICCV (2021)

    Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021)

    2021

  7. [7]

    In: ICLR (2026)

    Bansal, H., Peng, C., Bitton, Y., Goldenberg, R., Grover, A., Chang, K.W.: VideoPhy-2:Achallengingaction-centricphysicalcommonsenseevaluationinvideo generation. In: ICLR (2026)

    2026

  8. [8]

    In: ICCV (2021)

    Cai, J., Wang, Y., Hwang, J.N.: ACE: Ally complementary experts for solving long-tailed recognition in one-shot. In: ICCV (2021)

    2021

  9. [9]

    JAIR (2002)

    Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: SMOTE: synthetic minority over-sampling technique. JAIR (2002)

    2002

  10. [10]

    In: ECCV (2020)

    Chou, H.P., Chang, S.C., Pan, J.Y., Wei, W., Juan, D.C.: Remix: rebalanced mixup. In: ECCV (2020)

    2020

  11. [11]

    arXiv preprint (2025)

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint (2025)

    2025

  12. [12]

    In: CVPR (2019)

    Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V.: AutoAugment: Learn- ing augmentation strategies from data. In: CVPR (2019)

    2019

  13. [13]

    IEEE TPAMI (2022)

    Cui, J., Liu, S., Tian, Z., Zhong, Z., Jia, J.: ResLT: Residual learning for long-tailed recognition. IEEE TPAMI (2022)

    2022

  14. [14]

    In: CVPR (2019)

    Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR (2019)

    2019

  15. [15]

    IJCV (2022)

    Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Molti- santi, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. IJCV (2022)

    2022

  16. [16]

    In: CVPR (2021)

    Deng, Z., Liu, H., Wang, Y., Wang, C., Yu, Z., Sun, X.: PML: Progressive margin loss for long-tailed age classification. In: CVPR (2021)

    2021

  17. [17]

    In: CVPR (2023) 16 P

    Du, F., Yang, P., Jia, Q., Nan, F., Chen, X., Yang, Y.: Global and local mix- ture consistency cumulative learning for long-tailed visual recognitions. In: CVPR (2023) 16 P. Gatti et al

    2023

  18. [18]

    In: ICLR (2023)

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image gener- ation using textual inversion. In: ICLR (2023)

    2023

  19. [19]

    Google Blog (May 2025),https://deepmind.google/models/veo/, accessed: 2026-02-17

    Google DeepMind: Veo 3: Video generation with native audio. Google Blog (May 2025),https://deepmind.google/models/veo/, accessed: 2026-02-17

    2025

  20. [20]

    something something

    Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The "something something" video database for learning and evaluating visual common sense. In: ICCV (2017)

    2017

  21. [21]

    In: ICLR (2026)

    Gu, J., Liu, X., Zeng, Y., Nagarajan, A., Zhu, F., Hong, D., Fan, Y., Yan, Q., Zhou, K., Liu, M.Y., et al.: PhyWorldBench: A comprehensive evaluation of physical realism in text-to-video models. In: ICLR (2026)

    2026

  22. [22]

    In: CVPR (2019)

    Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)

    2019

  23. [23]

    In: ICLR (2024)

    Hasegawa, N., Sato, I.: Exploring weight balancing on long-tailed recognition prob- lem. In: ICLR (2024)

    2024

  24. [24]

    He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., QI, X.: Is synthetic data from generative models ready for image recognition? In: ICLR (2023)

    2023

  25. [25]

    In: ICML (2025)

    Hou, Y., Jia, Y.: A square peg in a square hole: Meta-expert for long-tailed semi- supervised learning. In: ICML (2025)

    2025

  26. [26]

    IEEE TMM (2023)

    Hu, Y., Gao, J., Xu, C.: Learning multi-expert distribution calibration for long- tailed video classification. IEEE TMM (2023)

    2023

  27. [27]

    Multimedia Systems (2025)

    Hu, Y., Zhang, Y., Zhang, L.: Long-tailed video recognition via majority-guided diffusion model. Multimedia Systems (2025)

    2025

  28. [28]

    In: CVPR (2023)

    Iscen, A., Fathi, A., Schmid, C.: Improving image recognition by retrieving from web-scale image-text data. In: CVPR (2023)

    2023

  29. [29]

    In: CVPR (2020)

    Jamal, M.A., Brown, M., Yang, M.H., Wang, L., Gong, B.: Rethinking class- balanced methods for long-tailed visual recognition from a domain adaptation perspective. In: CVPR (2020)

    2020

  30. [30]

    Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decouplingrepresentationandclassifierforlong-tailedrecognition.In:ICLR(2020)

    2020

  31. [31]

    arXiv preprint (2017)

    Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint (2017)

    2017

  32. [32]

    1 Kontext: Flow matching for in-context image generation and editing in latent space

    Labs,B.F.,Batifol,S.,Blattmann,A.,Boesel,F.,Consul,S.,Diagne,C.,Dockhorn, T., English, J., English, Z., Esser, P., et al.: FLUX. 1 Kontext: Flow matching for in-context image generation and editing in latent space. arXiv preprint (2025)

    2025

  33. [33]

    In: CVPR (2022)

    Li, B., Yao, Y., Tan, J., Zhang, G., Yu, F., Lu, J., Luo, Y.: Equalized focal loss for dense long-tailed object detection. In: CVPR (2022)

    2022

  34. [34]

    In: CVPR (2022)

    Li, B., Han, Z., Li, H., Fu, H., Zhang, C.: Trustworthy long-tailed classification. In: CVPR (2022)

    2022

  35. [35]

    In: CVPR (2022)

    Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: CVPR (2022)

    2022

  36. [36]

    In: AAAI (2024)

    Li, M., Zhikai, H., Lu, Y., Lan, W., Cheung, Y.m., Huang, H.: Feature fusion from head to tail for long-tailed visual recognition. In: AAAI (2024)

    2024

  37. [37]

    In: ECCV (2024)

    Li, R., Feng, Z., Xu, T., Li, L., Wu, X.J., Awais, M., Atito, S., Kittler, J.: C2C: Component-to-compositionlearningforzero-shotcompositionalactionrecognition. In: ECCV (2024)

    2024

  38. [38]

    In: IJCAI (2025) Gen2Balance 17

    Li, W., Luo, D., Yang, D., Li, Z., Wang, W., Zhou, Y.: The role of video generation in enhancing data-limited action understanding. In: IJCAI (2025) Gen2Balance 17

    2025

  39. [39]

    In: AAAI (2023)

    Li, X., Xu, H.: MEID: mixture-of-experts with internal distillation for long-tailed video recognition. In: AAAI (2023)

    2023

  40. [40]

    In: CVPR (2020)

    Li, Y., Wang, T., Kang, B., Tang, S., Wang, C., Li, J., Feng, J.: Overcoming classifier imbalance for long-tail object detection with balanced group softmax. In: CVPR (2020)

    2020

  41. [41]

    IJCV (2024)

    Lin, J., Liu, Z., Wang, W., Wu, W., Wang, L.: VLG: General video recognition with web textual knowledge. IJCV (2024)

    2024

  42. [42]

    In: ICCV (2017)

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: ICCV (2017)

    2017

  43. [43]

    IEEE Transactions on Systems, Man, and Cybernetics (2008)

    Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics (2008)

    2008

  44. [44]

    In: CVPR (2019)

    Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR (2019)

    2019

  45. [45]

    In: CVPR (2022)

    Long, A., Yin, W., Ajanthan, T., Nguyen, V., Purkait, P., Garg, R., Blair, A., Shen, C., Van den Hengel, A.: Retrieval augmented classification for long-tail visual recognition. In: CVPR (2022)

    2022

  46. [46]

    In: ICLR (2021)

    Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)

    2021

  47. [47]

    Midjourney, Inc.: Midjourney.https://www.midjourney.com(2022), accessed: 2026-02-17

    2022

  48. [48]

    arXiv preprint (2020)

    Miech, A., Alayrac, J.B., Laptev, I., Sivic, J., Zisserman, A.: RareAct: A video dataset of unusual interactions. arXiv preprint (2020)

    2020

  49. [49]

    In: AAAI (2023)

    Moon, W., Seong, H.S., Heo, J.P.: Minority-oriented vicinity expansion with at- tentive aggregation for video long-tailed recognition. In: AAAI (2023)

    2023

  50. [50]

    Motamed, S., Culp, L., Swersky, K., Jaini, P., Geirhos, R.: Do generative video models understand physical principles? In: WACV (2026)

    2026

  51. [51]

    OpenAI Blog (Sep 2025),https://openai.com/index/ sora-2/, accessed: 2026-02-17

    OpenAI: Sora 2 is here. OpenAI Blog (Sep 2025),https://openai.com/index/ sora-2/, accessed: 2026-02-17

    2025

  52. [52]

    In: CVPR (2023)

    Perrett, T., Sinha, S., Burghardt, T., Mirmehdi, M., Damen, D.: Use your head: Improving long-tail video recognition. In: CVPR (2023)

    2023

  53. [53]

    Google Blog (The Keyword) (Nov 2025),https://blog.google/innovation-and-ai/products/nano-banana-pro/, accessed: 2026-02-17

    Raisinghani, N.: Introducing Nano Banana Pro. Google Blog (The Keyword) (Nov 2025),https://blog.google/innovation-and-ai/products/nano-banana-pro/, accessed: 2026-02-17

    2025

  54. [54]

    NeurIPS (2020)

    Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long- tailed visual recognition. NeurIPS (2020)

    2020

  55. [55]

    In: CVPR (2022)

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

    2022

  56. [56]

    Runway Research (Dec 2025),https: //runwayml.com/research/introducing-runway-gen-4.5, accessed: 2026-02-17

    Runway AI: Introducing Runway Gen-4.5. Runway Research (Dec 2025),https: //runwayml.com/research/introducing-runway-gen-4.5, accessed: 2026-02-17

    2025

  57. [57]

    In: CVPR (2023)

    Sariyildiz, M.B., Alahari, K., Larlus, D., Kalantidis, Y.: Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In: CVPR (2023)

    2023

  58. [58]

    In: WACV (2021)

    Sevilla-Lara, L., Zha, S., Yan, Z., Goswami, V., Feiszli, M., Torresani, L.: Only time can tell: Discovering temporal data for temporal modeling. In: WACV (2021)

    2021

  59. [59]

    In: NeurIPS (2024)

    Shao, J., Zhu, K., Zhang, H., Wu, J.: DiffuLT: Diffusion for long-tail recognition without external knowledge. In: NeurIPS (2024)

    2024

  60. [60]

    arXiv preprint (2023)

    Shin, J., Kang, M., Park, J.: Fill-Up: Balancing long-tailed data with generative models. arXiv preprint (2023)

    2023

  61. [61]

    NeurIPS (2019) 18 P

    Shu, J., Xie, Q., Yi, L., Zhao, Q., Zhou, S., Xu, Z., Meng, D.: Meta-weight-net: Learning an explicit mapping for sample weighting. NeurIPS (2019) 18 P. Gatti et al

    2019

  62. [62]

    In: CVPR (2025)

    Sidhu, M., Chopra, H., Blume, A., Kim, J., Reddy, R.G., Ji, H.: Search and detect: Training-free long tail object detection via web-image retrieval. In: CVPR (2025)

    2025

  63. [63]

    In: CVPR 2024 Workshop SyntaGen: Harnessing Generative Models for Synthetic Visual Datasets (2024)

    Singh, K., Navaratnam, T., Holmer, J., Schaub-Meyer, S., Roth, S.: Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In: CVPR 2024 Workshop SyntaGen: Harnessing Generative Models for Synthetic Visual Datasets (2024)

    2024

  64. [64]

    arXiv preprint (2012)

    Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint (2012)

    2012

  65. [65]

    In: ICLR (2025)

    Sun, S., Lu, H., Li, J., Xie, Y., Li, T., Yang, X., Zhang, L., Yan, J.: Rethinking classifier re-training in long-tailed recognition: Label over-smooth can balance. In: ICLR (2025)

    2025

  66. [66]

    In: CVPR (2021)

    Tan, J., Lu, X., Zhang, G., Yin, C., Li, Q.: Equalization loss v2: A new gradient balance approach for long-tailed object detection. In: CVPR (2021)

    2021

  67. [67]

    In: CVPR (2020)

    Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J.: Equalization loss for long-tailed object recognition. In: CVPR (2020)

    2020

  68. [68]

    In: ICCV (2023)

    Tao, Y., Sun, J., Yang, H., Chen, L., Wang, X., Yang, W., Du, D., Zheng, M.: Local and global logit adjustments for long-tailed learning. In: ICCV (2023)

    2023

  69. [69]

    Thozhiyoor, V.V., Tripathi, S., Radhakrishnan, V.B., Bhattad, A.: Objects in gen- erated videos are slower than they appear: Models suffer sub-earth gravity and don’t know galileo’s principle... for now. In: CVPR Findings (2026)

    2026

  70. [70]

    NeurIPS (2020)

    Tian, J., Liu, Y.C., Glaser, N., Hsu, Y.C., Kira, Z.: Posterior re-calibration for imbalanced datasets. NeurIPS (2020)

    2020

  71. [71]

    NeurIPS (2022)

    Tong, Z., Song, Y., Wang, J., Wang, L.: VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS (2022)

    2022

  72. [72]

    In: ICLR 2019 Workshop Deep Gen- erative Models for Highly Structured Data (2019)

    Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: FVD: A new metric for video generation. In: ICLR 2019 Workshop Deep Gen- erative Models for Highly Structured Data (2019)

    2019

  73. [73]

    In: ICML (2019)

    Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Ben- gio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: ICML (2019)

    2019

  74. [74]

    arXiv preprint (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: WAN: Open and advanced large-scale video generative models. arXiv preprint (2025)

    2025

  75. [75]

    In: ICLR (2024)

    Wang, B., Wang, P., Xu, W., Wang, X., Zhang, Y., Wang, K., Wang, Y.: Kill two birds with one stone: Rethinking data augmentation for deep long-tailed learning. In: ICLR (2024)

    2024

  76. [76]

    In: NeurIPS (2024)

    Wang, P., Zhao, Z., Wen, H., Wang, F., Wang, B., Zhang, Q., Wang, Y.: LLM- autoDA: Largelanguage model-drivenautomatic dataaugmentation forlong-tailed problems. In: NeurIPS (2024)

    2024

  77. [77]

    In: ICLR (2020)

    Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.X.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2020)

    2020

  78. [78]

    In: ICLR (2024)

    Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: InternVid: A large-scale video-text dataset for multimodal understanding and generation. In: ICLR (2024)

    2024

  79. [79]

    arXiv preprint (2025)

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint (2025)

    2025

  80. [80]

    In: CVPR (2021)

    Wu, T., Liu, Z., Huang, Q., Wang, Y., Lin, D.: Adversarial robustness under long- tailed distribution. In: CVPR (2021)

    2021

Showing first 80 references.