Pith. sign in

REVIEW 3 major objections 3 minor 36 references

Synthetic Data Augmentation using Pre-trained Diffusion Models for Long-tailed Food Image Classification

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A two-stage pipeline that picks the most visually similar class as a negative prompt and blends diversity with contrastive guidance generates synthetic food data that beats prior methods on two long-tailed benchmarks.

desk verdict Plausible recipe for long-tailed food augmentation, but the key selection heuristic is never isolated from the CCFG component. read the letter →

arxiv 2506.01368 v1 pith:RODUHRRB submitted 2025-06-02 cs.CV

GaYeon Koh , Hyun-Jic Oh , Jeonghyun Noh , Won-Ki Jeong This is my paper
classification cs.CV
keywords long-tailedclassificationsyntheticdataaugmentationdiffusionmodelsstablefoodimagenegativepromptingcontrastiveguidanceclassimbalance
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

The paper addresses a failure mode in using pre-trained Stable Diffusion to augment long-tailed food datasets: when asked to generate images for visually similar classes (e.g., Biscuits vs Cookies), the model produces overlapping, confusable images, and naive negative prompts can produce unrelated scenery or people. The proposed two-stage pipeline first generates a reference set with a positive prompt and condition-annealed sampling (CADS) for diversity, then uses a pre-trained ResNet-18 to encode these images and selects, for each target class, the class with the most similar mean feature vector as a negative prompt. Stage two generates the augmentation set with DiSC-DS, a sampler that combines CADS-style diversity with CCFG-style contrastive guidance against the chosen negative prompt, with a time-dependent weighting. The paper reports the highest top-1 accuracy on both long-tailed benchmarks, beating the second-best method by 2.3% on Food101-LT and 7.9% on VFN-LT, with the largest gains on tail classes. The significance is that this is a training-free way (no diffusion fine-tuning, no balanced dataset needed) to turn a pre-trained generator into useful synthetic data for rare food classes.

What carries the argument

The central object is DiSC-DS, a sampling procedure that fuses two existing guidance mechanisms: Condition-Annealed Diffusion Sampling (CADS), which injects scheduled Gaussian noise into the text-conditioning vector so that identical prompts yield diverse images, and Contrastive Classifier-Free Guidance (CCFG), which uses a noise-contrastive estimate to pull samples toward a positive prompt and push them away from a negative prompt. The paper's addition is a weighted combination of the two noise estimates, with a per-timestep temperature $\tau_{t,\gamma} = \tau\sqrt{\gamma(t)}$ that shrinks as the annealing schedule $\gamma(t)$ decays, so contrastive separation is strong early and fades as diversity takes over. The negative prompt itself is selected by a separate, data-driven mechanism: a reference set is generated with a positive prompt, its images are encoded by a pre-trained ResNet-18, and the class whose mean feature vector has the highest cosine similarity to the target class's mean vector is chosen as the negative prompt. That selection is what gives the contrastive guidance a concrete and correct concept to suppress.

What would settle it

A direct test is to replace the similarity-selected negative prompt with a randomly chosen food class for every target class and rerun the pipeline on both benchmarks; if the overall accuracy stays within noise, the confusing-class selection is not what makes DiSC-DS work, and if accuracy drops, the selection is load-bearing. A complementary check is to compute the negative prompts from real training images instead of the synthetic reference set; if the chosen classes differ substantially for many targets, the reference set's representativeness is the sensitive step that would need to be controlled.

Watch

Extended reading notes

Core claim

The central claim is that inter-class confusion, not just scarcity, is what limits synthetic augmentation of long-tailed food data, and that this confusion can be largely removed by conditioning generation with an automatically chosen negative prompt: the class whose mean feature vector in a pre-trained ResNet-18 space is most similar to the target class. Building on that selection, the DiSC-DS sampler linearly interpolates a CADS-style noise estimate (which adds scheduled noise to the conditioning vector for intra-class diversity) and a CCFG-style contrastive noise estimate (which steers sampling toward the positive and away from the negative prompt), with the contrastive temperature annealed by the same schedule as the conditioning noise. In the paper's experiments this pipeline achieves the highest overall top-1 accuracy on Food101-LT (51.6%) and VFN-LT (59.1%), outperforming the second-best method by 2.3 and 7.9 percentage points respectively, and the ablation study shows that both the negative-prompt guidance and the dynamic (rather than fixed) contrastive weight are needed for the best result.

Load-bearing premise

The whole method rests on the mean feature vector of a synthetic reference set, computed with a pre-trained ResNet-18, correctly identifying the real class that is most visually confusable with the target class; if the reference set does not look like real food, the selected negative prompt could suppress the wrong features and the separation benefit would disappear.

Editorial extensions

If this is right

  • On both benchmarks, the full pipeline beats every compared method in overall top-1 accuracy, with a larger margin on the more imbalanced VFN-LT (7.9%) than on Food101-LT (2.3%).
  • The gains are concentrated in tail classes (tail accuracy 45.2% on Food101-LT and 52.9% on VFN-LT), while head-class accuracy is also the highest among compared methods, showing the approach does not trade head for tail performance.
  • The ablation shows that CADS alone (diversity without negative prompts) increases diversity but leaves classes overlapping and can hurt head accuracy, while adding the CCFG negative-prompt component in DiSC-DS recovers the loss and gives the best overall accuracy.
  • The paper's ablations also show that the dynamic, annealed version of the contrastive temperature $\tau$ is needed; fixed values of 0.2, 0.5, and 0.8 all give lower overall accuracy on at least one benchmark.

Reading between the lines

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

  • Because the negative-prompt selection only needs a feature encoder and cosine similarity, the same confusing-class selection could be applied on top of any prompt-conditional generator (newer diffusion or flow models) without retraining, as long as a text prompt can name the classes.
  • The method's choice of ResNet-18 as the similarity encoder is likely sensitive; a stronger or food-specific encoder might select a different negative class, and the paper does not test this, so the robustness of the selection step is an open question.
  • A concrete extension is to apply the same two-stage recipe to other long-tailed or fine-grained domains with visually confusable classes (e.g., birds, flowers, skin lesions), where inter-class separation is also a known failure mode.
  • If the claim holds, it suggests a general recipe for synthetic augmentation: let a pre-trained generator create a reference set, use data-driven feature similarity to pick the negative prompt, and anneal contrastive strength over sampling time to balance diversity and separation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper presents a two-stage synthetic data augmentation framework (DiSC-DS) for long-tailed food image classification. In Stage 1, the method generates a diverse reference set with a pre-trained Stable Diffusion model using CADS, encodes these images with a pre-trained ResNet-18, and selects, per class, the most visually similar other class (Eq. 7) to serve as a negative prompt. In Stage 2, it combines CADS and Contrastive CFG (CCFG) with a dynamic temperature (Eq. 10) and a weighted interpolation (Eq. 11) to generate synthetic training images. A ResNet-18 classifier is trained on real and synthetic images with Mixup. The authors report top-1 accuracy of 51.6% on Food101-LT and 59.1% on VFN-LT, outperforming previous methods by 2.3% and 7.9% respectively.

Significance. If the results are reproducible, the method is a practically useful, fine-tuning-free augmentation strategy for long-tailed food recognition; it does not require a balanced dataset for diffusion fine-tuning and it explicitly targets both intra-class diversity and inter-class separation. The paper includes a re-implementation of SYNAuG, ablations of the sampling components (Tables 2 and 3), and a feature-level analysis with t-SNE and Inception Score. However, the main novelty - confusing-class selection - is not isolated experimentally, and the evidence is based on single-run numbers and partly adopted baselines. The significance therefore depends on the missing control experiments, not on the architecture or loss definitions.

major comments (3)
  1. [Section 3.1.1, Eq. (7), Table 2] The proposed confusing-class selection is never isolated. Table 2 contrasts CADS (no negative prompt) with DiSC-DS (selected negative prompt), so the observed gain conflates the effect of the negative-prompt selection with the CCFG mechanism, the dynamic temperature, and the interpolation weight alpha. No experiment compares Eq. (7) against a randomly chosen negative class or a fixed generic negative prompt while holding all other components fixed. Without such a control, the central novelty of the paper is unsupported: if random negative prompts perform equivalently, the method reduces to CADS+CCFG with arbitrary conditioning, and the reported improvement does not depend on the confusing-class selection. Please add an ablation with random negative classes and a fixed generic negative prompt, keeping all other settings identical.
  2. [Section 4.1, Table 1] All reported accuracies are single runs without error bars or multiple seeds, and most baseline numbers are adopted from Food1Stage [12] rather than re-run under a common protocol. With margins of 2.3% and 7.9%, it is unclear whether the improvements are statistically reliable; the lack of variance estimates and the absence of ClusDiff on Food101-LT make the state-of-the-art claim difficult to assess. Please report mean and standard deviation over at least three seeds for the proposed method and the main baselines, or clearly state the variance reported in the source papers.
  3. [Sections 3.1 and 4.1] The manuscript does not specify the number of reference images per class in Stage 1 or the number of synthetic images per class in Stage 2, and it describes hyperparameters (dynamic tau, alpha, Mixup mode) as chosen per dataset without a validation protocol. These choices are load-bearing for reproducibility and for guarding against test-set overfitting; for example, Table 2 uses Mixup-random on Food101-LT and Mixup-all on VFN-LT, a choice made after seeing test accuracy. Please report the exact generation counts or ratios and a validation procedure that does not use the test set for model selection.
minor comments (3)
  1. [Algorithm 1] The notation is inconsistent: the Require line states 'tau_t > 0 (tau = 0.8)' but Eqs. (8)-(10) use tau as both a scalar and a time-dependent schedule; please define tau_t explicitly and clarify how the fixed value 0.8 is used in the dynamic schedule.
  2. [Eqs. (1)-(4)] The description of the rescaling in Eq. (3) is confusing; the text says 'we corrupt y_hat+ to y_hat' but the equation defines y_hat_rescaled, and the subsequent mixing in Eq. (4) is not clearly connected to the CADS formulation. Please rewrite this passage to distinguish the noisy condition, the rescaled condition, and the final condition used in Eq. (5).
  3. [Figure 4] The Inception Score is reported per class for only two classes; since IS is normally computed over a set of images and is sensitive to class statistics, please state how the score is computed and why per-class IS values are comparable across the four settings.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is evaluated on external test sets, and the method's components and negative-prompt selection are not defined in terms of the reported accuracy.

full rationale

The paper's derivation chain is self-contained. DiSC-DS is an empirical combination of externally published components (pre-trained Stable Diffusion [30], CADS [32], CCFG [6]) plus a novel confusing-class selection rule (Eqs. 6-7). That rule selects a negative prompt from cosine similarity of mean features of a synthetic reference set, but the reported top-1 accuracies are measured on real, held-out test sets (Food101-LT and VFN-LT), so no reported number is a refit of an input by construction. The only points where the authors choose settings based on their own results are the per-dataset Mixup mode (Rand vs. All) in Table 2 and the dynamic-tau schedule in Table 3; these are hyperparameter selections, not quantities derived from the benchmark labels and then renamed as predictions, and they do not make the central SOTA claim true by definition. No load-bearing self-citation appears: the authors' own prior works ([23-25]) are cited only in related work as examples of diffusion-based augmentation in pathology, and the load-bearing components CADS and CCFG are external. The absence of a random-negative-prompt ablation is a robustness gap concerning whether the controversial selection rule is necessary, but that is an experimental-support issue, not circularity. Overall, the evaluation is externally grounded, so the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The method is a hyperparameter-heavy empirical pipeline. The central accuracy claim depends on roughly a dozen hand-set hyperparameters (CADS schedule, CFG scale, CCFG temperature, interpolation weight, per-dataset Mixup mode) and on assumptions about the reliability of pre-trained components (Stable Diffusion, ResNet-18, CADS, CCFG). No new physical or conceptual entities are introduced.

free parameters (9)
  • CADS initial noise scale s = 0.1
    Controls the amount of condition noise in reference set generation (Eq. 1); set by hand.
  • CADS annealing thresholds tau1, tau2 = 0.5, 0.9
    Defined in Eq. 2 for the gamma(t) schedule; chosen by the authors, not derived.
  • CADS condition mixing factor psi = 1.0
    Eq. 4 weighs rescaled and original condition; set to 1.0 in experiments.
  • CFG guidance scale w = 2.0
    Eq. 5; balances conditioned and unconditioned noise estimates.
  • CCFG temperature tau = 0.8 (base, dynamically scaled)
    Eqs. 8-10; Table 3 compares fixed vs dynamic and selects dynamic tau=0.8 as best.
  • CADS/CCFG interpolation weight alpha = 0.8
    Eq. 11; trades diversity (CADS) against separation (CCFG).
  • Mixup strategy = Random for Food101-LT, All for VFN-LT
    Post-hoc choice from Table 2; different strategies on the two datasets.
  • Synthetic images per class = Not reported
    The number of generated samples per class is never stated, but it determines the augmentation strength.
  • Number of denoising steps = 50
    DPM-Solver++ with 50 steps; standard sampling budget.
assumptions (5)
  • domain assumption Pre-trained Stable Diffusion v1.4 can generate recognizable food images from text prompts
    The whole pipeline conditions on SD v1.4; if generation quality is poor for food, the method fails. The authors even note that naive SD generation gives unrealistic images for some classes (Fig. 1).
  • domain assumption ResNet-18 features on synthetic reference images encode visually relevant class similarity
    Negative prompt selection uses cosine similarity of mean features (Eqs. 6-7); no evidence that these similarities match real-image confusions.
  • domain assumption CCFG guidance derived from NCE loss correctly pulls samples away from negative concepts
    The paper adopts CCFG [6] as a black box; Eqs. 8-9 rely on the contrastive guidance formulation from the cited work.
  • domain assumption Mixup between real and synthetic images reduces the domain gap without introducing harmful label noise
    Training follows SYNAuG [34]; no analysis of whether mixed images retain correct labels.
  • domain assumption Reported baseline numbers from Food1Stage [12] are directly comparable to the authors' runs
    The paper adopts reported scores for most baselines (Sec. 4.1); if those used different training setups or backbones, the SOTA claim weakens.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synthetic Data Augmentation using Pre-trained Diffusion Models for Long-tailed Food Image Classification." pith.science (2026). https://pith.science/paper/RODUHRRB

@misc{pith2026250601368,
  author       = {Pith},
  title        = {Pith review of: Synthetic Data Augmentation using Pre-trained Diffusion Models for Long-tailed Food Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RODUHRRB}},
  note         = {Machine review of arXiv:2506.01368}
}
read the original abstract

Deep learning-based food image classification enables precise identification of food categories, further facilitating accurate nutritional analysis. However, real-world food images often show a skewed distribution, with some food types being more prevalent than others. This class imbalance can be problematic, causing models to favor the majority (head) classes with overall performance degradation for the less common (tail) classes. Recently, synthetic data augmentation using diffusion-based generative models has emerged as a promising solution to address this issue. By generating high-quality synthetic images, these models can help uniformize the data distribution, potentially improving classification performance. However, existing approaches face challenges: fine-tuning-based methods need a uniformly distributed dataset, while pre-trained model-based approaches often overlook inter-class separation in synthetic data. In this paper, we propose a two-stage synthetic data augmentation framework, leveraging pre-trained diffusion models for long-tailed food classification. We generate a reference set conditioned by a positive prompt on the generation target and then select a class that shares similar features with the generation target as a negative prompt. Subsequently, we generate a synthetic augmentation set using positive and negative prompt conditions by a combined sampling strategy that promotes intra-class diversity and inter-class separation. We demonstrate the efficacy of the proposed method on two long-tailed food benchmark datasets, achieving superior performance compared to previous works in terms of top-1 accuracy.

Figures

Figures reproduced from arXiv: 2506.01368 by the authors.

Figure 1
Figure 1. Synthetic images generated using pre-trained SD models. (a) Images generated using only positive prompts for “Biscuits” and “Cookies,” highlighting their visual similarity. (b) Images generated with additional randomly selected negative prompts (one or multiple classes), often resulting in unintended artifacts like scenery or people. selects a class with the most similar features as a negative prompt, to prevent int… view at source ↗
Figure 2
Figure 2. Overview of the Proposed Two-Stage Data Augmentation Framework. (a) To mitigate inter-class confusion, we generate a reference set and compute cosine similarity in the feature space to identify the most visually similar class for each target class. This class is used as a negative prompt in subsequent synthesis. (b) We then apply DiSC-DS, leveraging condition-annealed sampling to balance intra-class diversity and in… view at source ↗
Figure 3
Figure 3. Visual comparison of food generation results. We compare the generated images from four different sampling approaches—Pre-trained SD, CADS, CCFG, and DiSC-DS (Ours)—against the real input data. To analyze the effectiveness of each method, we focus on visually similar class pairs within each dataset: “Garlic bread” and “French toast” from the Food101-LT dataset, and “Pork chop” and “Pork rib” from the VFN-LT dataset.… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of synthetic samples for “Garlic bread” and “French toast,” in Food101-LT. Comparing Pre-trained SD, CADS, CCFG, and DiSC-DS (Ours) using the Inception Score (IS) [2], DiSC-DS achieves high-quality data and better class separation. Methods Mixup [36…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [12]

    Single-stage heavy-tailed food classification

    Jiangpeng He and Fengqing Zhu. Single-stage heavy-tailed food classification. In IEEE Int. Conf. Image Process., pages 1115–1119, 2023. 1, 3, 6

    work page 2023

  2. [1]

    Long-tailed recognition via weight balancing

    Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6897–6907,

    work page

  3. [2]

    A note on the inception score

    Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018. 6, 8

    arXiv 2018

  4. [3]

    Food-101–mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. In Eur. Conf. Comput. Vis., pages 446–461, 2014. 5

    work page 2014

  5. [4]

    A systematic study of the class imbalance problem in convolu- tional neural networks

    Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolu- tional neural networks. Neural Netw., 106:249–259, 2018. 1, 2, 6

    work page 2018

  6. [5]

    Learning imbalanced datasets with label- distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label- distribution-aware margin loss. Adv. Neural Inform. Process. Syst., 32, 2019. 1, 2, 6

    work page 2019

  7. [6]

    Con- trastive CFG: Improving CFG in diffusion models by con- trasting positive and negative concepts

    Jinho Chang, Hyungjin Chung, and Jong Chul Ye. Con- trastive CFG: Improving CFG in diffusion models by con- trasting positive and negative concepts. arXiv preprint arXiv:2411.17077, 2024. 2, 3, 5, 6, 7, 8

    arXiv 2024

  8. [7]

    Improved regular- ization of convolutional neural networks with cutout

    Terrance DeVries and Graham W Taylor. Improved regular- ization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 3

    arXiv 2017

Show all 36 references
  1. [8]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst., 34:8780–8794, 2021. 3

    2021

  2. [9]

    Noise-contrastive estimation: A new estimation principle for unnormalized sta- tistical models

    Michael Gutmann and Aapo Hyv ¨arinen. Noise-contrastive estimation: A new estimation principle for unnormalized sta- tistical models. In Int. Conf. Artif. Intell. Stat. , pages 297– 304, 2010. 5

    2010

  3. [10]

    Latent-based diffusion model for long-tailed recognition

    Pengxiao Han, Changkun Ye, Jieming Zhou, Jing Zhang, Jie Hong, and Xuesong Li. Latent-based diffusion model for long-tailed recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2639–2648, 2024. 3

    2024

  4. [11]

    Diffusion model with clustering-based con- ditioning for food image generation

    Yue Han, Jiangpeng He, Mridul Gupta, Edward J Delp, and Fengqing Zhu. Diffusion model with clustering-based con- ditioning for food image generation. In Int. Workshop Mul- timed. Assist. Dietary Manag., pages 61–69, 2023. 1, 3, 6

    2023

  5. [13]

    Multi-task image-based dietary assessment for food recognition and portion size es- timation

    Jiangpeng He, Zeman Shao, Janine Wright, Deborah Kerr, Carol Boushey, and Fengqing Zhu. Multi-task image-based dietary assessment for food recognition and portion size es- timation. In IEEE Conf. Multimed. Inf. Process. Retrieval , pages 49–54, 2020. 1

    2020

  6. [14]

    Long-tailed food classification

    Jiangpeng He, Luotao Lin, Heather A Eicher-Miller, and Fengqing Zhu. Long-tailed food classification. Nutrients, 15(12):2751, 2023. 1, 2, 3, 5, 6

    2023

  7. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770–778, 2016. 4, 6

    2016

  8. [16]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 4

    2022 arXiv

  9. [17]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Adv. Neural Inform. Process. Syst., 33:6840–6851, 2020. 3

    2020

  10. [18]

    Diffusemix: Label- preserving data augmentation with diffusion models

    Khawar Islam, Muhammad Zaigham Zaheer, Arif Mah- mood, and Karthik Nandakumar. Diffusemix: Label- preserving data augmentation with diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 27621– 27630, 2024. 3

    2024

  11. [19]

    Dynamic negative guid- ance of diffusion models

    Felix Koulischer, Johannes Deleu, Gabriel Raya, Thomas Demeester, and Luca Ambrogioni. Dynamic negative guid- ance of diffusion models. arXiv preprint arXiv:2410.14398,

    arXiv

  12. [20]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems , 35:5775–5787,

  13. [21]

    Visual aware hierarchy based food recognition

    Runyu Mao, Jiangpeng He, Zeman Shao, Sri Kalyan Yarla- gadda, and Fengqing Zhu. Visual aware hierarchy based food recognition. In Int. Conf. Pattern Recog. , pages 571–598,

  14. [22]

    Long-tail learning via logit adjustment

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 1, 3, 6

    2007 arXiv

  15. [23]

    Co- synthesis of histopathology nuclei image-label pairs using a context-conditioned joint diffusion model

    Seonghui Min, Hyun-Jic Oh, and Won-Ki Jeong. Co- synthesis of histopathology nuclei image-label pairs using a context-conditioned joint diffusion model. In Eur. Conf. Comput. Vis., pages 146–162, 2024. 3

    2024

  16. [24]

    Diffmix: Diffusion model- based data synthesis for nuclei segmentation and classifica- tion in imbalanced pathology image datasets

    Hyun-Jic Oh and Won-Ki Jeong. Diffmix: Diffusion model- based data synthesis for nuclei segmentation and classifica- tion in imbalanced pathology image datasets. In Med. Image Comput. Comput. Assist. Interv., pages 337–345, 2023

    2023

  17. [25]

    Controllable and efficient multi-class pathology nuclei data augmentation using text- conditioned diffusion models

    Hyun-Jic Oh and Won-Ki Jeong. Controllable and efficient multi-class pathology nuclei data augmentation using text- conditioned diffusion models. In Med. Image Comput. Com- put. Assist. Interv., pages 36–46, 2024. 3

    2024

  18. [26]

    Training lan- guage models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training lan- guage models to follow instructions with human feedback. Adv. Neural Inform. Process. Syst., 35:27730–27744, 2022. 3

    2022

  19. [27]

    Influence-balanced loss for imbalanced visual clas- sification

    Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual clas- sification. In Int. Conf. Comput. Vis., pages 735–744, 2021. 1, 2, 6

    2021

  20. [28]

    The majority can help the minority: Context-rich minority oversampling for long-tailed classifi- cation

    Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classifi- cation. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 6887–6896, 2022. 6

    2022

  21. [29]

    Balanced meta-softmax for long-tailed visual recog- nition

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recog- nition. Adv. Neural Inform. Process. Syst. , 33:4175–4186,

  22. [30]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684–10695, 2022. 1, 3, 5, 6, 7, 8

    2022

  23. [31]

    Focal loss for dense object detection

    T-YLPG Ross and GKHP Doll ´ar. Focal loss for dense object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2980–2988, 2017. 1, 2, 6

    2017

  24. [32]

    CADS: Unleashing the diversity of diffusion models through condition-annealed sampling

    Seyedmorteza Sadat, Jakob Buhmann, Derek Bradley, Ot- mar Hilliges, and Romann M Weber. CADS: Unleashing the diversity of diffusion models through condition-annealed sampling. arXiv preprint arXiv:2310.17347, 2023. 2, 3, 5, 6, 7, 8

    2023 arXiv

  25. [33]

    Experimental perspectives on learning from imbal- anced data

    Jason Van Hulse, Taghi M Khoshgoftaar, and Amri Napoli- tano. Experimental perspectives on learning from imbal- anced data. In Proc. Int. Conf. Mach. Learn., pages 935–942,

  26. [34]

    SYNAuG: Exploiting synthetic data for data imbalance problems

    Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. SYNAuG: Exploiting synthetic data for data imbalance problems. arXiv preprint arXiv:2308.00994, 2023. 1, 3, 5, 6

    2023 arXiv

  27. [35]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In Int. Conf. Comput. Vis., pages 6023–6032, 2019. 3

    2019

  28. [36]

    mixup: Beyond empirical risk minimiza- tion

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. arXiv preprint arXiv:1710.09412, 2017. 3, 5, 6, 8

    2017 arXiv

Pith tools

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