Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation

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

Pith's one-line read PoGDiff replaces the ground-truth diffusion target with a Product of Gaussians, combining a sample's own target with the model's prediction under a neighboring text embedding, to keep fine-tuning on imbalanced data from collapsing…

desk verdict The pseudonym/evaluation-name mismatch is a real confound that breaks the headline identity claims; the PoG consistency regularizer is a plausible idea that deserves a corrected test. read the letter →

arxiv 2502.08106 v3 pith:2RWJQ3AA submitted 2025-02-12 cs.LG cs.AIcs.CVstat.ML

classification cs.LGcs.AIcs.CVstat.ML
keywords imbalancedtext-to-imagegenerationdiffusionmodelsfine-tuningProductofGaussiansminorityclasseslong-taileddistributiongenerativerecall
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 diffusion models fine-tuned on imbalanced image-text datasets fail on minority groups because the standard denoising objective locks each text prompt to exactly its own images, leaving rare identities with almost no statistical support. It proposes PoGDiff, which replaces the ground-truth target in the denoising KL divergence with a Product of Gaussians built from the original target and the model's prediction for a neighboring text embedding. That change adds a consistency term to the training loss: the model is pulled to denoise the same image given similar prompts, with a weight that favors visually similar neighbors and rare text embeddings. On imbalanced variants of AgeDB, DigiFace, VGGFace2, and CIFAR-100, PoGDiff reports large gains over vanilla fine-tuning, CBDM, and T2H in identity accuracy, FID, DINO score, and a new coverage metric (gRecall). The paper claims this is the first general fine-tuning method for imbalanced text-to-image generation with natural language prompts.

What carries the argument

The machine is the Gaussian-product identity: $\mathcal{N}(\mu_1, \lambda_1^{-1}I) \circ \mathcal{N}(\mu_2, \lambda_2^{-1}I) = \mathcal{N}\!\left(\frac{\lambda_1\mu_1 + \lambda_2\mu_2}{\lambda_1 + \lambda_2}, (\lambda_1 + \lambda_2)^{-1}I\right)$. The paper uses it to combine the ground-truth denoising target with the model's prediction under a neighboring text embedding $y'$, producing a new target mean $\mu_{\mathrm{PoG}}$ that the model is pushed toward. The simplification of the resulting KL bound yields the two-term training loss, where the second term $\psi\|\epsilon_\theta(x_t,y)-\epsilon_\theta(x_t,y')\|^2$ enforces that similar prompts denoise to similar images; $\psi$ is computed as an image-cosine similarity (with a bonus when both images share an identity) times an inverse text density from a VAE ELBO, and the neighbor $y'$ is sampled from the $k$ nearest text embeddings with probability proportional to image similarity.

What would settle it

Evaluate with the same pseudonym prompts used for training (e.g., “An image of Lukas” for Einstein) instead of the real names, or test on invented identities that never appeared in the pretrained model's data. If PoGDiff's minority-identity accuracy then drops to near-baseline levels, the reported gains come from the pretrained model's familiarity with real names rather than from the learned text-to-identity mapping.

Watch

Extended reading notes

Core claim

PoGDiff's central claim is that the failure of diffusion fine-tuning on long-tailed data can be traced to the per-sample KL objective, and that swapping the ground-truth Gaussian for a Product of Gaussians fixes it. In the usual formulation, at each denoising step the model $p_\theta(x_{t-1}|x_t, y)$ is trained against $q(x_{t-1}|x_t, x_0, y)$; PoGDiff instead minimizes $D_{\mathrm{KL}}( q(x_{t-1}|x_t, x_0, y) \circ p_\theta(x_{t-1}|x_t, y') \,\|\, p_\theta(x_{t-1}|x_t, y) )$, where $y'$ is a text embedding sampled from the $k$ nearest neighbors of $y$ and $\circ$ is the Gaussian product. With the assumption $\lambda_y = \lambda_{\mathrm{PoG}} = \lambda_t + \lambda_{y'}$, this KL is upper bounded by $A(\lambda_t)\|\epsilon_\theta(x_t,y)-\epsilon\|^2 + A(\lambda_{y'})\|\epsilon_\theta(x_t,y)-\epsilon_\theta(x_t,y')\|^2$, and after absorbing time dependence into a similarity weight $\psi$, the final objective is $\|\epsilon_\theta(x_t,y)-\epsilon\|^2 + \psi\|\epsilon_\theta(x_t,y)-\epsilon_\theta(x_t,y')\|^2$. $\psi$ is the product of a cosine-similarity term between images (with a same-identity bonus) and an inverse text density estimated via a VAE's ELBO, so the objective upweights rare prompts and close neighbors. The paper shows on four imbalanced datasets that this objective simultaneously improves generation accuracy (human, GPT-4o, DINO scores) and quality (FID) over vanilla Stable Diffusion fine-tuning and two adapted baselines, with the biggest margins in the few-shot minority regime.

Load-bearing premise

The evaluation assumes the model can only learn each identity from the fine-tuning data, because training captions use pseudonyms while test prompts use real celebrity names; if Stable Diffusion's pretrained knowledge of those celebrities is doing the work, the reported accuracy gains are not a fair measure of the method.

Editorial extensions

If this is right

  • Fine-tuning a text-to-image diffusion model on a long-tailed dataset with PoGDiff should preserve identity accuracy for classes represented by as few as two images, a regime where vanilla fine-tuning and existing class-balancing baselines produce near-zero accuracy.
  • The two-term objective is a drop-in addition to any denoising-diffusion fine-tuning loop, with roughly a second forward pass at training time and no inference overhead, so it can be layered onto backbone models beyond Stable Diffusion.
  • Combining generation accuracy with a coverage metric like gRecall gives a more complete picture of imbalanced text-to-image performance than FID alone, since FID only captures the mean and variance of the generated distribution.
  • If the bound in Proposition 3.1 is tight enough in practice, PoGDiff training is standard diffusion training plus a consistency regularizer on the prompt space, making its behavior predictable when the text encoder is frozen.

Reading between the lines

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

  • Because $\psi$ uses identity labels to boost same-person neighbors, PoGDiff implicitly requires identity supervision at training time; an unsupervised variant that replaces identity with clustering is plausible but untested, and its effectiveness is an open question.
  • The mechanism is modality-agnostic: any conditioned generative model with a frozen embedding space could borrow neighbor information the same way, so extending PoGDiff-style objectives to video or time-series generation is a natural next step the paper lists but does not try.
  • A stress test suggested by the paper's own failure-case discussion is to measure how often the sampled neighbor $y'$ belongs to a different identity when the minority class has only two images; the method's robustness may degrade as the neighbor pool becomes semantically crowded.
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 / 4 minor

Summary. The paper proposes PoGDiff, a fine-tuning method for imbalanced text-to-image generation. Instead of minimizing the KL divergence between the predicted and ground-truth distributions at each denoising step, PoGDiff replaces the ground-truth target with a Product of Gaussians formed by combining the original target with the model's own prediction conditioned on a neighboring text embedding. The resulting objective (Eq. 13) adds to the standard diffusion noise-matching term a regularizer ψ‖εθ(xt,y)−εθ(xt,y′)‖² that encourages consistent predictions for similar text embeddings, with ψ combining image similarity and an inverse text density estimated by a VAE. Experiments are reported on imbalanced versions of AgeDB, DigiFace, VGGFace2, and CIFAR-100, using FID, DINO, human evaluation, GPT-4o, and a proposed gRecall metric. The paper claims that PoGDiff improves both generation accuracy and quality, especially for minority classes.

Significance. If the empirical claims were validated, PoGDiff would address a real and underexplored problem: fine-tuning text-to-image diffusion models on long-tailed data with natural-language prompts. The proposed gRecall metric is a sensible attempt to measure diversity under an accuracy constraint, and the proof of Proposition 3.1 provides a valid upper bound under the stated Gaussian assumptions. The manuscript also provides detailed dataset construction and implementation descriptions, which is useful for reproducibility. However, the paper's central empirical evidence is compromised by a train/evaluation prompt mismatch, and the theoretical derivation contains a pivotal ad hoc step; as a result, the contribution currently rests on a heuristic regularizer whose claimed benefits are not convincingly established.

major comments (4)
  1. [Sec. 3.2.2, Eq. (5)–(6)] The transition from Eq. (5) to Eq. (6) is not a derivation: after rewriting λy′/λt as σt²/σy′², the paper 'hypothetically defines' σy′² = σt²/ψ to eliminate the time-step dependence. This choice is not implied by the Gaussian model, and it forces the regularizer weight to equal the arbitrarily constructed ψ in Eq. (12), which depends on image similarity, a VAE density estimate, and unreported hyperparameters a1, a2, a3. In addition, Proposition 3.1 relies on the assumption λy = λPoG = λt + λy′, which is stated without justification. The upper bound itself is fine, but the claimed equivalence between PoGDiff and a normal diffusion objective plus a consistency regularizer is not established.
  2. [Appendix C vs. Appendix E] The empirical protocol confounds the fine-tuning effect with Stable Diffusion's pretrained knowledge of celebrity identities. Training captions for AgeDB are pseudonymized (e.g., 'Albert Einstein' is replaced with 'Lukas'), while all generation and evaluation prompts are 'An image of {p}' with p set to the real name. Because Stable Diffusion v1.5 was pretrained on LAION-5B and AgeDB contains well-known public figures, high DINO, human, and GPT-4o scores on real-name prompts can be obtained by retrieving the pretrained representation of that celebrity, without requiring the pseudonym-conditioned fine-tuning to have learned anything. This is the only evidence for the headline accuracy gains on the face datasets, and it directly contradicts the stated setting in Appendix H that the method targets data the model has 'not encountered during pre-training.' A control experiment using pseudonym prompts at evaluation, or a dataset of genuinely novel identities, is needed before the accuracy claims can be accepted.
  3. [Tables 1–5] No error bars, confidence intervals, or significance tests are reported for any of the metrics, and the checklist response marks statistical significance as 'NA.' This is insufficient for the claim that PoGDiff 'consistently outperforms all baselines,' particularly on the small datasets (AgeDB-IT2I-S has 32 images and 2 identities). For example, DINO scores in Table 2 are reported to two decimals without any measure of run-to-run variability, and the gRecall scores in Table 5 are single point estimates. The paper should either provide repeated-run statistics or substantially temper the comparative claims.
  4. [Sec. 4.1, gRecall definition] The gRecall metric depends on a cosine-similarity threshold set to 0.7, but the paper gives no analysis of how sensitive the results in Table 5 are to this threshold or to the choice of DINOv2 as the embedding. Since Table 5 is the primary evidence for the diversity claim, the arbitrary threshold undermines the strength of that claim. A threshold sweep or a discussion of the metric's stability should be provided.
minor comments (4)
  1. [Eq. (12) and Appendix F] The values of the hyperparameters a1, a2, a3 and the number of neighbors k are never reported; Appendix F lists only learning rates, training steps, and batch sizes. These parameters are part of the method and must be disclosed for reproducibility.
  2. [Eq. (9)] The notation sa1+a2·1[I(x)≠I(x′)] is ambiguous: it is presumably an exponent, but the paper does not state the value of the base s when s can be close to 0, and the max(0, ·) wrapper is redundant if s is a cosine similarity in [0,1]. Please clarify the intended operation.
  3. [Appendix G, Table 8] In the T2H row of Table 8, the DINO score for CIFAR-100 appears as '0345', which is missing a decimal point.
  4. [Eq. (10)] The approximation p(y) ≈ eELBOVAE(y) is stated without explaining why the ELBO is an adequate density estimate for text embeddings, nor is the VAE architecture and training procedure described in enough detail to reproduce the inverse text density; Appendix J.7 gives only a brief description.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity in the PoG-to-objective derivation; the psi weight is an explicit construction, and the only self-citation (VIR) is not load-bearing. The main weakness is an evaluation confound, not circular reasoning.

full rationale

The central derivation (Eq. 1 -> Eq. 4) is a genuine upper bound proved in Appendix A under a stated assumption (lambda_y = lambda_PoG = lambda_t + lambda_y'); the conclusion is not used as an input. The step from Eq. 5 to Eq. 6 is explicitly flagged as a construction: 'we hypothetically define sigma^2_y' = sigma^2_t / psi ... thereby effectively removing the time step dependency', so the regularizer weight is chosen, not derived, and the paper then proposes an independent heuristic for psi (image cosine similarity and VAE-ELBO inverse text density) with ablations (Table 6) showing neither psi component is essential for all metrics. The only self-citation is ref. [26] (VIR, same authors), used to motivate inverse text density reweighting alongside external ref. [22] (DIR); it is not a uniqueness theorem and does not forbid alternative weightings, so it is not load-bearing. A separate evaluation concern does not affect circularity: AgeDB captions are pseudonymized at training (Appendix C) but evaluated with real names (Appendix E), while Appendix H states the method targets data SD 'has not encountered during pre-training' - a real confound for the face identity claims, but the reported metrics are not equal by construction to the fitted psi or to any self-cited result. Therefore the derivation chain is self-contained with respect to circularity; the score reflects only a minor, non-load-bearing self-citation.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The paper's central claim rests on several free hyperparameters (a1, a2, a3, k, gRecall threshold) whose values are not reported, on the ad hoc definition sigma^2_y' = sigma^2_t / psi that converts the derived bound into a hand-chosen regularizer, and on the unstated evaluation assumption that real-name prompts are a valid test after pseudonym training. The neighbor-consistency mechanism and the VAE ELBO density estimate are domain assumptions without independent validation.

free parameters (5)
  • a1 = not reported
    Controls the exponent of cosine similarity in psi (Eqn 9); value never specified in main text or appendix.
  • a2 = not reported
    Controls the penalty for different identities in psi (Eqn 9); value never specified.
  • a3 = not reported
    Scales the inverse text density term in psi (Eqn 11); value never specified.
  • k (number of neighbors) = not reported
    Number of nearest neighbors sampled in Eqn (8); value not reported.
  • gRecall threshold = 0.7
    Cosine similarity threshold for classifying a generated image as 'correct' in gRecall; chosen without sensitivity analysis.
assumptions (6)
  • standard math Product of two Gaussians is a Gaussian with precision-weighted mean
    Used in Eqn (2) and Lemma A.1.
  • ad hoc to paper Assumption lambda_y = lambda_PoG = lambda_t + lambda_y'
    Stated in Prop 3.1; required for the upper bound in Eqn (4) but not justified by data or prior theory.
  • ad hoc to paper sigma^2_y' = sigma^2_t / psi can be hypothesized to remove time-step dependency
    Section 3.2.2; this step defines away the derivation and lets any psi be inserted.
  • domain assumption Similar images have text embeddings whose conditional denoising predictions should be consistent
    Core mechanism of borrowing neighbor information (Fig 3); not proven, only motivated.
  • domain assumption VAE ELBO approximates text embedding density p(y)
    Eqn (10); no evidence that the ELBO is calibrated for inverse-density reweighting.
  • domain assumption Evaluation with real names is a valid test of the fine-tuned mapping despite pseudonym training
    Appendix C vs Appendix E; this is an unstated assumption that, if false, confounds the results with pretrained knowledge.
invented entities (2)
  • Product-of-Gaussians (PoG) target distribution
    purpose: Replaces the standard ground-truth diffusion target q(x_{t-1}|x_t,x_0,y) with q composed with p_theta(x_{t-1}|x_t,y') to encourage neighbor consistency.
    A mathematical construction internal to the method; no external falsifiable handle beyond the paper's own experiments.
  • gRecall metric
    purpose: Measures generative diversity when accuracy is enforced, by counting covered training images per identity.
    A new evaluation metric defined by the authors; its threshold and aggregation are arbitrary and not benchmarked against established diversity metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation." pith.science (2026). https://pith.science/paper/2RWJQ3AA

@misc{pith2026250208106,
  author       = {Pith},
  title        = {Pith review of: PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2RWJQ3AA}},
  note         = {Machine review of arXiv:2502.08106}
}
read the original abstract

Diffusion models have made significant advancements in recent years. However, their performance often deteriorates when trained or fine-tuned on imbalanced datasets. This degradation is largely due to the disproportionate representation of majority and minority data in image-text pairs. In this paper, we propose a general fine-tuning approach, dubbed PoGDiff, to address this challenge. Rather than directly minimizing the KL divergence between the predicted and ground-truth distributions, PoGDiff replaces the ground-truth distribution with a Product of Gaussians (PoG), which is constructed by combining the original ground-truth targets with the predicted distribution conditioned on a neighboring text embedding. Experiments on real-world datasets demonstrate that our method effectively addresses the imbalance problem in diffusion models, improving both generation accuracy and quality.

Figures

Figures reproduced from arXiv: 2502.08106 by the authors.

Figure 1
Figure 1. PoGDiff for imbalanced text-to-image generation. Existing methods, e.g., Stable Diffusion [7] and CBDM [15], fall short for minority data (Low Density). In contrast, Our PoGDiff successfully generates high-quality images even for minority data, outperforming all baselines. have a very limited number of them. Since none of the employees appear in the LAION-5B dataset, generating photos of them requires finetuning the… view at source ↗
Figure 2
Figure 2. Overview of our PoGDiff. During fine￾tuning, PoGDiff collects k neighbors of the current text embedding y and samples one y ′ from them based on Eqn. (8). Both y and y ′ will then be em￾ployed to denoise the current image xt to xt−1. where ◦ represents the product of two Gaus￾sian distributions, y ′ is a selected neighbor￾ing embedding from other samples in the training dataset (more details below), and pθ(xt−1|xt, … view at source ↗
Figure 3
Figure 3. Comparing denoising networks of typi￾cal diffusion models [1, 7] and our PoGDiff. Left: In conditional text-to-image diffusion models, a data point (i.e., x) is mainly affected by its text embedding (besides random latent codes). Right: In PoGDiff, neighbors participate to modulate the final effective density. Here, y denotes the text prompts, which are the embeddings of the text descriptions of the images; x denote… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Example generated images from different methods. Our PoGDiff outperforms the baselines [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Example generated images from different methods. Our PoGDiff outperforms the baselines [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Overview of label distributions for six IT2I datasets in bar plots. The x-axis corresponds to [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: TSNE visualization for all the methods for an example individual in the AgeDB-IT2M-large [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Comprehensive Survey on Imbalanced Data Learning

    cs.LG 2025-02 conditional novelty 3.0 of 10

    A structured survey and benchmark that groups imbalanced data learning methods into data re-balancing, feature representation, training strategy, and ensemble learning.

Reference graph

Works this paper leans on

59 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

  2. [2]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

  3. [3]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

  4. [4]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR, 2021

  5. [5]

    Conditional stochastic interpolation for generative learning

    Ding Huang, Jian Huang, Ting Li, and Guohao Shen. Conditional stochastic interpolation for generative learning. arXiv preprint arXiv:2312.05579, 2023

  6. [6]

    Cascaded diffusion models for high fidelity image generation

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1–33, 2022

  7. [7]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

  8. [8]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479–36494, 2022

Show all 59 references
  1. [9]

    Classifier-free diffusion guidance

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

  2. [10]

    Diffwave: A versatile diffusion model for audio synthesis

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020

  3. [11]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837–2845, 2021

  4. [12]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022

  5. [13]

    Mvdream: Multi- view diffusion for 3d generation

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi- view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023

  6. [14]

    Brain imaging generation with latent diffusion models

    Walter HL Pinaya, Petru-Daniel Tudosiu, Jessica Dafflon, Pedro F Da Costa, Virginia Fernandez, Parashkev Nachev, Sebastien Ourselin, and M Jorge Cardoso. Brain imaging generation with latent diffusion models. In MICCAI Workshop on Deep Generative Models, pages 117–126. Springer, 2022

  7. [15]

    Class-balancing diffusion models

    Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18434–18443, 2023

  8. [16]

    Laion- 5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion- 5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Inform...

  9. [17]

    Smote: synthetic minority over-sampling technique

    Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research , 16: 321–357, 2002. 11

  10. [18]

    Learning from imbalanced data

    Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263–1284, 2009

  11. [19]

    Smote for regression

    Luís Torgo, Rita P Ribeiro, Bernhard Pfahringer, and Paula Branco. Smote for regression. In Portuguese conference on artificial intelligence, pages 378–389. Springer, 2013

  12. [20]

    Smogn: a pre-processing approach for imbalanced regression

    Paula Branco, Luís Torgo, and Rita P Ribeiro. Smogn: a pre-processing approach for imbalanced regression. In First international workshop on learning with imbalanced domains: Theory and applications, pages 36–50. PMLR, 2017

  13. [21]

    Rebagg: Resampled bagging for imbalanced regression

    Paula Branco, Luis Torgo, and Rita P Ribeiro. Rebagg: Resampled bagging for imbalanced regression. In Second International Workshop on Learning with Imbalanced Domains: Theory and Applications, pages 67–81. PMLR, 2018

  14. [22]

    Delving into deep imbalanced regression

    Yuzhe Yang, Kaiwen Zha, Yingcong Chen, Hao Wang, and Dina Katabi. Delving into deep imbalanced regression. In International conference on machine learning, pages 11842–11851. PMLR, 2021

  15. [23]

    Balanced mse for imbalanced visual regression

    Jiawei Ren, Mingyuan Zhang, Cunjun Yu, and Ziwei Liu. Balanced mse for imbalanced visual regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7926–7935, 2022

  16. [24]

    Ranksim: Ranking similarity regularization for deep imbalanced regression

    Yu Gong, Greg Mori, and Frederick Tung. Ranksim: Ranking similarity regularization for deep imbalanced regression. arXiv preprint arXiv:2205.15236, 2022

  17. [25]

    Conr: Contrastive regularizer for deep imbalanced regression

    Mahsa Keramati, Lili Meng, and R David Evans. Conr: Contrastive regularizer for deep imbalanced regression. arXiv preprint arXiv:2309.06651, 2023

  18. [26]

    Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing

    Ziyan Wang and Hao Wang. Variational imbalanced regression: Fair uncertainty quantification via probabilistic smoothing. Advances in Neural Information Processing Systems, 36, 2024

  19. [27]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500...

  20. [28]

    Multi- concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi- concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931–1941, 2023

  21. [29]

    Photomaker: Customizing realistic human photos via stacked id embedding

    Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640–8650, 2024

  22. [30]

    Long-tailed diffusion models with oriented calibration

    Tianjiao Zhang, Huangjie Zheng, Jiangchao Yao, Xiangfeng Wang, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Long-tailed diffusion models with oriented calibration. In The Twelfth International Conference on Learning Representations, 2024

  23. [31]

    Label-noise robust diffusion models

    Byeonghu Na, Yeongmin Kim, HeeSun Bae, Jung Hyun Lee, Se Jung Kwon, Wanmo Kang, and Il-Chul Moon. Label-noise robust diffusion models. arXiv preprint arXiv:2402.17517, 2024

  24. [32]

    Finetuning text-to-image diffusion models for fairness

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. arXiv preprint arXiv:2311.07604, 2023

  25. [33]

    Training unbiased diffusion models from biased dataset

    Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il-Chul Moon. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations, 2024

  26. [34]

    Fair generative modeling via weak supervision

    Kristy Choi, Aditya Grover, Trisha Singh, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In International Conference on Machine Learning , pages 1887–1898. PMLR, 2020. 12

  27. [35]

    Agedb: the first manually collected, in-the-wild age database

    Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages...

  28. [36]

    Digiface-1m: 1 million digital face images for face recognition

    Gwangbin Bae, Martin de La Gorce, Tadas Baltrušaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer V...

  29. [37]

    Vggface2: A dataset for recognising faces across pose and age

    Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67–74. IEEE, 2018

  30. [38]

    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. Advances in neural information processing systems, 32, 2019

  31. [39]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

  32. [40]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

  33. [41]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

  34. [42]

    Re- thinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Re- thinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

  35. [43]

    Albert Einstein

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36, 2024. 13 A Proofs for Proposition 3.1 Proposition A.1. Assume λy = λPoG ≜ λt + λy′, we have our loss function LPoGDiff t−1 = Eq λy 2 ∥µθ(xt...

  36. [44]

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: It can be found in the Sec. 1. Guidelines: • The answer NA means that the abstract and introduction do not include the...

  37. [45]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: In Appendix H. 24 Guidelines: • The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but th...

  38. [46]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] Justification: Although we do not have theoretical result like a theory paper, but we provide the proof for...

  39. [47]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  40. [48]

    For the code, we will release it once this paper is accepted

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: They are all public, and det...

  41. [49]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: Yes, they are all discussed i...

  42. [50]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [NA] Justification: The error bars are not applicable in our settings. ...

  43. [51]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: They are all discussed ...

  44. [52]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [NA] Justification: It is not applicable to our settings. Guidelines: • The answer NA means that the...

  45. [53]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: They are discussed in the appendix. Guidelines: • The answer NA means that there is no societal impact of...

  46. [54]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: It is not appli...

  47. [55]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We cite all the ...

  48. [56]

    Guidelines: • The answer NA means that the paper does not release new assets

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] Justification: It is not applicable to our settings. Guidelines: • The answer NA means that the paper does not release new assets. •...

  49. [57]

    29 Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  50. [58]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  51. [59]

    Answer: [NA] Justification: It is not applicable to our settings

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.