Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A lightweight module can reduce bias in pre-trained text-to-image diffusion models without any bias annotations by steering sampling toward invariant semantic information.

desk verdict A practical, annotation-free debiasing recipe for diffusion models with real empirical results, but the theory section overclaims and the environment-inference step lacks identifiability. read the letter →

arxiv 2412.08480 v1 pith:C5QRUWER submitted 2024-12-11 cs.CV cs.IRcs.LG

classification cs.CVcs.IRcs.LG
keywords DiffusionModelDebiasInvariantLearningFairnessText-to-ImageGenerationBiasMitigationUnknownGuidance
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 claims that bias in pre-trained text-to-image diffusion models can be reduced without any bias labels, by learning a small module that guides the sampling process toward invariant semantic information. The core idea is to infer the hidden bias structure in the training data from how a frozen biased denoiser's loss varies per sample, then use a max-min objective to make the model focus on what stays the same across environments rather than spurious correlations. Only a lightweight module $G_\psi$ is trained, so the pre-trained model is left untouched. On Waterbirds, CelebA, and FairFace the method lowers the bias metric while keeping FID, Recall, and CLIP-T comparable to the biased backbone, and the paper offers a theoretical argument that the procedure reduces an upper bound on generalization error.

What carries the argument

The load-bearing object is the lightweight gradient-estimator module $G_\psi$ used as a mean-shift term in the reverse sampling process, plus the max-min training game that produces it. In the first phase, a learnable matrix $W\in\mathbb{R}^{N\times E}$ is optimized by maximizing $\mathrm{Var}_e(L_e)+\omega\min_e(L_e)$, where $L_e$ is the group-weighted denoising loss of the frozen biased model; this infers potential bias annotations without labels. In the second phase, the loss $\mathbb{E}\|\epsilon-\epsilon_\theta(x_t,t,y)+\Delta G_\psi(x_t,\Phi(y),t)\|^2+\lambda\,\mathrm{Var}_e(L_e)$ is minimized, with $\Phi(y)$ an encoder of the text prompt that plays the role of the feature encoder in invariant learning. The module shifts the predicted noise so sampling tracks invariant semantic information, and because only $G_\psi$ is trained the pre-trained diffusion model stays fixed.

What would settle it

Take a synthetic dataset with a known spurious correlation and set the group count $E$ to a value different from the number of sensitive-attribute combinations; if the soft grouping $W$ does not separate the known bias groups, or the bias metric stays near the biased baseline while generation quality collapses, the central claim would be refuted.

Watch

Extended reading notes

Core claim

InvDiff's central claim is that debiasing a diffusion model can be reframed as invariant learning for the sampling process. The authors model a biased model's output as fitting $P(X|Y)$ that includes spurious correlations, and an ideal model as fitting the invariant component $P(X^{\mathrm{inv}}|Y)$. They fill the gap between the two with a gradient-estimator module $G_\psi(x_t, \Phi(y), t)$ that approximates $\nabla_{x_t}\log p(x^{\mathrm{inv}}_y|x_t)$, taking the place of classifier guidance. A two-stage max-min objective first maximizes the variance of group-weighted denoising losses $L_e$ to infer a soft grouping matrix $W$ (bias annotations), then minimizes that same variance as an Environment Invariance Constraint while optimizing reconstruction, so the encoder $\Phi(y)$ captures invariant semantics. The authors prove, by bounding the error on an unseen target distribution with an $H$-divergence argument, that this procedure is equivalent to reducing the upper bound of generalization error.

Load-bearing premise

The load-bearing premise is that the per-sample denoising loss of the frozen biased model is a reliable measure of how spurious that sample is, so the groups found by maximizing the variance of $L_e$ across a user-chosen number of environments match the true bias structure.

Editorial extensions

If this is right

  • Debiasing can be done post hoc on an already-trained diffusion model, with no bias annotations and no unbiased training data, by fine-tuning only a small guidance network.
  • The same objective transfers beyond text-to-image models: on the AusElec time-series benchmark, InvDiff improves CRPS over the TimeGrad backbone and lowers prediction variance across domains.
  • Without any bias labels, the soft-grouping variant (InvDiff-Part-Soft) reaches bias reductions close to the hand-annotated variant on Waterbirds and FairFace, indicating that hidden bias structure can be recovered from loss variance alone.
  • Because the guidance module can have only 15M parameters and still lower bias, the method keeps memory and training time low enough for practical fine-tuning on a single GPU.

Reading between the lines

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

  • Editorial inference: if the loss-variance proxy for spuriousness holds, the same grouping mechanism should apply to biases with no obvious sensitive attribute, such as texture or style shortcuts, because it only needs the per-sample denoising loss to vary.
  • Editorial inference: the theory linking debiasing to a generalization upper bound predicts that bias reduction should track worst-group accuracy under subpopulation shift, a relation the paper does not test directly but could be checked on a dataset with multiple spurious correlations.
  • Editorial inference: the user-chosen group count $E$ is a practical tuning knob; the paper finds $E=2$ fails while $E=4$ and $E=8$ work, so an automatic selection of $E$ from the loss-variance profile would make the method more robust in deployment.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes InvDiff, a framework to reduce spurious correlations and bias in pre-trained text-to-image diffusion models without requiring explicit bias annotations. The method first infers pseudo-environment memberships by maximizing the variance of per-sample denoising losses of the frozen biased model (Eq. 8-9), then trains a lightweight module G_psi that provides gradient guidance and is regularized by an invariant-learning-style variance penalty (Eq. 10). Experiments on Waterbirds, CelebA, and FairFace show reduced bias metrics with largely maintained generation quality, and additional experiments examine data augmentation and time-series forecasting. The paper also claims a theoretical guarantee that implementing InvDiff is equivalent to reducing an upper bound on generalization error.

Significance. If fully substantiated, the paper addresses a practically important problem: debiasing generative models when the bias structure is unknown and bias labels are unavailable. The empirical evaluation is reasonably extensive, spans three benchmark datasets, and includes useful ablations and a data-augmentation comparison. The code is released, and the lightweight-module design is a practical strength. However, the central theoretical claim is not actually derived, and the environment-inference step has no identifiability guarantee; the paper's own Figure 9 indicates that the inferred groups do not isolate the constructed spurious attribute. As written, the contribution is best characterized as an empirically effective heuristic with an overclaimed theoretical justification.

major comments (4)
  1. [Section 4.4, Eq. (12)] The claimed theoretical guarantee is not established. Proposition 1 is quoted from prior work [35] and then mapped to InvDiff by informal analogy: the text asserts that maximizing L_e increases distributional disparity between groups and that minimizing L_e reduces the H-divergence term, but L_e is a scalar denoising loss per environment, not a distribution over X, and no formal bridge is provided between the objective in Eq. (10) and the four terms of the bound in Eq. (12). The abstract's statement that 'the implementation of InvDiff is equivalent to reducing the error upper bound of generalization' is therefore unsupported as written. The authors should either supply a real derivation or clearly downgrade this to an intuitive motivation.
  2. [Section 4.3, Eqs. (8)-(9); Appendix B, Figure 9] The environment-inference mechanism lacks identifiability. Maximizing variance of the frozen model's scalar denoising loss can separate easy from hard samples, frequent from rare prompts, or any other axis of difficulty, without necessarily recovering the intended spurious attribute. Figure 9 shows that for CelebA with omega=5 every inferred group contains a mixture of all four (hair color, gender) combinations, so the soft grouper does not isolate the constructed gender-hair color correlation. The discussion in Section 5.2, which attributes weaker CelebA results to 'many latent biases,' further concedes this point. The empirical bias reduction is still meaningful, but the mechanistic claim that the grouper recovers invariant versus spurious information needs support or substantial reframing.
  3. [Section 5.3 and Section 5.1.1] The no-annotation claim is weakened by the choice of E. The paper states that E can be set as the product of the numbers of categories in sensitive attributes, and the experiments use E=4,4,8 for Waterbirds, CelebA, and FairFace, matching the known gender/hair-color or gender/race structure. This means the proposed method relies on knowing which attributes are sensitive and how many categories they have, even if per-sample bias labels are not used. The authors should either justify that E can be selected without such knowledge (e.g., by showing robustness across a wide range of E without tuning to the true structure) or moderate the claim of operating 'without auxiliary bias annotations.'
  4. [Section 4.2, Eq. (7); Section 4.3, Eq. (10)] The relationship between the learned module G_psi and the intended invariant score is not specified. The text introduces the intractable term \nabla_{x_t} \log p(x_inv_y | x_t) and then states that a gradient estimator G_psi is used to simulate it, but no loss term, constraint, or analysis ensures that G_psi actually approximates this score. The objective in Eq. (10) only penalizes the variance of a denoising loss; there is no mechanism that ties G_psi to invariant semantic information x_inv_y. This is a load-bearing gap for the paper's central framing, even though the empirical results may still be explained by the regularization effect of the variance penalty.
minor comments (6)
  1. [Section 3.2] The subsection heading contains a typo: 'Inviariant Learning' should be 'Invariant Learning.'
  2. [Appendix A.1.1] The text says 'For the network architecture of \Delta, we choose a UNet,' but \Delta is a scalar hyperparameter; the trainable module is G_psi. This should be corrected.
  3. [Table 1] The values in parentheses for Bias and CLIP-T are presumably variances or standard deviations, but this is not stated in the table caption. Please define the notation.
  4. [Figure 9] The axis labels BLN-M, BLN-F, BLK-M, BLK-F are not expanded in the caption; please spell out (blond male, blond female, black-haired male, black-haired female) for readability.
  5. [Section 5.4] The AusElec time-series experiment would benefit from a citation to the WOODS benchmark and a brief description of how InvDiff is adapted to the TimeGrad backbone; currently the adaptation is only implicit.
  6. [General] The manuscript contains several formatting artifacts from the submission template, including placeholder page headers, ACM copyright placeholders, and glyph errors such as '/glyph1197ose' in Figure 4; these should be cleaned before publication.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation identified: the debiasing objective is fit to the frozen model's denoising loss, while the bias evaluation uses external classifiers; the Section 4.4 'guarantee' is an unsupported mapping rather than a definitional identity.

full rationale

The paper's central training objective is self-contained: Eq. 8-9 infer environment assignments W by maximizing the variance of the frozen biased model's denoising loss L_e, and Eq. 10 adds an invariance penalty over those groups. This is an EIIL-style min-max procedure, not a reduction of the predicted bias to the fitted quantity. The evaluation in Table 1 measures bias with external classifiers (hair color, race, background) on generated images, so the main empirical claim is not forced by construction. The theoretical section imports Proposition 1 from [35] and asserts that maximizing/minimizing L_e increases/decreases H-divergence terms; since L_e is a scalar loss, this mapping is not derived, but it is an unsupported assumption rather than Eq. X = Eq. Y, so it is a correctness risk, not circularity. The only self-citation ([10]) is an example of diffusion use in temporal modeling in the Introduction and is not load-bearing for the debiasing claim. The choice E=4, 4, 8 as the product of sensitive-attribute category counts (Section 5.3) weakens the 'no annotation' framing, but it is a group cardinality, not per-sample bias labels, and does not make the prediction equivalent to the fit. Hence no circular step; score 2 reflects the minor self-citation and the partially assumption-laden E selection, not a circular reduction.

Assumptions & free parameters 5 free parameters · 3 assumptions · 2 invented entities

The method depends on the unobserved decomposition of images into invariant and spurious parts, a hand-chosen set of hyperparameters (E, Delta, Lambda, Omega, G_psi size), and a transfer of invariant-learning assumptions to the generative setting. None of these are derived from first principles or validated independently.

free parameters (5)
  • E (number of environments) = 4, 4, 8 for Waterbirds, CelebA, FairFace
    Chosen a priori from the product of the number of categories of known sensitive attributes (Section 5.3), despite the claim of not using bias annotations; tuned via grid search over {2, 4, 8}.
  • Delta (guidance scale) = Range searched {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.8, 0.9, 1, 2, 10}
    Controls the magnitude of the guidance module's contribution to the noise prediction in Eq. (10); final per-dataset values not reported.
  • Lambda (regularization weight) = Range searched {0.2, 0.4, 0.6, 0.8, 1, 2, 5, 10, 20, 50, 100}
    Regularization weight for the environment invariance penalty in Eq. (10); affects the bias-quality trade-off.
  • Omega (dispersion degree) = not reported
    Dispersion degree in the grouping objective Eq. (8); controls the min-group coverage term; no final value given.
  • G_psi parameter quantity = chosen among {860M, 551M, 220M, 56M, 15M}
    The paper reports that 860M-551M maintain quality while 220M degrades; this is a design choice tuned on the task.
assumptions (3)
  • domain assumption Invariant learning assumption: environments vary spurious correlations while preserving causal mechanisms (EIC, Eq. 5), and this carries over to generative diffusion modeling.
    Section 3.2 introduces EIC for discriminative models and Section 4.3 transfers it to generative models without proving the analogous property holds for the diffusion denoising loss.
  • domain assumption The domain generalization bound of Proposition 1 (from [35]) applies to the diffusion model setting, and the grouping and regularization in InvDiff reduce its terms.
    Section 4.4 restates the bound and verbally assigns each term to a component of InvDiff; the required conditions (e.g., Eq. 11 defining O) are not shown to hold.
  • ad hoc to paper The text encoder Phi(y) captures invariant semantic information relevant to the prompt.
    Section 4.3 replaces the image encoder of invariant learning with an encoder on the prompt y; no argument is given that prompt-derived features are sufficient to separate invariant from spurious image content.
invented entities (2)
  • invariant semantic information x_inv_y (and its score p(x_inv_y | x_t))
    purpose: Conceptual target of the guidance module G_psi; the diffusion process is steered toward this unobservable quantity to remove spurious correlations.
    The paper defines x_inv_y as the part of the image that defines the prompt y, but never observes or measures it directly. G_psi is trained without any supervision from x_inv_y, so there is no falsifiable handle outside the method.
  • gradient estimator G_psi
    purpose: Approximates the intractable gradient of log p(x_inv_y | x_t) and is added to the noise prediction of the frozen diffusion model (Eq. 10).
    G_psi is a UNet trained with a modified denoising loss; the paper does not show that its output equals or approximates the claimed score, only asserts it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models." pith.science (2026). https://pith.science/paper/C5QRUWER

@misc{pith2026241208480,
  author       = {Pith},
  title        = {Pith review of: InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5QRUWER}},
  note         = {Machine review of arXiv:2412.08480}
}
read the original abstract

As one of the most successful generative models, diffusion models have demonstrated remarkable efficacy in synthesizing high-quality images. These models learn the underlying high-dimensional data distribution in an unsupervised manner. Despite their success, diffusion models are highly data-driven and prone to inheriting the imbalances and biases present in real-world data. Some studies have attempted to address these issues by designing text prompts for known biases or using bias labels to construct unbiased data. While these methods have shown improved results, real-world scenarios often contain various unknown biases, and obtaining bias labels is particularly challenging. In this paper, we emphasize the necessity of mitigating bias in pre-trained diffusion models without relying on auxiliary bias annotations. To tackle this problem, we propose a framework, InvDiff, which aims to learn invariant semantic information for diffusion guidance. Specifically, we propose identifying underlying biases in the training data and designing a novel debiasing training objective. Then, we employ a lightweight trainable module that automatically preserves invariant semantic information and uses it to guide the diffusion model's sampling process toward unbiased outcomes simultaneously. Notably, we only need to learn a small number of parameters in the lightweight learnable module without altering the pre-trained diffusion model. Furthermore, we provide a theoretical guarantee that the implementation of InvDiff is equivalent to reducing the error upper bound of generalization. Extensive experimental results on three publicly available benchmarks demonstrate that InvDiff effectively reduces biases while maintaining the quality of image generation. Our code is available at https://github.com/Hundredl/InvDiff.

Figures

Figures reproduced from arXiv: 2412.08480 by the authors.

Figure 1
Figure 1. Proportion and samples of water/land bird in ter [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of InvDiff. We first design a novel debiasing objective L𝑒 for diffusion models. Then we propose a max-min game with the debiasing objective. We first infer potential bias annotations by maximizing the objective. Given the annotations, we finetune the biased model to unbiased by minimizing the proposed objective. biased model to unbiased by minimizing the proposed objective. The overview of InvDiff is sh… view at source ↗
Figure 3
Figure 3. Images sampled from our unbiased model [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: InvDiff ’s results on multiple unknown biases. Bias Metric [33], which assesses the extent of bias in the results of the generative model. For every prompt P, we compute the bias(P) = 1 𝐾 (𝐾−1)/2 Í 𝑖,𝑗 ∈ [𝐾]:𝑖<𝑗 | freq(𝑖) − freq(𝑗)|, where freq(𝑖) is class𝑖’s frequency…
Figure 5
Figure 5. Figure 5: The Impact of 𝑮𝜓 Parameters on Model Performance and Training Time. attributes using 200,000 images collected from the CelebFaces At￾tributes Dataset 1 . We then assessed whether the images generated by our models on CelebA can mitigate these potential biases in Stable…
Figure 6
Figure 6. Figure 6: InvDiff for Time Series Forecasting. 5.3 Hyperparameter Sensitivity Analysis (RQ2) • Impact of Parameter Quantity of 𝑮𝜓 . In our settings, we fix the pretrained conditional diffusion model 𝝐𝜃 and train the lightweight learnable module 𝑮𝜓 . We investigate the impact of …
Figure 7
Figure 7. Figure 7: Analysis of debiasing at different levels of bias in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: Correspondence between real features and soft [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 8
Figure 8. Figure 8: Changes in bias under different Δ . E1 E2 E3 E4 BLK-F BLK-M BLN-F BLN-M E4-W0 E1 E2 E3 E4 E5 E6 E7 E8 BLK-F BLK-M BLN-F BLN-M E8-W0 E1 E2 E3 E4 BLK-F BLK-M BLN-F BLN-M E4-W3 E1 E2 E3 E4 E5 E6 E7 E8 BLK-F BLK-M BLN-F BLN-M E8-W3 E1 E2 E3 E4 BLK-F BLK-M BLN-F BLN-M E4-W5…
Figure 11
Figure 11. Figure 11: Samples from baseline TIW, trained on Waterbirds [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: The effect of 𝜆 on debias and image quality. B Experiments B.1 Hyperparameter Sensitivity Analysis • Analysis of the dispersion degree 𝜔 of the grouping. When the dispersion degree 𝜔 is set larger, data with the same true label are more likely to be dispersed into dif…

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. TarDiff: Target-Oriented Diffusion Guidance for Synthetic Electronic Health Record Time Series Generation

    cs.LG 2025-04 conditional novelty 6.0 of 10

    TarDiff guides diffusion-based synthetic EHR generation with a gradient-alignment signal computed from a guidance set, reporting improved downstream mortality and ICU-stay classification versus prior generative models.

Reference graph

Works this paper leans on

50 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [35]

    Anthony Sicilia, Xingchen Zhao, and Seong Jae Hwang. 2023. Domain adversarial neural networks for domain generalization: When it works and how to improve. Machine Learning 112, 7 (2023), 2685–2721

  2. [1]

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019)

  3. [2]

    Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. 2022. How well can Text-to-Image Generative Models understand Ethical Natural Language Inter- ventions?. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (Eds.). Association for Computational Linguistics,...

  4. [3]

    Bloomberg. 2023. The Bias in Generative AI: 2023 Report. https://www.bloomberg. com/graphics/2023-generative-ai-bias/ Accessed: 2024-08-06

  5. [4]

    Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, and Yanyan Lan. 2022. When Does Group Invariant Learning Survive Spurious Correlations?. In Advances in Neural Information Processing Systems, Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (Eds.). https://openreview.net/forum?id=ripJhpwlA2v

  6. [5]

    Marc Cheong, Ehsan Abedin, Marinus Ferreira, Ritsaart Reimann, Shalom Chal- son, Pamela Robinson, Joanne Byrne, Leah Ruppanner, Mark Alfano, and Colin Klein. 2024. Investigating Gender and Racial Biases in DALL-E Mini Images. 1, 2, Article 13 (jun 2024), 20 pages. https://doi.org/10.1145/3649883

  7. [6]

    Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 3043–3054

  8. [7]

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

Show all 50 references
  1. [8]

    Elliot Creager, Jörn-Henrik Jacobsen, and Richard Zemel. 2021. Environment inference for invariant learning. In International Conference on Machine Learning . PMLR, 2189–2200

  2. [9]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

  3. [10]

    Xinyao Fan, Yueying Wu, Chang Xu, Yuhao Huang, Weiqing Liu, and Jiang Bian. 2024. MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process. In The Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=CZiY6OLkt...

  4. [11]

    Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. 2023. Fair diffusion: Instruct- ing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893 (2023)

  5. [12]

    Jean-Christophe Gagnon-Audet, Kartik Ahuja, Mohammad Javad Darvishi Bayazi, Pooneh Mousavi, Guillaume Dumas, and Irina Rish. 2023. WOODS: Benchmarks for Out-of-Distribution Generalization in Time Series. Transactions on Machine Learning Research (2023). https://openreview.net/...

  6. [13]

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

  7. [14]

    Jonathan Ho and Tim Salimans. 2021. Classifier-Free Diffusion Guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications . https://openreview.net/forum?id=qw8AKxfYbI

  8. [15]

    Nathalie Japkowicz. 2000. The class imbalance problem: Significance and strate- gies. In Proc. of the Int’l Conf. on artificial intelligence , Vol. 56. 111–117

  9. [16]

    Kimmo Karkkainen and Jungseock Joo. 2021. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision . 1548–1558

  10. [17]

    Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. 2023. De-stereotyping text-to-image models through prompt tuning. ICML Workshop on Challenges in Deployable Generative AI (2023)

  11. [18]

    Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, and Il chul Moon. 2024. Training Unbiased Diffusion Models From Biased Dataset. In The Twelfth International Conference on Learning Representations . https://openreview.net/forum?id=39cPKijBed

  12. [19]

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. 2021. Out-of- distribution generalization via risk extrapolation (rex). InInternational Conference on Machine Learning. PMLR, 5815–5826

  13. [20]

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. 2019. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems 32 (2019)

  14. [21]

    Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. 2023. More control for free! image synthesis with semantic diffusion guidance. InProceedings of the IEEE/CVF Winter Conference on Applications of C...

  15. [22]

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. 2015. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision. 3730–3738

  16. [23]

    Wang Lu, Jindong Wang, Xinwei Sun, Yiqiang Chen, Xiangyang Ji, Qiang Yang, and Xing Xie. 2024. Diversify: A General Framework for Time Series Out-of- Distribution Detection and Generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 6 (2024), 4534–45...

  17. [24]

    Ranjita Naik and Besmira Nushi. 2023. Social Biases through the Text-to-Image Generation Lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’23) . Association for Computing Machinery, New York, NY, USA, 786–808. https://doi.org/10.1145/360021...

  18. [25]

    Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. 2020. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33 (2020), 20673–20684

  19. [26]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings...

  20. [27]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning . Pmlr, 8821–8831

  21. [28]

    Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. 2021. Au- toregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International Conference on Machine Learning . PMLR, 8857–8868

  22. [29]

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

  23. [30]

    Hashimoto, and Percy Liang

    Shiori Sagawa*, Pang Wei Koh*, Tatsunori B. Hashimoto, and Percy Liang. 2020. Distributionally Robust Neural Networks. InInternational Conference on Learning Representations. https://openreview.net/forum?id=ryxGuJrFvS

  24. [31]

    Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting

  25. [32]

    Preethi Seshadri, Sameer Singh, and Yanai Elazar. 2024. The Bias Amplification Paradox in Text-to-Image Generation. Annual Conference of the North American Chapter of the Association for Computational Linguistics (2024)

  26. [33]

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. 2023. Finetuning Text-to-Image Diffusion Models for Fairness.arXiv preprint arXiv:2311.07604 (2023)

  27. [34]

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. 2024. Finetuning Text-to-Image Diffusion Models for Fairness. In The Twelfth International Conference on Learning Representations . https: //openreview.net/forum?id=hnrB5YHoYu

  28. [36]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

  29. [37]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli

  30. [38]

    Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradients of the data distribution . Curran Associates Inc., Red Hook, NY, USA

  31. [39]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. In International Conference on Learning Representations. https://openreview.net/forum?id=PxTIG12RRHS

  32. [40]

    Vladimir N Vapnik. 1999. An overview of statistical learning theory. IEEE transactions on neural networks 10, 5 (1999), 988–999

  33. [41]

    Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip S. Yu. 2023. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Knowledge and Data Engineering 35, 8 (2023), 8052–8072. https://doi...

  34. [42]

    Shuliang Wang, Xinyu Pan, Sijie Ruan, Haoyu Han, Ziyu Wang, Hanning Yuan, Jiabao Zhu, and Qi Li. 2024. DiffCrime: A Multimodal Conditional Diffusion Model for Crime Risk Map Inference. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3212–3221

  35. [43]

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2023. Diffusion models: A comprehensive survey of methods and applications. Comput. Surveys 56, 4 (2023), 1–39

  36. [44]

    Tao Yang, Yuwang Wang, Yan Lu, and Nanning Zheng. 2023. DisDiff: Unsu- pervised Disentanglement of Diffusion Probabilistic Models. In Thirty-seventh Conference on Neural Information Processing Systems . https://openreview.net/ forum?id=3ofe0lpwQP

  37. [45]

    Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. 2023. Change is hard: a closer look at subpopulation shift. In Proceedings of the 40th International Conference on Machine Learning (ICML’23) . JMLR.org, Article 1652, 39 pages

  38. [46]

    Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. 2022. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning . PMLR, 25407–25437

  39. [47]

    Dauphin, and David Lopez-Paz

    Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. 2018. mixup: Beyond Empirical Risk Minimization. InInternational Conference on Learn- ing Representations. https://openreview.net/forum?id=r1Ddp1-Rb

  40. [48]

    CompVis/stable-diffusion-v1-4

    Zijian Zhang, Zhou Zhao, and Zhijie Lin. 2022. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in Neural Information Processing Systems 35 (2022), 22117–22130. A Appendix: Experimental Details A.1 Training Configuration A.1.1 Mode...

  41. [2015]

    In International conference on machine learning

    Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning . PMLR, 2256–2265

  42. [2023]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22522–22531

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.