Pith. sign in

REVIEW 6 major objections 6 minor 1 cited by

Contrastive Conditional-Unconditional Alignment for Long-tailed Diffusion Model

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-loss training scheme for class-conditional diffusion models—negatives-only contrastive repulsion on unconditional latents plus timestep-weighted alignment of conditional to unconditional noise estimates—improves tail-class diversity…

desk verdict A genuinely new regularizer for long-tailed diffusion with large reported gains, but the evaluation protocol—hyperparameters chosen on the test set, single-seed runs—means the headline numbers are not yet hard evidence. read the letter →

arxiv 2507.09052 v3 pith:ATHJPJXQ submitted 2025-07-11 cs.CV cs.LG

Fang Chen , Alex Villa , Gongbo Liang , Fuxing Li , Xiaoyi Lu , Meng Tang This is my paper
classification cs.CVcs.LG
keywords long-tailedimagegenerationdiffusionmodelsclass-conditionalcontrastivelearningmodecollapsetail-classdiversityconditional-unconditionalalignmentclassifier-freeguidance
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

Class-conditional diffusion models trained on long-tailed data collapse for rare classes: the few training images get memorized, and generated images become near-copies. This paper proposes a two-loss training scheme, CCUA, that pushes the unconditional branch of the model to spread its samples apart in latent space (an unsupervised contrastive loss with negatives only) and then squeezes the conditional branch toward the unconditional branch at the early, class-agnostic denoising timesteps (a timestep-weighted alignment loss). The claim is that this implicitly diversifies the conditional latents without the shortcut solution that direct conditional contrastive losses invite. The paper reports that on ImageNet-LT at 256×256, CCUA lowers overall FID from 19.9 to 15.1 and tail-class FID from 33.9 to 22.5 compared with SiT at 900k steps, with similar gains on TinyImageNet-LT, Places-LT, and CIFAR-LT using DDPM.

What carries the argument

The machinery is a pair of additive losses on the noise-estimation network, which is split into an encoder e(θ) (taking x_t and timestep t, with or without class c) that outputs a latent h, and a decoder d(θ) that outputs the noise estimate ε. The Unsupervised Contrastive Loss L_ucl is a negatives-only InfoNCE loss on unconditional latents h(∅), treating each sample as its own positive (numerator exp(1/τ)) and all other batch samples as negatives, which pushes latents apart uniformly over the hypersphere. The Alignment Loss L_al is the squared L2 distance between conditional and unconditional noise estimates, ||ε_θ(x_t,t,c) − ε_θ(x_t,t,∅)||², weighted by t/T so early denoising steps dominate; it is derived from the KL divergence between conditional and unconditional posteriors. The claimed synergy is that L_ucl diversifies the unconditional latent population that L_al then copies into the conditional branch, avoiding the shortcut solution of maximizing inter-class mutual information.

What would settle it

Compute, on the base pre-alignment model, the average early-timestep distance between conditional and unconditional noise estimates for the tail classes of ImageNet-LT; if those distances are already large, the alignment loss is collapsing class-specific structure rather than sharing diversity, and the reported FIDtail gains would not replicate on datasets where early denoising steps carry strong class identity.

Watch

Extended reading notes

Core claim

The paper's central claim is that mode collapse in long-tailed class-conditional diffusion can be cured by a contrastive conditional–unconditional alignment: instead of regularizing conditional latents directly, one diversifies unconditional latents with a negatives-only InfoNCE loss and then aligns conditional noise estimates to unconditional ones with weight t/T, so the initial denoising stages become class-agnostic. The authors argue that direct contrastive loss on conditional latents maximizes I(h_c;c) (inter-class variance) with a trivial per-class-constant solution, whereas contrastive loss on unconditional latents forces diversity across all images and the alignment loss distills that diversity into the conditional branch. They demonstrate the method on both U-Net and Diffusion Transformer backbones, reporting consistent FID, IS, KID, and FIDtail gains across five long-tailed datasets, with the largest relative gains on tail classes.

Load-bearing premise

The method assumes that in the first part of the denoising process, different classes look alike, so making conditional and unconditional generation match during that phase shares variety from common classes to rare classes instead of blurring away what makes each class distinct.

Editorial extensions

If this is right

  • On ImageNet-LT 256×256 with SiT, CCUA improves overall FID from 19.9 to 15.1 and tail-class FID from 33.9 to 22.5 at 900k steps, with gains already visible at 250k steps.
  • The recipe transfers across architectures: the same two losses improve DDPM on TinyImageNet-LT (FID 18.7→15.2, FIDtail 40.1→30.4) and Places-LT (FID 13.9→12.0), and on CIFAR10/CIFAR100-LT.
  • CCUA outperforms prior long-tailed diffusion treatments (CBDM, OCLT, DiffROP) and the concurrent Dispersive Loss, even beating the balanced-data DDPM reference under extreme 10-step DDIM sampling on CIFAR-LT.
  • Because the contrastive loss acts on the unconditional branch, unconditional generation also improves (e.g., CIFAR100-LT FID 18.53→15.97), indicating the diversity gain is not conditional-only.

Reading between the lines

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

  • A natural extension would be to make the alignment schedule adaptive: replacing the fixed t/T weight with a per-timestep or per-class schedule learned from the measured divergence between conditional and unconditional outputs could preserve class information when the class-agnosticness assumption weakens.
  • The mutual-information decomposition suggests a direct diagnostic: estimate I(h_c;c) and I(h_c;x|c) before and after training; CCUA predicts the intra-class term grows while the inter-class term is contained, whereas Dispersive Loss would predict a large inter-class jump.
  • The negatives-only contrastive loss depends on batch composition, so coupling CCUA with class-balanced negative sampling, rather than plain batch resampling, could concentrate repulsion where it is most needed; the paper does not test this.
  • The same two-loss template—diversify an unconditional side branch, align the conditional branch early in the process—could apply to other conditional generators, such as text-to-image models or class-conditioned video diffusion, whenever a long-tailed or few-shot condition collapses.
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

6 major / 6 minor

Summary. The paper proposes CCUA, a training-time regularization for class-conditional diffusion models trained on long-tailed data. The method adds two losses to the standard diffusion loss: an unsupervised contrastive loss (UCL) applied to the latents of the unconditional denoising stream, and an alignment loss (AL) that, at large timesteps, pulls the conditional noise estimate toward the unconditional noise estimate. The authors argue that the unconditional stream is easier to diversify and that aligning conditional generation to it transfers diversity from head classes to tail classes. They evaluate CCUA on ImageNet-LT, TinyImageNet-LT, Places-LT, and CIFAR-LT, with both SiT/DiT and DDPM/U-Net backbones, reporting consistent improvements over the baselines and several specialized long-tailed diffusion methods, with the headline result being FID 15.1 vs. 19.9 and FIDtail 22.5 vs. 33.9 for SiT on ImageNet-LT at 900k steps.

Significance. If the reported gains are real, CCUA would be a practically useful and easily implementable plug-in for long-tailed class-conditional generation. The paper is unusually thorough in covering two architecture families, four datasets, several imbalance factors, sampling-step regimes, and multiple ablations, and it reports training-time overhead. The core idea of aligning conditional and unconditional denoising only at early timesteps is a clean and interesting adaptation of the UTLO/transitional-GAN idea to diffusion models. However, the central claim is empirical, and the current evaluation protocol leaves room for selection on the test set and for seed noise; the significance of the paper therefore hinges on whether the comparison can be made more rigorous.

major comments (6)
  1. [Sec. 4.3, Table 5 and Appendix A.1] The headline hyperparameters are selected on the same ImageNet-LT test set used for the main comparison. Table 5 states that alpha = 0.05 and gamma = 0.05 were chosen 'for the best FID and IS', and Appendix A.1 states that a grid search over CFG guidance strength omega is run for each method on the evaluated test set. With no held-out validation split and, apparently, a single seed per configuration, the FID/FIDtail gains in Table 1 may partly reflect test-set selection rather than a robust property of CCUA. The authors should either adopt a validation split for hyperparameter and guidance-strength selection, or report the main results as mean and standard deviation over multiple seeds under a fixed protocol.
  2. [Sec. 4.1, Tables 1 and 2] FIDtail is computed on the last 30% of classes, where the number of real per-class images is very small; e.g., a tail class such as 'red wine' has 13 training images. FID estimates with such small reference sets are known to have high variance, and the paper reports no confidence intervals or repeated runs for any FIDtail value. The single-run gap of 22.5 vs. 33.9 in Table 1 therefore needs uncertainty quantification before it can support the central claim that CCUA improves tail-class fidelity and diversity.
  3. [Sec. 3.3, Eq. (7)] The derivation from Eq. (5) to Eq. (7) is not a constant-proportionality simplification. Substituting the mean parameterizations of Eq. (6) into Eq. (5) yields a t-dependent factor beta_t^2 / (alpha_t (1 - alpha_bar_t) sigma_t^2) multiplying the squared noise difference, not simply ||eps_cond - eps_uncond||^2. The timestep weight t/T used in Eq. (8) is therefore a heuristic choice rather than a direct consequence of the KL divergence. This should be stated explicitly; the empirical justification for t/T is otherwise fine.
  4. [Sec. 3.1 and Fig. 4] The alignment loss relies on the assumption that low-frequency components are similar across different classes at large timesteps, so that aligning conditional with unconditional generation transfers diversity without losing class-specific information. The paper supports this only with qualitative visualizations (Fig. 4 and Fig. 11) and cites FreeU. No quantitative evidence is given that this overlap holds for the specific tail classes in ImageNet-LT, TinyImageNet-LT, or Places-LT. Because the claimed mechanism is load-bearing for the method's motivation, the authors should provide a quantitative measure of cross-class low-frequency or latent similarity at large t for the datasets used.
  5. [Sec. 3.2, Eq. (2)] The mutual-information argument is informal. The claim that maximizing I(h^c; c) 'admits a trivial solution of having an identical latent for all images of the same class' is not proved, and the decomposition I(h^c; x, c) = I(h^c; c) + I(h^c; x | c) is presented without the needed caveats about the InfoNCE estimator. This does not invalidate the method, but it should be described as intuition rather than as a theorem, or the relevant conditions should be stated.
  6. [Sec. 3.4 and Eq. (9)] The final loss in Eq. (9) is written as Lccua = alpha * Lucl + gamma * Lal, but Algorithm 1 applies the same scaling factor 1/|B| to Lddpm, Lucl, and Lal. If the intended reading is that alpha and gamma already contain the normalization, this should be stated; otherwise the algorithm and the equation disagree about what is averaged over the batch.
minor comments (6)
  1. [Eq. (1)] The definition of pi_anc as exp(1/tau) implicitly assumes the latents are normalized; this should be stated before Eq. (1) rather than only in the prose after it.
  2. [Table 1] There is a stray superscript '5' in the CCUA row at 700k steps ('CCUA (ours) 5 140.5'), which appears to be a formatting artifact and should be removed.
  3. [Appendix A.1] The term 're-balanced factor 0.1' is used without a definition; the paper should specify how the factor enters the resampling probabilities.
  4. [Algorithm 1] The phrases 'Unconditional Training of CFG' and 'Conditional Training of CFG' are confusing; the branches refer to whether the current sample is trained with the null class or with its class label under classifier-free guidance, not to two kinds of CFG training.
  5. [Sec. 4.2] The sentence 'Our method achieves about 20% improvement on overall FID, 30% improvement on IS Score and FIDtail' is imprecise because the relative improvements vary by training step; please report the ranges or the specific steps to which these percentages refer.
  6. [Appendix A.4] In Table 12, the red parentheses for 'decline v.s. SiT baseline' are not applied consistently; e.g., Recall at 250k for '+ BRS' is shown with a red value but Precision at the same row is not color-coded, making the table harder to read.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed losses are defined independently of the reported metrics, and the empirical FID comparisons are external measurements; self-citations are motivational and not load-bearing.

full rationale

The core CCUA derivation is self-contained. The alignment loss is derived from a KL divergence between conditional and unconditional posteriors (Eqs. 3-5), simplified to an MSE between noise estimates (Eq. 7), and weighted by t/T (Eq. 8); the unsupervised contrastive loss is a standard negative-only InfoNCE on unconditional latents (Eq. 1). Neither loss is defined in terms of FID, FIDtail, IS, or any reported evaluation metric, so the reported improvements on ImageNet-LT, TinyImageNet-LT, Places-LT, and CIFAR-LT are external measurements rather than quantities forced by construction. The invocation of UTLO [20] (an author's prior GAN work) supplies motivation and a conceptual analogy, but the diffusion-specific alignment loss and its timestep weighting are implemented and evaluated independently; DiffROP [43] (also by an author) is used only as a baseline. The FreeU [37] low-frequency-similarity observation is an external citation supporting an assumption, not an imported conclusion. Concerns about hyperparameters (alpha, gamma) and CFG guidance being selected on the test set (Tables 5 and Appendix A.1) and single-seed runs are evaluation-validity issues, not circularity: they affect whether the measured gain is robust, but they do not make the reported metric equal to the loss or to a fitted parameter by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or architectural entities. It does introduce two new loss terms with several hyperparameters fitted to the test dataset, and its theoretical motivation relies on unproven assumptions about latent statistics during early denoising steps.

free parameters (7)
  • alpha (weight for unsupervised contrastive loss) = 0.05
    Selected by ablation on ImageNet-LT (Table 5) as the best FID and IS combination; controls the strength of repulsive diversity on unconditional latents.
  • gamma (weight for alignment loss) = 0.05
    Selected by ablation on ImageNet-LT (Table 5) together with alpha; controls how strongly conditional and unconditional denoising are aligned at large timesteps.
  • tau (temperature in UCL) = 0.1
    Kept at 0.1 as the default setting (Sec. 3.2); a hyperparameter of the InfoNCE-style loss, not tuned in the paper.
  • batch resample re-balanced factor = 0.1
    Used for ImageNet-LT and TinyImageNet-LT (Appendix A.1); controls how strongly tail classes are oversampled within each mini-batch.
  • latent layer for UCL (SiT block index) = N/4-th block for main results
    Chosen based on Table 13 ablation on ImageNet-LT; the N-th block gives better recall and sFID but worse FID, so N/4 was selected for best overall FID.
  • alignment loss time scheduler = linear (t/T)
    Chosen based on Table 14 ablation on ImageNet-LT after 900k steps; constant and quadratic schedules were worse on FID.
  • CFG guidance strength omega = grid-searched per method
    Appendix A.1 states 'we perform a grid search algorithm for each method to determine the optimal guidance strength'; this evaluation parameter affects all reported FID numbers.
assumptions (5)
  • standard math Standard DDPM and SiT training framework: forward diffusion, reverse process, and classifier-free guidance with label dropout.
    The method is built on the established diffusion formulation used throughout Sec. 3.3 and Algorithm 1.
  • standard math InfoNCE loss is a lower bound on mutual information, and minimizing it maximizes a lower bound on the mutual information between data and latents.
    Invoked in Sec. 3.2 to argue that contrastive losses increase diversity; the chain rule I(h;x,c)=I(h;c)+I(h;x|c) is used to criticize conditional contrastive loss.
  • domain assumption For the initial denoising steps (large t), low-frequency components of generated images are similar across different classes when starting from the same noise.
    Core motivation for the alignment loss; supported by Fig. 4 and cited FreeU [37], but not quantitatively verified on the specific datasets used.
  • domain assumption Aligning conditional and unconditional noise predictions at large t transfers diversity from head classes to tail classes.
    This is the mechanism behind the alignment loss; the paper provides no quantitative proof that knowledge transfer occurs, only the resulting FID improvements.
  • domain assumption Batch resampling with a rebalanced factor improves tail-class learning without degrading head classes when used with the proposed losses.
    Batch re-sampling is applied for ImageNet-LT and TinyImageNet-LT (Appendix A.1) and is shown to be needed for best FIDtail in the ablation (Table 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Conditional-Unconditional Alignment for Long-tailed Diffusion Model." pith.science (2026). https://pith.science/paper/ATHJPJXQ

@misc{pith2026250709052,
  author       = {Pith},
  title        = {Pith review of: Contrastive Conditional-Unconditional Alignment for Long-tailed Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ATHJPJXQ}},
  note         = {Machine review of arXiv:2507.09052}
}
abstract

Training data for class-conditional image synthesis often exhibit a long-tailed distribution with limited amount of images for tail classes. Such an imbalance causes mode collapse and reduces the diversity of synthesized images for tail classes. For class-conditional diffusion models trained with imbalanced data, we aim to improve the diversity and fidelity of tail class images without compromising the quality of head class images. We propose contrastive conditional-unconditional alignment (CCUA), which comprises two synergistic loss functions. Our first loss is an Alignment Loss (AL) that aligns class-conditional generation with unconditional generation at large timesteps. Alignment loss makes the denoising process insensitive to class conditions for the initial steps, which enriches tail classes through knowledge sharing from head classes. Secondly, we diversify unconditional generation via an Unsupervised Contrastive Loss (UCL) to increase the distance/dissimilarity among synthetic images. We combine the two losses to implicitly diversify conditional generation. Our framework is easy to implement as demonstrated on both U-Net based architecture and Diffusion Transformer. Our method outperforms vanilla denoising diffusion probabilistic models, score-based diffusion model, and alternative contrastive methods for class-imbalanced image generation across various datasets, in particular ImageNet-LT with 256$\times$256 resolution.

Figures

Figures reproduced from arXiv: 2507.09052 by the authors.

Figure 1
Figure 1. Generated images for a tail class (‘red wine’) with only 13 training images by (a) standard SiT [25] with diffusion loss and (b) proposed CCUA framework. Both models are trained on long-tailed ImageNet dataset for 900k steps with 256x256 resolution. (red wine) when trained using a long-tailed version of ImageNet. We aim to increase the fidelity and diversity of tail class images while maintaining the quality of head… view at source ↗
Figure 2
Figure 2. Variants of contrastive learning for regularizing latents of conditional & uncon￾ditional generation and our proposed CCUA framework. The black and blue dots are latents from class-conditional generation and grey dots are from unconditional genera￾tion. The latents represent intermediate features from a U-Net or transformer-based denoising network. Rather than applying standard supervised contrastive learning (a) an… view at source ↗
Figure 3
Figure 3. Implementation of the proposed CCUA framework. The noise estimation network is divided into a latent encode network e(∗) and a decode network d(∗). e(∗) encodes input xt to a low-dimensional latent h, which is decoded to noise ϵ by d(∗). We increase the distance of unconditional latents by Lucl with negative samples only, and align unconditional and conditional generation from the same sample x i t and utilize Lal t… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The leftmost column shows synthetic images with different classes but with the same initial noise or random seed. We visualize the low-frequency component and the high-frequency component from the reverse processing. It shows that low-frequency components are similar d…
Figure 5
Figure 5. Figure 5: Synthetic images of SiT and it with CCUA for ImageNet-LT tail classes ‘espresso’ and ‘window shade’. All methods start the denoising process from the same Gaussian noise at corresponding grid cells. CCUA shows consistently better diversity and fidelity [PITH_FULL_IMAG…
Figure 6
Figure 6. Figure 6: Model details of CCUA framework with UNet-based Model and Diffusion Transformer. As shown in [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: (a) Visualization of low-dimensional embeddings of five training images x for a tail class (‘worm’ in CIFAR100-LT) and synthetic images generated by DDPM [13] and our method. (b) We also show the distributions of real images and synthetic images generated by our method…
Figure 8
Figure 8. Figure 8: Synthetic images of SiT and CC UA for ImageNet-LT tail classes (from top-left to right-bottom: ‘bubble’, ‘redwine’, ‘comic book’ and ‘yawl’). All methods start the denoising process from the same Gaussian noise at corresponding grid cells. CCUA shows consistently highe…
Figure 9
Figure 9. Figure 9: More synthetic results for TinyImageNet-LT tail classes (from top-left to right￾bottom: ‘teapot’, ‘water tower’, ‘pretzel’, ‘mushroom’, ‘orange’, and ‘pizza’). Images in corresponding grid cell for DDPM and our method are initialized from the same Gaussian noise. Our m…
Figure 11
Figure 11. Figure 11: The low-frequency images with the same initial noise are very similar for [PITH_FULL_IMAGE:figures/full_fig_p028_11.png]
Figure 10
Figure 10. Figure 10: To see overfitting on tail classes, we find Top-10 nearest neighbors (3rd to 12th columns sorted by distances) among 1000 synthetic images to an anchor image (Leftmost column) in the training set. KNN is based on L2 distances of Inception V3 embeddings. For each examp…
Figure 11
Figure 11. Figure 11: Top: reverse process starting from the same initial Gaussian noise but with different class conditions (2nd-4th rows) or without condition (1st row). Middle: low￾frequency components of each noisy image. Bottom: high-frequency components of each noisy image [PITH_FUL…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. GRASP: Guided Residual Adapters with Sample-wise Partitioning

    cs.CV 2025-12 unverdicted novelty 5.0 of 10

    GRASP applies deterministic conditioning-space partitioning and sample-wise residual adapters to improve tail-class fidelity, diversity, and downstream utility in flow matching models, outperforming full fine-tuning a...

Reference graph

Works this paper leans on

47 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: International Conference on Artificial Intelligence and Statistics

    Ai, Q., Wang, P., He, L., Wen, L., Pan, L., Xu, Z.: Generative oversampling for imbalanced data via majority-guided vae. In: International Conference on Artificial Intelligence and Statistics. pp. 3315–3330. PMLR (2023)

    work page 2023

  2. [2]

    arXiv preprint arXiv:2112.03126 (2021)

    Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126 (2021)

    arXiv 2021

  3. [3]

    arXiv preprint arXiv:1801.01401 (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)

    arXiv 2018

  4. [4]

    In: ICLR (2019)

    Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. In: ICLR (2019)

    work page 2019

  5. [5]

    Advances in neural information processing systems32(2019)

    Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems32(2019)

    work page 2019

  6. [6]

    In: ICML (2020)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)

    2020

  7. [7]

    arXiv preprint arXiv:2210.11427 (2022)

    Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)

    arXiv 2022

  8. [8]

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

    Dalva, Y., Yanardag, P.: Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24209– 24218 (2024)

    work page 2024

Show all 47 references
  1. [9]

    Advances in neural information processing systems34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  2. [10]

    arXiv preprint arXiv:2407.00783 (2024)

    Fuest, M., Ma, P., Gui, M., Fischer, J.S., Hu, V.T., Ommer, B.: Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783 (2024)

    2024 arXiv

  3. [11]

    In: CVPR (2020)

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)

    2020

  4. [12]

    arXiv preprint arXiv:2210.02303 (2022)

    Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)

    2022 arXiv

  5. [13]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  6. [14]

    arXiv preprint arXiv:2207.12598 (2022)

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    2022 arXiv

  7. [15]

    Advances in Neural Information Processing Systems35, 8633–8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in Neural Information Processing Systems35, 8633–8646 (2022)

    2022

  8. [16]

    In: ICML (2021)

    Jiang, Z., Chen, T., Mortazavi, B.J., Wang, Z.: Self-damaging contrastive learning. In: ICML (2021)

    2021

  9. [17]

    Advances in neural information processing systems33, 12104–12114 (2020)

    Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. Advances in neural information processing systems33, 12104–12114 (2020)

    2020

  10. [18]

    In: CVPR (2019)

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)

    2019

  11. [19]

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

    Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007–6017 (2023) 16 Fang et al

    2023

  12. [20]

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

    Khorram, S., Jiang, M., Shahbazi, M., Danesh, M.H., Fuxin, L.: Taming the tail in class-conditional gans: Knowledge sharing via unconditional training at lower resolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7580–7590 (2024)

    2024

  13. [21]

    Advances in neural information processing systems33, 18661–18673 (2020)

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems33, 18661–18673 (2020)

    2020

  14. [22]

    Foundations and Trends®in Machine Learning12(4), 307–392 (2019)

    Kingma, D.P., Welling, M., et al.: An introduction to variational autoencoders. Foundations and Trends®in Machine Learning12(4), 307–392 (2019)

    2019

  15. [23]

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

    Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R.S., Indyk, P., Katabi, D.: Targeted supervised contrastive learning for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6918–6928 (2022)

    2022

  16. [24]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Stella, X.Y.: Open long-tailed recognition in a dynamic world. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)

    2022

  17. [25]

    In: European Conference on Computer Vision

    Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: European Conference on Computer Vision. pp. 23–40. Springer (2024)

    2024

  18. [26]

    Journal of machine learning research9(11) (2008)

    Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research9(11) (2008)

    2008

  19. [27]

    arXiv preprint arXiv:1807.03748 (2018)

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  20. [28]

    In: International Conference on Machine Learn- ing

    Ouyang, Y., Xie, L., Cheng, G.: Improving adversarial robustness through the contrastive-guided diffusion process. In: International Conference on Machine Learn- ing. pp. 26699–26723. PMLR (2023)

    2023

  21. [29]

    arXiv (2022)

    Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv (2022)

    2022

  22. [30]

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

    Qin, Y., Zheng, H., Yao, J., Zhou, M., Zhang, Y.: Class-balancing diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18434–18443 (2023)

    2023

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Rangwani, H., Bansal, L., Sharma, K., Karmali, T., Jampani, V., Babu, R.V.: Noisytwins: Class-consistent and diverse image generation through stylegans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 5987–5996 (2023)

    2023

  24. [32]

    In: European Conference on Computer Vision

    Rangwani, H., Jaswani, N., Karmali, T., Jampani, V., Babu, R.V.: Improving gans for long-tailed data through group spectral regularization. In: European Conference on Computer Vision. pp. 426–442. Springer (2022)

    2022

  25. [33]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  26. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22500–22510 (2023)

    2023

  27. [35]

    In: International Confer- ence on Learning Representations (2022),https://openreview.net/forum?id= 7TZeCsNOUB_ CCUA: Long-tailed Diffusion Models Training 17

    Shahbazi, M., Danelljan, M., Paudel, D.P., Gool, L.V.: Collapse by condition- ing: Training class-conditional GANs with limited data. In: International Confer- ence on Learning Representations (2022),https://openreview.net/forum?id= 7TZeCsNOUB_ CCUA: Long-tailed Diffusion Mode...

    2022

  28. [36]

    Shi, J.X., Wei, T., Xiang, Y., Li, Y.F.: How re-sampling helps for long-tail learning? Advances in Neural Information Processing Systems36, 75669–75687 (2023)

    2023

  29. [37]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4733–4743 (2024)

    2024

  30. [38]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  31. [39]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep- tion architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818–2826 (2016)

    2016

  32. [40]

    arXiv preprint arXiv:2012.04842 (2020)

    Tan, S., Shen, Y., Zhou, B.: Improving the fairness of deep generative models without retraining. arXiv preprint arXiv:2012.04842 (2020)

    2020 arXiv

  33. [41]

    arXiv preprint arXiv:2506.09027 (July 2025)

    Wang, R., He, K.: Diffuse and disperse: Image generation with representation regularization. arXiv preprint arXiv:2506.09027 (July 2025)

    2025 arXiv

  34. [42]

    arXiv preprint arXiv:2404.13040 (2024)

    Wang, X., Dufour, N., Andreou, N., Cani, M.P., Abrevaya, V.F., Picard, D., Kalogeiton, V.: Analysis of classifier-free guidance weight schedulers. arXiv preprint arXiv:2404.13040 (2024)

    2024 arXiv

  35. [43]

    arXiv preprint arXiv:2402.10821 (2024)

    Yan, D., Qi, L., Hu, V.T., Yang, M.H., Tang, M.: Training class-imbalanced diffusion model via overlap optimization. arXiv preprint arXiv:2402.10821 (2024)

    2024 arXiv

  36. [44]

    arXiv preprint arXiv:2410.06940 (2024)

    Yu, S., Kwak, S., Jang, H., Jeong, J., Huang, J., Shin, J., Xie, S.: Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940 (2024)

    2024 arXiv

  37. [45]

    arXiv preprint arXiv:1912.04486 (2019)

    Zhang, J., Liu, L., Wang, P., Shen, C.: To balance or not to balance: A simple- yet-effective approach for learning with long-tailed distributions. arXiv preprint arXiv:1912.04486 (2019)

    2019 arXiv

  38. [46]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836–3847 (2023)

    2023

  39. [47]

    In: The Twelfth International Conference on Learning Representations (2024) 18 Fang et al

    Zhang, T., Zheng, H., Yao, J., Wang, X., Zhou, M., Zhang, Y., Wang, Y.: Long- tailed diffusion models with oriented calibration. In: The Twelfth International Conference on Learning Representations (2024) 18 Fang et al. A Technical Appendices and Supplementary Material A.1 Imp...

    2024

Pith tools

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