acceptodds
Withdrawn from ICLR 2027

One Class at a Time: Single-Class Batch Sampling for Training Class-Conditional GANs on Limited Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Training conditional generative adversarial networks on limited data remains a challenging open problem, with severe mode collapse resulting from discriminator overfitting and insufficient per-class supervision. Existing approaches address this through data augmentation, regularisation, and knowledge sharing between classes, but the role of batch composition in shaping the discriminator's learned representations has received little attention. We investigate the effect of single-class batch sampling, wherein each training batch is composed exclusively of images from a single class, on conditional GAN training under limited data regimes. We argue that standard mixed-class batches allow the discriminator to exploit coarse inter-class differences as a shortcut for real/fake discrimination, reducing pressure to capture intra-class features and consequently weakening the signal provided to the generator. Single-class batches eliminate this shortcut, forcing the discriminator to attend to intra-class structure. Experiments across multiple limited data benchmarks demonstrate consistent improvements in quality and diversity. Our results suggest that batch composition is an overlooked but practically significant axis of design in conditional GAN training in limited data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
gan, conditional gan, limited data

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_576aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the training instability and mode collapse in class conditional GANs (cGANs) under limited data setup. The authors propose a simple approach based on single-class batch training to prevent conditional collapse and to improve the quality and diversity of the generated images. The authors evaluate their method on different dataset in balanced and imbalanced data setups, comparing to StyleGAN2+ADA as their main baseline, together with additional ablation studies. The paper suggests the proposed method is beneficial for training cGANs with balanced low-regime datasets. In the case of imbalanced dataset, single-class training is helpful when applied to represented classes, selected using a threshold hyperparameter.

Critical strengths
  • Simple and practical method: The proposed single-class batch strategy is easy to implement, does not require architectural changes, and yields consistent improvements in quality and diversity across several low-data settings.

  • Strong results in low-data regime: The improvements are substantial on some datasets, especially CUB-200-2011, suggesting the promise and effectiveness of the method.

  • Useful analysis of balanced vs. imbalanced regimes: The paper identifies when the method can hurt tail classes and proposes a threshold-based hybrid strategy for long-tailed datasets.

  • Robustness: The paper includes ablations over batch size, threshold choice, and split-class batching, supporting the robustness of the method to hyperparameters.

Critical weaknesses
  • Limited baseline comparison: The method is mainly compared against StyleGAN2+ADA, but the paper discusses several methods specifically designed for limited-data and long-tailed cGAN training (Shahbazi et al., ICLR 2022, Khorram et al., CVPR 2024). Including those baselines would help clarify the complementary effect of the proposed method and assess the claimed state-of-the-art performance.

  • Limited support for architecture-agnostic behavior: All experiments use StyleGAN2+ADA as the backbone, whose discriminator includes a minibatch standard-deviation layer. Since the proposed method directly changes batch statistics, part of the gain may be StyleGAN2-specific. Testing other architectures, e.g., R3GAN (Huang et al., NeurIPS 2024), Projected GAN (Sauer et al., NeurIPS 2021), StyleGAN-XL (Sauer et al., 2022), or StyleGAN3 (Karras et al., NeurIPS 2021), and/or ablating the impact of this layer would better support the architecture-agnostic claim.

Questions for authors

Related to critical weaknesses:

  • How does the method perform against prior approaches specifically designed for limited-data or long-tailed cGAN training, such as Shahbazi et al. (2022) and Khorram et al. (2024)?

  • Have the authors tested the proposed method on architectures other than StyleGAN2+ADA?

Reviewer_GdutOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a new technique for training class-conditional GANs with limited training data. The technique, called SCB (single-class batches) groups training data by the labeled class for more effective training of the GAN. The authors perform experiments with a single GAN architecture - StyleGAN2, and on several datasets, including ones with class imbalance. In the case of class imbalance, the authors find that SCB only helps in the classes with sufficient data, so they do not apply SCB on classes with limited data, where the minimum training data is defined by a threshold.

Critical strengths

Generative modelling on limited training data is an impactful problem that this work tackles.

Critical weaknesses

Core premise not well founded
The premise is not clear from the abstract/introduction. In the abstract the authors claim the mode collapse “resulting from discriminator overfitting and insufficient per-class supervision” but it's not well founded. There is a brief assertion in Section 2.2, but with no evidence or citation, and it describes memorization rather than the diversity collapse shown in Fig. 1. If mode collapse here is about diversity of generated images, it’s not completely clear to me why discriminator overfitting results in that. And if the mode collapse is about remembering training data, none of the metrics (FID, precision/recall, etc.) measure that aspect.

I’m unsure about the reasoning in section 4 - it seems like the authors blame confusion on early gradients. It says “If some of these directions are more salient or easier to learn (e.g., from classes with a greater number of samples), they tend to dominate earlier gradient signal to the generator.”. But I’m not sure how starting from fully unconditional training will help in that case, as it does in Transitional (Shahbazi et al. (2022)), which the paper admits in Introduction.

Section 4.3 is not entirely convincing, the claim is that tail classes need knowledge sharing but other classes with enough samples don’t. It seems like tail batches are only mixed with other tail batches and not the more common classes, why? Also, having a parameter for the absolute number of minimum samples per class is not ideal.

Experiments not sufficient
The authors cite previous work but never compare to any, I am not convinced all previous works are orthogonal, and even if they are, they should include that in experiments. Need experiments such as:

  • SCB vs Transitional,
  • SCB + Transitional, to test the orthogonality claim.

The paper should also ideally experiment with multiple GAN architectures. Especially because the paper claims knowledge sharing approaches are not architecture agnostic. I’m not convinced that this method would perform similarly on other architectures. In a BigGAN-style generator, single-class batches would change the conditional BatchNorm statistics, so agnosticism can't be assumed. StyleGAN2's discriminator has a minibatch-std layer that compares samples within a batch. That can interact with SCB.

Questions for authors
  1. Did you run experiments comparing with (or in combination with) other work towards training CGANs with limited training data?
  2. Did you run experiments on any other GAN architectures?
Reviewer_kqJcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates the problem of mode collapse and reduced generation quality in conditional generative adversarial networks (cGANs) trained with limited data. The authors propose using single-class batches (SCB) during training instead of conventional mixed-class batches. They argue that when the discriminator is trained on mixed-class batches, it may exploit shortcut features to distinguish between real and fake samples rather than focusing on fine-grained intra-class differences. By using single-class batches, the authors aim to encourage the discriminator to learn more precise intra-class real/fake distinctions and improve generation quality.

The proposed approach is evaluated using StyleGAN2 on several datasets. The authors report results using multiple evaluation metrics, including CLIP-FID, intra-class CLIP-FID,iFID-CLIP, precision, recall, and CMMD, to assess the quality and diversity of generated samples.

Critical strengths
  1. The paper is well-written, clearly presented, and easy to follow.
  2. The paper proposes a simple and intuitive approach to improving the training of conditional GANs using single-class batches.
  3. The authors evaluate the proposed method across multiple datasets and report improvements using various evaluation metrics, supporting its empirical effectiveness.
Critical weaknesses

Weaknesses:

  1. The methodological novelty seems limited, the proposed single batch training procedure is a simple modification of CGAN architecture.
  2. The experiments are conducted only using StyleGAN2. Evaluating the proposed method on other conditional GAN architectures, such as BigGAN, would help establish its application to other architectures.
  3. It would be good to report FID also as it can provide direct comparison to StyleGAN2 results.
Questions for authors

Please refer weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.