Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training

1 Introduction

• Conditional GANs use dataset category information and include classifier-based and projection-based approaches.The introduction describes conditioning as a standard approach for high-quality image generation.Source: §1 · p. 1
• ACGAN can collapse early as class counts rise and generate easily classified images at the cost of diversity.Projection-based GANs have achieved strong results on many-category datasets, motivating the paper’s comparison with them.Source: §1 · pp. 1–2
• ReACGAN combines unit-hypersphere projections with D2D-CE, using data-to-class, data-to-data, and mini-batch class-to-class relationships for supervision.The authors attribute early gradient explosion to unbounded input features and poor classifier performance during early training.Source: §1 · p. 2
• ReACGAN beats classifier- and projection-based GANs, improving state-of-the-art FID by 2.5%, 15.8%, 5.1%, and 14.5% on CIFAR10, Tiny-ImageNet, CUB200, and ImageNet, respectively.The paper also reports benefits from consistency regularization and differentiable augmentations, and compatibility between D2D-CE and StyleGAN2.Source: §1 · p. 2

2 Background: Generative Adversarial Networks

• A GAN maps latent variables into data space with a generator, while a discriminator distinguishes real samples from generated ones.The paper characterizes GANs as implicit generative models.Source: §2 · p. 2
• Conditional GANs divide into classifier-based and projection-based families; ACGAN uses softmax cross-entropy but can become unstable as class counts grow and lose diversity.The text notes that earlier ACGAN variants do not resolve these issues on datasets with many categories.Source: §2 · p. 3
• Projection discriminators improved ImageNet generation and motivated projection-based GANs as a standard, while this paper revisits ACGAN to address its instability.The paper presents ReACGAN as its approach to high-quality and diverse image generation.Source: §2 · p. 3

3.1 Feature Normalization

• ACGAN’s early classifier gradients can explode as feature norms grow and class probabilities remain low, causing collapse; unit-hypersphere normalization resolves it.The paper also reports that ACGAN’s FID stops decreasing after its gradient norms fall.Source: §3.1 · pp. 3–4

3.2 Data-to-Data Cross-Entropy Loss (D2D-CE)

• D2D-CE replaces proxy similarities in the denominator with similarities to different-label mini-batch samples and adds margins to suppress easy positives and negatives.A false-negative mask is used when calculating negative-sample similarities.Source: §3.2 · pp. 4–5

3.3 Useful Properties of Data-to-Data Cross-Entropy Loss

• D2D-CE implicitly mines hard negatives, while sufficiently easy positive and negative samples produce no gradient updates.The classifier therefore focuses on hard examples rather than being dominated by easy ones.Source: §3.3 · p. 5

3.4 Rebooted Auxiliary Classifier Generative Adversarial Networks (ReACGAN)

• ReACGAN jointly trains an adversarial objective and D2D-CE, alternating discriminator and generator updates to produce diverse, high-quality images without early-training collapse.The discriminator uses real images for the D2D-CE objective; the generator seeks images that deceive it and yield low D2D-CE.Source: §3.4 · pp. 5–6

4.1 Datasets and Evaluation Metrics

• Experiments cover CIFAR10, Tiny-ImageNet, CUB200, ImageNet, and AFHQ, with IS, FID, F0.125, and F8 used to assess generation.IS and FID are used together because IS tends to measure fidelity while FID tends to weight diversity; F0.125 and F8 estimate precision and recall.Source: §4.1 · p. 6

4.2 Experimental Details

• Comparisons use shared backbones and generator conditioning where possible, a common evaluation protocol, and three-run averages except for ImageNet and StyleGAN2 experiments.The authors search temperature and positive-margin settings, reporting robustness to hyperparameter choice; the reference split differs for CIFAR10 and CUB200.Source: §4.2 · p. 7, §4.3 · p. 7

4.3 Evaluation Results

• ReACGAN surpasses previous methods by 2.5%, 5.1%, 15.8%, and 14.5% in FID on CIFAR10, Tiny-ImageNet, CUB200, and ImageNet (256 B.S.), respectively.At ImageNet batch size 2048, ReACGAN reaches FID 8.23, comparable with 7.89 for the StudioGAN BigGAN implementation; the ReACGAN result uses two runs.Source: Table 1 · p. 7, §4.3 · p. 8, Table 2 · p. 8
• D2D-CE gives consistent results across datasets, with the lowest FID and highest F8 and F0.125 in most cases among the compared conditioning losses.AC- and multi-hinge-based models collapse early on Tiny-ImageNet and CUB200, while projection, conditional contrastive, and D2D-CE losses perform satisfactorily across datasets.Source: §4.3 · p. 8, Table 3 · p. 9
• ReACGAN outperforms projection-discriminator and conditional-contrastive counterparts with three adversarial losses, but performs worse than the projection discriminator with least-squares loss.The authors speculate that least-squares training may affect feature-map norms and impair D2D-CE classifier training.Source: §4.3 · p. 8, Table 4 · p. 9
• On CIFAR10, StyleGAN2 with D2D-CE produces a 13.9% better generation result than conditional StyleGAN2, and ADA or DiffAug can reinforce it for limited data.StyleGAN2 with DiffAug, D2D-CE, and tuning achieves the best reported CIFAR10 performance in that comparison.Source: §4.3 · p. 9, Table 5 · p. 9

4.4 Ablation Study

• Feature and weight normalization stabilize ACGAN training and improve Tiny-ImageNet and CUB200 results; D2D-CE adds data-to-data relationships and easy-sample suppression.Suppression does not work well with feature-normalized cross-entropy; ACGAN has high ImageNet IS but worse FID, F0.125, and F8 than the model with all three components.Source: §4.4 · pp. 9–10, Table 6 · p. 10, Figure 5 · p. 10

5 Conclusion

• ACGAN instability as class counts rise is attributed to exploding classifier gradients caused by unbounded input features and weak early-stage classification.The authors derive the classifier gradient analytically and numerically check its values.Source: §5 · p. 10
• The paper proposes D2D-CE and ReACGAN to alleviate ACGAN instability and reinforce its generation performance.Source: §5 · p. 10
• Experiments report ReACGAN superiority over classifier- and projection-based GANs on five benchmarks, with robustness to hyperparameter selection and compatibility with architectures and differentiable augmentations.Source: §5 · p. 10

A Algorithm

• The training algorithm alternates discriminator and generator updates using adversarial and D2D-CE losses, then applies Adam to their combined objectives.The discriminator's D2D-CE term uses real images, while the generator's uses generated images; both use the stated temperature and margin settings.Source: A Algorithm · p. 15

B Software: PyTorch-StudioGAN

• PyTorch-StudioGAN expands StudioGAN to address limited reliable open-source implementations and configuration support for GAN losses, architectures, regularizations, augmentations, and metrics.The library lists implementations including ReACGAN, adversarial losses, differentiable augmentations, and evaluation metrics.Source: B Software: PyTorch-StudioGAN · p. 15
• The software appendix also lists training utilities such as mixed precision, distributed and data parallelism, synchronized batch normalization, and latent-sampling tools.Source: B Software: PyTorch-StudioGAN · p. 15

C.1 Datasets

• CIFAR10 contains 60k 32×32 RGB images across 10 classes, split into 50k training and 10k test images.Source: §C.1 · p. 16
• Tiny-ImageNet has 120k 64×64 RGB images across 200 categories, with 500 training images per class, making it more challenging than CIFAR10.Source: §C.1 · p. 16
• CUB200 provides about 12k fine-grained bird images across 200 classes, resized to 128×128 to test generation in a limited-data setting.Images are center-cropped using a square box matching each image’s short side.Source: §C.1 · p. 16
• ImageNet has about 1,281k training and 50k validation RGB images; AFHQ has 14,630 training and 1,500 validation 512×512 images across three animal classes.All datasets use horizontal flips and pixel normalization to [-1, 1].Source: §C.1 · p. 16

C.2 Hyperparameter Setup

• Hyperparameters follow specified settings or StudioGAN configurations; ReACGAN’s temperature and positive margin are selected by linear search, with stable performance across tested values.The search tests τ ∈ {0.125, 0.25, 0.5, 0.75, 1.0} and mp ∈ {0.5, 0.75, 0.9, 0.95, 0.98, 1.0}; λ equals τ except for ImageNet.Source: §C.2 · pp. 16–17, Table A1 · p. 16, Figure A1 · p. 17

• D2D-CE uses normalized reference and proxy embeddings, along with similarities to negative samples in the mini-batch.N(i) denotes the indices of negative samples in the mini-batch.Source: D Proofs of Properties of D2D-CE · p. 17
• When one negative similarity exceeds another, its loss derivative is at least as large, so D2D-CE performs hard-negative mining.The derivative grows exponentially as the negative similarity increases.Source: D Proofs of Properties of D2D-CE · pp. 17–18
• D2D-CE suppresses positive-similarity gradients when sq ≥ mp and negative-similarity gradients when sq,r ≤ mn.Source: D Proofs of Properties of D2D-CE · pp. 17–18
• When both suppression conditions hold, D2D-CE reaches its stated global minimum, log(1 + |N(i)|).Source: D Proofs of Properties of D2D-CE · pp. 17–18

• Feature normalization resolves the early-training collapse on CUB200, while adding data-to-data relationships provides further performance gain.The experiments also attribute the gain to easy-sample suppression.Source: E Additional Experimental Results · p. 18, Figure A2 · p. 19
• Classifier target probabilities continue rising even when FID stops improving, suggesting ACGAN can become biased toward label classification over adversarial training.Source: E Additional Experimental Results · pp. 18–19, Figure A3 · p. 19
• ReACGAN fails to approximate a 1-D mixture of Gaussians with overlapping supports, but adding TAC enables exact approximation.The reported limitation is tied to conditional distributions with non-overlapping supports being accurately approximated.Source: E Additional Experimental Results · p. 19, Figure A4 · p. 20
• On CIFAR10 and Tiny-ImageNet, ReACGAN + TAC gives comparable or marginally better results than ReACGAN; the authors suggest the benchmarks may have non-overlapping conditional distributions.Source: E Additional Experimental Results · p. 20, Table A2 · p. 20
• ReACGAN provides consistent generation results with Deep CNN and ResNet-style architectures on CIFAR10 and Tiny-ImageNet.Source: E Additional Experimental Results · p. 20, Table A3 · p. 20
• More negative samples benefit D2D-CE generation results; the authors also report gradient signals growing exponentially as negative-sample counts increase linearly.Source: E Additional Experimental Results · pp. 20–21, Table A4 · p. 21
• Lowering classification strength or normalizing feature maps prevents ACGAN collapse, but classifier gradient clipping does not; normalization and D2D-CE perform better than the other tested cures.The authors speculate that clipping may constrain class probabilities and lead to inaccurate gradients.Source: E Additional Experimental Results · p. 21, Table A5 · p. 21
• On ImageNet with batch size 2048, ReACGAN takes 18m 52s per 100 generator updates, versus 17m 37s for BigGAN and 18m 24s for ContraGAN.The authors describe this as a slight computational overhead.Source: E Additional Experimental Results · p. 21

F Analysis of the differences between ReACGAN and ContraGAN

• ContraGAN’s 2C loss derives from NT-Xent and includes multiple positive samples and false negatives, while D2D-CE avoids easy-positive mining and false-negative attenuation.Source: F Analysis of the differences between… · pp. 21–23
• On ImageNet, ReACGAN has lower FID and higher IS than BigGAN and ContraGAN, but slightly lower Top-1 and Top-5 accuracy than BigGAN; ContraGAN’s Top-1 accuracy is 2.866%.The authors interpret ReACGAN as approximating the overall distribution with a slight loss of exact conditioning; they report ContraGAN can produce visually satisfactory but undesirably conditioned images.Source: F Analysis of the differences between… · p. 23, Table A6 · p. 23

G Potential Negative Societal Impacts

• Photorealistic GAN outputs can be misused for deepfakes, sexual harassment, and attacks on machine-vision applications; proposed responses include detection methods, policies, and ethical awareness.Source: G Potential Negative Societal Impacts · p. 24

H Computation resources

• Table A7 reports approximate total training time in days for reference.Source: Table A7 · p. 24

• Experiments generally average best performances across three random-seed runs; ImageNet is the exception, and standard deviations are provided for reference.Source: I Standard Deviations of Experiments · p. 24
• Table A8 compares classifier-based and projection-based GANs on CIFAR10, Tiny-ImageNet, and CUB200 using IS, FID, F0.125, and F8, with standard deviations across three runs.Source: Table A8 · p. 25
• Tables A9 and A10 examine D2D-CE effectiveness against other conditioning losses and its consistency across adversarial-loss choices.Source: Table A9 · p. 25, Table A10 · p. 25
• Table A11 ablates normalization, data-to-data consideration, and suppression of easy negative and positive samples.Source: Table A11 · p. 26
• Table A12 investigates D2D-CE across architectures on CIFAR10 and Tiny-ImageNet.Source: Table A12 · p. 26
• Table A13 ablates the number of negative samples.Source: Table A13 · p. 26

• ReACGAN with StyleGAN2 and ADA has FID=4.95 on AFHQ, while StyleGAN2 with ADA has FID=4.99.Source: Figure A5 · p. 26, Figure A6 · p. 27
• At batch size 2048 on ImageNet, ReACGAN has FID=8.23 and BigGAN has FID=7.89.Source: Figure A7 · p. 27, Figure A8 · p. 28
• At batch size 256 on ImageNet, ReACGAN has FID=13.98, ContraGAN 25.16, BigGAN 16.36, and ACGAN 26.35.Source: Figure A9 · p. 28, Figure A10 · p. 29, Figure A11 · p. 29
• On CUB200, ReACGAN has FID=14.67, compared with ContraGAN at 20.89, BigGAN at 17.80, and ACGAN at 61.29.Source: Figure A13 · p. 30, Figure A14 · p. 31, Figure A15 · p. 31
• On Tiny-ImageNet, ReACGAN has FID=26.82, compared with ContraGAN at 28.41, BigGAN at 31.92, and ACGAN at 61.50.Source: Figure A17 · p. 32, Figure A18 · p. 33, Figure A19 · p. 33
• Figure A21 presents generated images on CIFAR10.Source: Figure A21 · p. 34