Research map / Image generation / GANs
Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UTLO, which makes the low-resolution generator layers of class-conditional GANs unconditional (with an extra unconditional GAN loss) so tail classes borrow head-class structure, improving long-tail cGAN generation.
Problem
Class-conditional GANs trained on long-tailed data mode-collapse on tail classes, producing low-quality and low-diversity samples when tail classes have very few training images.
Approach
Split the generator so low-resolution layers are trained with an unconditional GAN objective (no class information) while only high-resolution layers receive class conditioning, letting tail classes share class-independent low-resolution features learned from head classes; also adds few-shot FID/KID metrics for long-tail evaluation.
Key results
- On AnimalFaces-LT, UTLO reaches FID 26.2 and FID-FS 48.4 vs. 29.4 / 50.2 for NoisyTwins on StyleGAN2-ADA.
- On CIFAR10-LT with imbalance ratio 100, UTLO improves FID-FS to 13.4 and KID-FS to 5.4 over the StyleGAN2-ADA baseline (24.2 / 9.7).
- On LSUN5-LT (128x128, rho=1000), UTLO reduces FID from 15.0 to 10.9 and FID-FS from 60.2 to 43.6 over the Projected GAN baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add unconditional training at lower resolutions | StyleGAN2-ADA · AnimalFaces-LT | FID-FS: 50.2 -> 48.4 | Helped |
| Add unconditional training at lower resolutions | StyleGAN2-ADA · CIFAR10-LT (ρ=50) | FID-FS: 12.2 -> 11.8 | Helped |
| Add unconditional training at lower resolutions | StyleGAN2-ADA · CIFAR10-LT (ρ=100) | FID-FS: 14.1 -> 13.4 | Helped |
| Add unconditional training at lower resolutions | StyleGAN2-ADA · CIFAR100-LT | FID-FS: 22.5 -> 21.8 | Helped |
| Add unconditional training at lower resolutions | FastGAN · LSUN5-LT | FID-FS: 60.2 -> 43.6 | Helped |
| Use 8 × 8 for unconditional training resolution (ablation) | UTLO on StyleGAN2-ADA · AnimalFaces-LT | FID-FS: 50.3 -> 48.4 | Helped |
| Use 16 × 16 for unconditional training resolution (ablation) | UTLO on StyleGAN2-ADA · AnimalFaces-LT | FID-FS: 48.4 -> 50.3 | Hurt |
| Use 32 × 32 for unconditional training resolution (ablation) | UTLO on StyleGAN2-ADA · AnimalFaces-LT | FID-FS: 48.4 -> 64.9 | Hurt |
Surprising findings
- Unconditional StyleGAN2-ADA scores better on standard FID/KID on AnimalFaces-LT yet is much worse on few-shot metrics, suggesting FID/KID alone can be misleading for long-tail evaluation.
Limitations
- Requires that some tail and head classes be at least coarsely similar for low-resolution sharing to help.
- Tuning of low-resolution choice (8x8 vs 16x16 vs 32x32) and loss weight lambda is needed; performance drops when unconditional layers extend to 32x32.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT LSUN5-LT AnimalFaces-LT Flowers-LT iNaturalist2019
Benchmarks: CIFAR10-LT CIFAR100-LT LSUN5-LT AnimalFaces-LT
Metrics: FID KID FID-FS KID-FS
Baselines: StyleGAN2-ADA Transitional-cGAN GSR NoisyTwins StyleGAN2-ADA UnCond. PGAN (FastGAN)+DA
Builds on: StyleGAN2-ADA Projected GAN Transitional-cGAN Group Spectral Regularization NoisyTwins
Closest papers by meaning
- Efficient Conditional GAN Transfer With Knowledge Propagation Across Classes (CVPR 2021 · similarity 0.91)
- Collapse by Conditioning: Training Class-conditional GANs with Limited Data (ICLR 2022 · similarity 0.91)
- Sequential Training of GANs Against GAN-Classifiers Reveals Correlated "Knowledge Gaps" Present Among Independently Trained GAN Instances (CVPR 2023 · similarity 0.90)
- Self-Supervised GANs via Auxiliary Rotation Loss (CVPR 2019 · similarity 0.90)
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (ICLR 2019 · similarity 0.90)
- Structured Generative Adversarial Networks (NeurIPS 2017 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.