GANs need large datasets; with little data the discriminator overfits and training collapses. This study isolates one variable — the adversarial loss — and measures what happens as data runs out.
Setup. A fixed DCGAN backbone (≈3.58 M generator / 2.77 M discriminator parameters) trained on class-balanced Tiny-ImageNet-200 subsets of 1,000, 5,000, 10,000 and 25,000 images: 4 losses × 4 sizes = 16 runs, identical optimiser and schedule, fixed validation set.
Findings.
- LSGAN is the most consistently stable — lowest generator-loss variance at every size (0.886 → 0.075 from 1k to 25k).
- WGAN with weight clipping collapses at 1,000 images (critic loss ≈ −5392, generator-loss std ≈ 918), despite being stable with more data.
- Vanilla GAN stays numerically stable but saturates at a persistently high generator loss (≈ 5–6).
- WGAN-GP regularises strongly but is expensive: ~557 s per epoch at 25k images versus ~27 s for LSGAN.
Takeaway. There is no universal winner. In a strictly low-data regime, LSGAN is the best compromise between adversarial stability and compute.