Diffusion versus GAN Synthesis for Wafer Map Defect Classification: A Preliminary Cross-Backbone Study on WM-811K

Synthetic data augmentation is the dominant strategy for class imbalance in wafer map defect classification on the WM-811K benchmark, with generative adversarial networks (GANs) and diffusion models reported in separate prior studies. To our knowledge, no work has compared these two families under a single controlled protocol on this benchmark. We present a cross-backbone evaluation of a two-stage hinge-loss GAN (TS-HingeGAN) and Stable Diffusion fine-tuned with class-conditioned LoRA (DB-SD-LoRA) on eight ImageNet-pretrained backbones. Holding split, seed, per-class synthetic budget, and evaluation set fixed across 24 cells (8 backbones x 3 conditions) under stratified 5-fold CV with real-only test partitions, we vary only the synthesis source and the backbone. DB-SD-LoRA attains pooled Clean-FID 51.0 versus 106.4 for TS-HingeGAN. Downstream gains are backbone-dependent: lightweight CNNs gain 2.5-3.6 Macro-F1 points from diffusion, while the strongest backbones (TinyViT-21M, ConvNeXtV2-T) gain only 0.6-0.8 points - a saturation pattern that prior single-backbone studies cannot reveal. Our best cell, TinyViT-21M + DB-SD-LoRA + ECOC-SVM, attains 0.9459 +/- 0.0074 Macro-F1 under 5-fold CV, exceeding the matched-protocol 0.9399 of SCRBLAA-Net.