Enhancing SCGAN's Disentangled Representation Learning with Contrastive SSIM Similarity Constraints

SCGAN introduces a similarity constraint between generated images and conditions as a regularization term on generative adversarial networks. This constraint guides the generator network to recognize the differences in representations based on conditions. We have gained an in-depth understanding of how SCGAN operates, revealing that the similarity constraint behaves similarly to the contrastive loss function. To improve SCGAN, we modified the model by using SSIM to measure image similarity and incorporating contrastive loss principles within the similarity constraint. This resulted in enhanced performance on FID and FactorVAE metrics. On the MNIST and Fashion-MNIST datasets, the modified model demonstrates enhanced performance, achieving log-likelihood values that exceed those of the SCGAN model by approximately 1.0% and 2.6%, re-spectively, with scores of 234.8 and 332.6. The modified model demonstrates substantial improvements in FID scores compared to SCGAN. On the MNIST dataset, it achieves an FID of 3.42, representing a 16.79% reduction. On the Fashion-MNIST dataset, the modified model exhibits a smaller but still significant improvement, with an FID of 12.97, an 11.35% reduction over SCGAN's scores of 4.11 and 14.63, respectively. Regarding disentanglement, the modified model demonstrates significant enhancements, recording increases of approximately 15.6% and 2.2% in the disentanglement scores for the MNIST and Fashion-MNIST datasets, respectively. The modified model demonstrates improved disentangled representation and better generalizability compared to other models. ¹¹The code is available at https://github.com/Iman-yazdanpanah/contrastive-SSIM-SCGAN.