Efficient generative adversarial networks using linear additive-attention Transformers

Although deep generative models such as Diffusion Models (DMs) and Generative Adversarial Networks (GANs) achieve remarkable image generation performance, they often rely on computationally expensive architectures that limit accessibility and increase training and inference costs. We introduce LadaGAN, a hybrid Transformer-convolutional GAN built around Ladaformer, an efficient linear-attention Transformer block. Its linear additive-attention mechanism computes a single attention vector per head, avoiding the quadratic complexity of dot-product attention. By combining Ladaformer with convolutional layers in both the generator and discriminator, LadaGAN reduces computational complexity while improving the training stability of Transformer-based GANs. LadaGAN matches or surpasses the compared convolutional and Transformer GANs on benchmark datasets while being substantially more efficient, and remains competitive with multi-step generative models using orders of magnitude fewer computational resources. While maintaining competitive FID scores, LadaGAN achieves over 600x higher inference throughput than ADM (333 vs. 0.50 images/s on CelebA 64x64) and requires over 100x fewer FLOPs. Unlike one-step Consistency Models, LadaGAN can be trained efficiently on a single GPU.