Efficient One-Step Diffusion Restoration Model with Compact Token Compression and Linear Attention
SANA-SR is an efficient one-step image super-resolution model that tackles high computational costs by using a deep compression autoencoder to drastically reduce latent tokens. It then uses a linear-attention DiT backbone with LoRA fine-tuning for restoration. This approach achieves superior quality on real images while maintaining fast inference speeds suitable for mobile deployment.