Q-DiT4SR: Exploration of Detail-Preserving Diffusion Transformer Quantization for Real-World Image Super-Resolution

Recently, Diffusion Transformers (DiTs) have emerged in Real-World Image Super-Resolution (Real-ISR) to generate high-quality textures, yet their heavy inference burden hinders real-world deployment. While Post-Training Quantization (PTQ) is a promising solution for acceleration, existing methods in super-resolution mostly focus on U-Net architectures, whereas generic DiT quantization is typically designed for text-to-image tasks. Directly applying these methods to DiT-based super-resolution models leads to severe degradation of local textures. Therefore, we propose **Q-DiT4SR**, the first PTQ framework specifically tailored for DiT-based Real-ISR. We propose **H-SVD**, a hierarchical SVD that integrates a global low-rank branch with a local block-wise rank-1 branch under a matched parameter budget. We further propose **V**ariance-**a**ware **S**patio-**T**emporal **M**ixed **P**recision: **VaSMP** allocates cross-layer weight bit-widths in a data-free manner based on rate-distortion theory, while **VaTMP** schedules intra-layer activation precision across diffusion timesteps via dynamic programming (DP) with minimal calibration. Experiments on multiple real-world datasets demonstrate that our Q-DiT4SR achieves SOTA performance under both **W4A6** and **W4A4** settings. Notably, the W4A4 quantization configuration reduces model size by **5.8**$\times$ and computational operations by **6.14**$\times$. Our code and models will be available at https://github.com/xunzhang1128/Q-DiT4SR.

Many real-world photos are blurry, noisy, or low-resolution, and image super-resolution aims to turn them into sharper and more detailed images. Recently, Diffusion Transformers have shown strong ability to generate realistic textures for this task, but they are usually large and computationally expensive, which makes them difficult to deploy in real-world applications. This paper proposes Q-DiT4SR, a compression framework designed specifically for Diffusion Transformer-based real-world image super-resolution. Directly using existing compression methods can noticeably damage local details such as hair, fabric, text, or building edges. To address this, our method carefully preserves important image details while reducing the numerical precision used inside the model. It decides which parts of the model need higher precision and which parts can be compressed more aggressively, and it also uses a lightweight correction strategy to better maintain both overall image structure and fine local textures. Experiments on several real-world image super-resolution benchmarks show that Q-DiT4SR produces higher-quality results than previous compression methods. Under strong compression settings, it significantly reduces both model size and computational cost, making Diffusion Transformer-based image enhancement more practical for real-world use.