teasr: training-efficient any-step diffusion transformer for real-world image super-resolution

2026-06-15

AI Summary

This work addresses the limitations of existing diffusion models in real-world image super-resolution, which suffer from slow inference, reliance on external teacher models, and inflexible sampling steps. The authors propose TEASR, a novel framework featuring the first self-adversarial distillation mechanism within a single model, eliminating the need for additional teacher models or discriminators while supporting arbitrary-step sampling. TEASR introduces a timestep-aware correction strategy and a timestep-decoupled dual-branch diffusion Transformer architecture, significantly enhancing both training efficiency and reconstruction quality. Experiments demonstrate that TEASR outperforms state-of-the-art methods across multiple benchmarks, enabling efficient distillation of a 20-billion-parameter model on a single GPU and achieving high-quality results in both one-step and multi-step super-resolution settings.

Abstract

Diffusion models excel in Real-World Image Super-Resolution (Real-ISR) due to their powerful generative priors but suffer from slow iterative sampling. Although existing one-step distillation methods accelerate inference, they typically require auxiliary teacher models that inflate training memory and restrict scalability to large-scale architectures. Furthermore, these fixed-step models lack the flexibility to trade off speed for quality. In this paper, we propose TEASR, a training-efficient any-step diffusion framework for Real-ISR that enables both one-step and multi-step restoration within a unified model. Our key idea is to perform self-adversarial distillation within a single diffusion model, eliminating the need for auxiliary teachers or discriminators. Specifically, we propose a timestep-aware rectification strategy that stabilizes one-step generation across noise levels. These two designs further enables the distillation of 20B-parameter diffusion models on a single GPU, significantly improving training efficiency. Moreover, we introduce a dual-branch diffusion transformer with decoupled timestep condition to separate the current noise state and the denoising target to enhance sampling quality. Extensive experiments demonstrate that TEASR supports seamless any-step sampling and consistently outperforms state-of-the-art methods across multiple datasets.

Problem

• diffusion models
• real-world image super-resolution
• training efficiency
• any-step sampling
• model scalability

Innovation

• any-step diffusion
• self-adversarial distillation
• timestep-aware rectification
• dual-branch diffusion transformer
• training-efficient super-resolution