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

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

TEASR, a new training-efficient any-step diffusion transformer, addresses the slow iterative sampling of diffusion models in Real-World Image Super-Resolution (Real-ISR). Unlike existing one-step distillation methods that require auxiliary teacher models, TEASR employs self-adversarial distillation within a single diffusion model, eliminating the need for extra teachers or discriminators. This approach, combined with a timestep-aware rectification strategy, stabilizes one-step generation across noise levels and significantly improves training efficiency, enabling the distillation of 20B-parameter diffusion models on a single GPU. Furthermore, TEASR introduces a dual-branch diffusion transformer with a decoupled timestep condition to enhance sampling quality by separating the current noise state and the denoising target. Experiments show TEASR supports seamless any-step sampling and outperforms state-of-the-art methods across multiple datasets.

Key takeaway

For Machine Learning Engineers developing Real-ISR solutions, TEASR offers a compelling alternative to traditional diffusion models. If you are constrained by GPU memory or need flexible inference speeds, consider adopting TEASR's self-adversarial distillation to train large models efficiently on a single GPU. This approach allows you to achieve superior image quality with adaptable sampling steps, optimizing both resource use and output fidelity for your applications.

Key insights

TEASR uses self-adversarial distillation within a single diffusion model for efficient, flexible Real-ISR, eliminating auxiliary teachers.

Principles

• Self-adversarial distillation reduces training overhead.
• Any-step sampling allows speed-quality trade-offs.
• Decoupled timestep conditions enhance sampling quality.

Method

TEASR performs self-adversarial distillation within a single diffusion model, employing a timestep-aware rectification strategy to stabilize one-step generation across noise levels.

In practice

• Distill 20B-parameter models on a single GPU.
• Balance inference speed with image quality.
• Enhance Real-ISR performance on datasets.

Topics

• Diffusion Models
• Image Super-Resolution
• Model Distillation
• Training Efficiency
• Any-Step Sampling
• Generative AI

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• TOC-SR: Task-Optimal Compact diffusion for Image Super Resolution — TOC-SR creates efficient one-step super-resolution models via compact diffusion backbone discovery and distillation.
• High-Fidelity Two-Step Image Generation via Teacher-Aligned End-to-End Distillation — Tailored distillation strategies can significantly improve the quality-efficiency trade-off for few-step image generation.
• TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution — TUDSR uses a two-stage, chunk-based diffusion framework to achieve high-quality super-resolution at resolutions up to 2048^2.
• Stage-wise Distortion-Perception Traversal in Zero-shot Inverse Problems with Diffusion Models — A stage-wise diffusion model framework enables flexible distortion-perception tradeoff traversal in zero-shot inverse problems.
• Adapting Diffusion Language Models for Lossless Pixel-Level Image Transmission — DDM-SSCC adapts diffusion language models for lossless pixel-level image transmission, achieving superior exact recovery via novel coding and denoising strategies.
• Image Restoration via Diffusion Models with Dynamic Resolution — Dynamic resolution DMs accelerate image restoration by operating in lower-dimensional subspaces, avoiding VAE overhead.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.