TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

2026-06-08 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

TUDSR, a Twice Upsampling-Diffusion framework, addresses the challenge of generating high-quality images at resolutions like 2048^2 using diffusion-based super-resolution models. Existing methods often produce poor results when upsampling ratios, such as ×8, exceed a model's native support, like ×4, or when the target resolution surpasses the model's native capabilities. Training models natively for high resolutions incurs significant computational and GPU memory costs. TUDSR mitigates this by employing a two-stage process: initial training at R-resolution, followed by a looped chunk-based training strategy at NR-resolution. Both stages utilize a one-step GAN architecture. TUDSR-S, built upon SD2.1-base, demonstrates leading performance across multiple benchmarks, successfully generating high-quality images at 1024^2 and 2048^2, outperforming current approaches.

Key takeaway

For Machine Learning Engineers developing high-resolution image generation systems, you should consider TUDSR's two-stage diffusion framework. This approach allows you to achieve high-quality super-resolution up to 2048^2 without the prohibitive computational costs of training native high-resolution models. Implement the looped chunk-based training strategy to overcome limitations when upsampling ratios exceed your model's native capabilities, enhancing output quality for demanding applications.

Key insights

TUDSR uses a two-stage, chunk-based diffusion framework to achieve high-quality super-resolution at resolutions up to 2048^2.

Principles

• Upsampling beyond native ratios degrades quality.
• High-resolution training is resource-intensive.
• Staged training can overcome resolution limits.

Method

TUDSR trains in two stages: first at R-resolution, then with a looped chunk-based strategy at NR-resolution. Each stage uses a one-step GAN.

In practice

• Generate 2048^2 images from lower resolution inputs.
• Utilize SD2.1-base for high-quality SR.
• Apply two-stage training to overcome GPU limits.

Topics

• Image Super-Resolution
• Diffusion Models
• Generative Adversarial Networks
• High-Resolution Imaging
• TUDSR Framework
• SD2.1-base

Code references

• wuer5/TUDSR
• shangwei5/ST-AVSR
• shangwei5/BasicAVSR

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

Related on AIssential

• The Fourth Challenge on Image Super-Resolution ($\times$4) at NTIRE 2026: Benchmark Results and Method Overview — The NTIRE 2026 SR challenge benchmarks $\times$4 image super-resolution across fidelity and perceptual quality.
• Image Restoration via Diffusion Models with Dynamic Resolution — Dynamic resolution DMs accelerate image restoration by operating in lower-dimensional subspaces, avoiding VAE overhead.
• A Scale-Adaptive Framework for Joint Spatiotemporal Super-Resolution with Diffusion Models — A scale-adaptive framework reuses one architecture for spatiotemporal super-resolution across diverse scaling factors.
• 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.
• Any Resolution Any Geometry: From Multi-View To Multi-Patch — URGT refines high-resolution depth and normal maps using a multi-patch transformer with global coherence.
• Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion — Real2SAM2Real uses editable 3D caches from lifting models to provide robust geometric scaffolds for Video Diffusion Models, improving control and consistency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.