Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment
Summary: arXiv:2505.18600v3 Announce Type: replace-cross
Abstract: Modern single-image super-resolution (SISR) models deliver photo-realistic results at the scale factors on which they are trained, but collapse when asked to magnify far beyond that regime. We address this scalability bottleneck with Chain-of-Zoom (CoZ), a model-agnostic framework that factorizes SISR into an autoregressive chain of intermediate scale-states with multi-scale-aware prompts. CoZ repeatedly re-uses a backbone SR model, decomposing the conditional probability into tractable sub-problems to achieve extreme resolutions without additional training. Because visual cues diminish at high magnifications, we augment each zoom step with multi-scale-aware text prompts generated by a vision-language model (VLM). The prompt extractor itself is fine-tuned using Generalized Reward Policy Optimization (GRPO) with a critic VLM, aligning text guidance towards human preference. Experiments show that a standard 4x diffusion SR model wrapped in CoZ attains beyond 256x enlargement with high perceptual quality and fidelity.
Introduction
Single-image super-resolution (SISR) has gone through significant advancements in recent years, enabling the creation of high-quality images from low-resolution inputs. However, these models often falter when tasked with magnifying images beyond their trained scale factors. The newly introduced Chain-of-Zoom (CoZ) framework aims to overcome this limitation by leveraging an autoregressive approach that allows for extreme resolutions without the need for additional training.
Key Features of Chain-of-Zoom
The Chain-of-Zoom framework incorporates several innovative features:
- Model-Agnostic Framework: CoZ is designed to work with any existing SISR model, making it versatile and accessible for various applications.
- Autoregressive Chain of Intermediate Scale-States: The framework breaks down the super-resolution task into manageable segments, allowing for gradual enhancement of image quality.
- Multi-Scale-Aware Prompts: To counteract the loss of visual details at higher magnifications, CoZ uses prompts generated by a vision-language model, enhancing the contextual understanding of the image.
- Fine-Tuning with Generalized Reward Policy Optimization: This process aligns the text guidance with human preferences, ensuring that the final output meets aesthetic and quality standards.
Experimental Results
In various experimental setups, a standard 4x diffusion super-resolution model integrated within the Chain-of-Zoom framework has successfully achieved enlargements exceeding 256x. The results demonstrate not only the capability to maintain high perceptual quality but also to preserve fidelity in the final images. This breakthrough indicates a significant advancement in the field of computer vision and image processing.
Conclusion
The Chain-of-Zoom framework represents a significant leap forward in the domain of single-image super-resolution. By addressing the existing limitations in scalability and quality through innovative methodologies, CoZ opens new avenues for applications in photography, digital media, and beyond. Researchers and practitioners interested in exploring the capabilities of this new framework can visit the project page for more information.