Paper Overview
- Research Area: Computer Vision (CV)
- Authors: Rongyuan Wu, Lingchen Sun, Zhengqiang Zhang, et al.
- Published: 2026-04-03
- arXiv: 2604.03225
- arXiv page: https://arxiv.org/abs/2604.03225
Summary
Most recent generative image super-resolution (SR) methods rely on adapting large text-to-image (T2I) diffusion models pretrained on web-scale text-image data. While effective, this paradigm starts from a generic T2I generator, even though SR is fundamentally a low-resolution (LR) input-conditioned image restoration task. This work investigates whether an SR model trained purely on visual data can rival T2I-based approaches.
To this end, the authors propose VOSR, a Vision-Only generative framework for SR:
1. Visual semantic guidance: VOSR extracts semantically rich and spatially grounded features from the LR input using a pretrained vision encoder. 2. Revisited classifier-free guidance: The paper re-examines classifier-free guidance for training generative restoration models, showing that the standard unconditional branch is ill-suited to restoration (the abstract is truncated at this point in the source post). 3. Efficiency and quality: VOSR requires less than one-tenth of the training cost of representative T2I-based SR methods, yet achieves competitive or better perceptual quality and efficiency in both multi-step and single-step settings. 4. Fidelity: On both synthetic and real-world benchmarks, VOSR produces more faithful structures and fewer hallucinations.
Links
*Automatically collected on 2026-04-06.*