Bidirectional Reward-Guided Diffusion for Real-World Image Super-Resolution

Zihao Fan, Xin Lu, Yidi Liu, Jie Huang, Dong Li, Xueyang Fu †, Zheng-Jun Zha · 2026

Read Bidirectional Reward-Guided Diffusion for Real-World Image Super-Resolution by Zihao Fan, Xin Lu, Yidi Liu, Jie Huang, Dong Li, Xueyang Fu †, Zheng-Jun Zha (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07069.txt.

Description

Zihao Fan and six co-authors sharpen diffusion-based super-resolution for real photos. Diffusion models trained on synthetic low- and high-resolution pairs excel at inventing fine detail, yet they often falter on real low-resolution photos, where the data distribution shifts. Their system, Bird-SR, frames super-resolution as trajectory-level preference tuning through reward feedback learning, using synthetic pairs and real low-resolution photos together.

Early diffusion steps tune structure directly on synthetic pairs. Later steps apply quality-guided rewards to both synthetic and real photos for perceptual gain. To curb reward hacking, synthetic rewards sit in a relative advantage space bounded by ground truth, while real-photo tuning is reined in by a semantic alignment rule. A dynamic weighting scheme leans to structure early and to perception later. On real-world benchmarks, Bird-SR topped top methods on perceptual quality while holding structural consistency.

Catalog record

CREATORZihao Fan, Xin Lu, Yidi Liu, Jie Huang, Dong Li, Xueyang Fu †, Zheng-Jun Zha
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07069

Browse the full library · About the Tower of Babel