New FDIR architecture balances image restoration fidelity and human preference

Researchers have introduced FDIR, a novel two-stage architecture designed to improve image restoration quality by balancing pixel-level fidelity, human perception, and downstream machine preference. The architecture first uses Quality-Guided One-Step Flow Matching (QO-Flow) to recover global semantic structure and then employs Flow-Conditioned Detail Refinement (FCDR) to restore high-frequency textures while minimizing generative hallucinations. Experiments indicate that FDIR surpasses existing methods in fidelity and offers a favorable perceptual-fidelity balance with competitive machine preference. AI

IMPACT This research offers a new method for image restoration that balances multiple evaluation criteria, potentially improving applications in areas requiring high fidelity and perceptual quality.

RANK_REASON The cluster contains a research paper detailing a new technical approach to image restoration. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Flow-Conditioned Detail Refinement
• Gotit.pub
• Hugging Face
• QO-Flow
• ScienceCast

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →