Problem
Research questions and friction points this paper is trying to address.
image restoration
lossy compression
fidelity
human perception
machine preference
FDIR: Harmonizing Fidelity and Human-Machine Preference in Lossy Compression Image Restoration
🤖 AI Summary
Existing lossy compressed image restoration methods struggle to simultaneously achieve high pixel-level fidelity, perceptual quality for human observers, and performance on downstream machine vision tasks, often resulting in trade-offs among these objectives. This work proposes FDIR, a two-stage framework that first recovers global semantic structure in latent space via quality-guided one-step flow matching (QO-Flow), followed by deterministic high-frequency texture reconstruction in pixel space using a flow-conditioned detail refinement module (FCDR) to effectively suppress hallucinations. FDIR is the first unified approach to jointly optimize all three goals through a decoupled design that balances their inherent conflicts, thereby avoiding excessive smoothing and distortion. Experiments demonstrate that FDIR achieves state-of-the-art or competitive performance across fidelity, perceptual quality, and machine task metrics.
📝 Abstract
Image restoration quality can be evaluated along three complementary facets: pixel-level fidelity, human perception, and downstream machine preference. However, existing lossy compression restoration methods optimize for at most one of these criteria: fidelity-oriented models often regress toward conditional means and produce over-smoothed outputs, while generative approaches hallucinate plausible but factually incorrect textures that degrade both ground-truth fidelity and downstream task accuracy. To navigate this three-way tradeoff, we propose FDIR, a two-stage architecture that decouples the conflicting demands through complementary inductive biases: Quality-Guided One-Step Flow Matching (QO-Flow) recovers global semantic structure in latent space via a single forward pass, while Flow-Conditioned Detail Refinement (FCDR) deterministically restores high-frequency textures and suppresses generative hallucinations in pixel space. Extensive experiments demonstrate that FDIR achieves superior fidelity, with a favorable perceptual-fidelity balance and competitive machine preference.
Innovation
Methods, ideas, or system contributions that make the work stand out.
FDIR
flow matching
image restoration
fidelity-perception tradeoff
detail refinement
🔎 Similar Papers
No similar papers found.
