Reproduce “Flow Straight to Reality: Perceptually Consistent Flow Matching for Efficient Image Restoration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sangwoo Jo, Donggeun Ko, Jayeon Kang, Youngsang Kwak, Jaehwa Kwak, Sungjoon Choi
PCFlow introduces a unified latent transport framework for image restoration, leveraging latent consistency flow matching and a novel conflict-free gradient alignment to balance distortion and perception. It achieves competitive perceptual quality and structural fidelity across various restoration tasks with significantly fewer parameters and inference steps than prior generative methods.
Key insights
- Direct continuous transport in a latent space, when guided by consistency flow objectives, enables efficient few-step image restoration while maintaining structural fidelity.
- Perceptual guidance can be effectively integrated into continuous flow models using an internal perceptual loss derived from the model's own features, which aligns better with latent dynamics than external networks.
- Asymmetric gradient projection and dynamic weighting of perceptual objectives are crucial for stably reconciling conflicting structural and perceptual goals, particularly during early, noisy stages of the restoration process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.