Reproduce “Edge-Aware Image Manipulation via Diffusion Models with a Novel Structure-Preservation Loss” 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
Minsu Gong, Nuri Ryu, Jungseul Ok, Sunghyun Cho
Researchers from POSTECH and Planby Technologies introduced a training-free framework for edge-aware image manipulation using latent diffusion models, integrating a novel bidirectional Structure Preservation Loss (SPL) into the denoising process. This approach enables text-prompt-driven edits while faithfully preserving pixel-level edge structures, achieving superior structural fidelity compared to existing LDM-based methods without sacrificing prompt adherence.
Key insights
- The bidirectional formulation of the Structure Preservation Loss (SPL) effectively disentangles structural differences from appearance changes, providing a robust and differentiable guide for generative models.
- Iterative guidance with SPL directly within the LDM's denoising pipeline is crucial for preventing structural errors from forming, as merely applying it during post-processing is insufficient to correct severe distortions.
- Combining different color spaces—HSI for SPL (intensity) and YCbCr for optional Color Preservation Loss (chroma)—empirically yields high-quality results by separating luminance-based structure guidance from hue preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.