Reproduce “LGCC: Enhancing Flow Matching Based Text-Guided Image Editing with Local Gaussian Coupling and Context Consistency” 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

Fangbing Liu, Pengfei Duan, Wen Li, Yi He

LGCC enhances flow matching-based text-guided image editing by introducing local Gaussian coupling and a content consistency loss, leading to improved detail preservation and reduced over-editing. The framework achieves a 1.60% improvement in local detail scores and a 2x-5x speedup in inference time compared to prior state-of-the-art models.

Key insights

  • Initializing flow matching with structured local perturbations, rather than pure noise, shortens latent space trajectories, enabling faster inference and superior detail preservation.
  • Explicitly enforcing semantic alignment between text instructions and image modifications through a consistency loss effectively mitigates context loss and prevents over-editing.
  • Enhancements can be effectively integrated into powerful pre-trained generative models via fine-tuning with specific noise coupling and loss functions, optimizing their practical utility without extensive retraining.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.