Reproduce “Multi-History-Step SDE Inversion for Image Editing with Superior Regional Awareness” 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
Haiyan Wei, Yunlong Wang, Huaibo Huang, Zhenan Sun, Kunbo Zhang
MIEdit introduces a training-free image editing framework that leverages multi-history-step SDE inversion with a power-interpolated beta schedule for enhanced stability and editing plasticity. The system incorporates an Inversion-Time Automatic Semantic Angle Masking (IASM) mechanism, which automatically generates precise semantic masks during the inversion phase, leading to superior regional awareness and instruction fidelity for text-guided image manipulation with improved computational efficiency.
Key insights
- Leveraging a multi-history-step predictor–corrector mechanism within SDE inversion significantly improves both reconstruction accuracy and computational efficiency for image editing.
- Stabilizing the latent trajectory's variance through a power-interpolated beta schedule during inference is crucial for enhancing editing plasticity and preventing artifact generation, particularly for large-scale image modifications.
- Classifier-Free Guidance, when repurposed during inversion for semantic angle mask computation, can automatically provide consistent and fine-grained regional control for edited and unedited areas without explicit user input.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.