ReGO: Reference-Guided Outpainting for Scenery Image

Area of Science:

• Computer Vision
• Image Generation
• Artificial Intelligence

Background:

• Existing image outpainting methods struggle with blurry textures and generative artifacts, limiting visual appeal.
• Synthesizing semantically coherent content remains a challenge in image completion tasks.

Purpose of the Study:

• To introduce ReGO (Reference-Guided Outpainting), a novel method to improve texture richness and authenticity in sketch-guided image outpainting.
• To address limitations of current outpainting techniques by leveraging neighboring reference images for pixel transfer.

Main Methods:

• ReGO utilizes neighboring reference images for texture synthesis through pixel transfer.
• An Adaptive Content Selection (ACS) module facilitates texture compensation in the target image.
• A style ranking loss is employed to maintain stylistic consistency and prevent reference image influence.

Main Results:

• ReGO significantly enhances texture richness and authenticity in outpainting results.
• Experimental results on NS6K, NS8K, and SUN Attribute benchmarks demonstrate superior performance compared to prior art.
• Integration with state-of-the-art outpainting models validates ReGO's effectiveness.

Conclusions:

• ReGO offers a model-agnostic learning paradigm for superior image outpainting.
• The method effectively overcomes common artifacts like blurriness, producing visually appealing and authentic results.
• ReGO represents a significant advancement in sketch-guided image outpainting technology.