Zero-shot cross-domain image composition via self attention injection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion models have advanced image composition.
• Existing methods struggle with balancing object structure and background style.
• Limitations include poor structure-style disentanglement and reliance on global alignment.

Purpose of the Study:

• To propose a novel framework for image composition that overcomes the structure-style trade-off.
• To enhance both structural fidelity and stylistic harmonization in generated images.
• To introduce a training-free approach for improved image composition.

Main Methods:

• Developed a tri-branch denoising framework.
• Implemented attention manipulation to decouple structure from style.
• Introduced Semantic Injection using self-attention maps.
• Adapted attention-based style transfer for composition via Style Guidance.

Main Results:

• The proposed method outperforms existing state-of-the-art approaches.
• Achieved consistent improvements in structural consistency.
• Demonstrated enhanced stylistic coherence between objects and backgrounds.
• The framework is training-free, simplifying the composition process.

Conclusions:

• The novel framework effectively decouples structure and style in image composition.
• Attention manipulation is key to resolving the structure-style dilemma.
• The method offers superior performance in both structural and stylistic aspects of image generation.