AccDiffusion v2: Toward More Accurate Higher-Resolution Diffusion Extrapolation.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion models often produce repetitive objects and local distortions when generating images at resolutions different from their training data.
• Existing methods struggle with accurate high-resolution image extrapolation without extensive retraining.

Purpose of the Study:

• To introduce AccDiffusion v2, a novel method for accurate patch-wise, higher-resolution diffusion extrapolation without requiring additional training.
• To address object repetition and local distortion issues inherent in high-resolution image generation from diffusion models.

Main Methods:

• Decoupling image-content prompts into patch-content-aware prompts for precise local descriptions.
• Integrating auxiliary local structural information via ControlNet to mitigate distortions.
• Employing dilated sampling with window interaction to enhance global semantic understanding.

Main Results:

• AccDiffusion v2 achieves state-of-the-art performance in training-free image generation extrapolation.
• Quantitative and qualitative experiments demonstrate significant suppression of repetitive generation and local distortion.
• The method effectively improves image quality during resolution extrapolation.

Conclusions:

• AccDiffusion v2 offers an effective solution for high-resolution image generation extrapolation from diffusion models.
• The proposed techniques of patch-specific prompting, ControlNet integration, and global semantic enhancement are crucial for accurate extrapolation.
• This work advances the capabilities of diffusion models for diverse resolution image synthesis.