Deep learning image generation technology for enhancing the presentation effect of image art based on artificial...

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Art

Background:

• Current AI Image Generation Technology (IGT) shows potential but struggles with precise style control and high-resolution output.
• Existing methods face challenges in maintaining artistic texture during Super-Resolution (SR) processing.
• Limitations include inaccurate style control, limited resolution, and texture loss in AI art generation.

Purpose of the Study:

• To propose an innovative framework, StyleDiffusion-HD, addressing shortcomings in AI IGT for artistic creation.
• To achieve precise bimodal control of text and visual style in AI-generated art.
• To enhance image resolution while preserving style consistency and artistic texture.

Main Methods:

• Integration of a Latent Diffusion Model (LDM) with Style Injection Attention (SIA) for bimodal style control.
• Introduction of a Super-Resolution (SR) module based on Flow Matching (FM) for resolution enhancement.
• Utilizing multi-source high-quality artistic datasets for comprehensive evaluation.

Main Results:

• StyleDiffusion-HD outperforms mainstream models on objective metrics like Fréchet Inception Distance (FID), CLIP Score (CS), and Style Loss (SL).
• The framework achieves high scores in subjective evaluations for generation quality, style consistency, and aesthetics.
• Demonstrated effectiveness in improving artistic presentation and maintaining style consistency during SR.

Conclusions:

• StyleDiffusion-HD offers a feasible technical solution for key challenges in current AI art generation.
• The study provides practical references for developing high-quality AI-assisted artistic creation.
• The proposed framework enhances precise style control and high-fidelity image output in AI art.