CondFoodGen: A Conditional Two-Stream Network for Controllable Food Image Generation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Food Computing

Background:

• Existing food image generation methods struggle with intra-class variability and food-specific challenges, leading to limited diversity and accuracy.
• Current approaches often lack optimizations for texture, shape, and color fidelity in food images.

Purpose of the Study:

• To propose Cond-FoodGen, a diffusion-based two-stream network for controllable food image generation.
• To enhance the realism, diversity, and accuracy of generated food images.

Main Methods:

• Utilized a two-stream network architecture with a control stream for conditional guidance and a generation stream.
• Introduced the Bidirectional Adaptive Gating (BAG) mechanism for optimizing stream interactions and feedback.
• Developed the Wavelet-Guided Hierarchical Attention (WGHA) module to improve fine-grained texture and structural realism.
• Implemented a progressive multi-stage training strategy for stable optimization.

Main Results:

• Cond-FoodGen consistently generates high-quality and diverse food images across three datasets.
• Achieved an average improvement of 11.0% over existing food image generation methods.
• Demonstrated an average improvement of 16.2% compared to leading conditional generation approaches.

Conclusions:

• Cond-FoodGen effectively addresses limitations in current food image generation techniques.
• The proposed methods (BAG and WGHA) significantly enhance image fidelity and diversity.
• The approach offers a robust solution for controllable and realistic food image synthesis.