Mitigating inappropriate concepts in text-to-image generation with attention-guided Image editing

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Text-to-image generative models are advancing rapidly, creating diverse visuals from text prompts.
• Concerns exist regarding the generation of inappropriate, offensive, or explicit content by these models.

Purpose of the Study:

• To develop a method for selectively suppressing inappropriate concepts during text-to-image generation.
• To address safety concerns without compromising image integrity or computational efficiency.

Main Methods:

• Leveraging attention maps to identify and suppress inappropriate concepts.
• Evaluating the method through quantitative assessments and human perceptual studies.
• Focusing on a simple, effective approach without additional model training.

Main Results:

• Demonstrated effective reduction of inappropriate content.
• Preserved the integrity and context of the original images.
• Achieved high computational efficiency compared to existing methods.

Conclusions:

• The proposed attention map-based method offers an effective and efficient solution for mitigating inappropriate content in generative models.
• The technique maintains image quality and context, requiring no additional training or significant engineering effort.