OMR-diffusion: Optimizing multi-round enhanced training in diffusion models for improved intent understanding

doi:10.1016/j.neucom.2025.131452

OMR-diffusion: Optimizing multi-round enhanced training in diffusion models for improved intent understanding

2025-10-09

0

PRE

AI

DOI：10.1016/j.neucom.2025.131452

Original

Original request for help

Save

Abstract

En 中文

• Optimizing User Preference in Image Generation: This work introduces a Visual Co-Adaptation (VCA) framework that integrates human feedback and reinforcement learning (RL) to align generated images with personalized user preferences while maintaining consistency across multiple turns. • Improving Generation Through Multi-Round Feedback: Dialogue-based refinement leverages user feedback at each step to enhance image diversity, structural consistency, and semantic alignment, significantly advancing text-to-image generation. • Enhancing Interaction with Feedback Loops: A dynamic reward system balances diversity, consistency, and mutual information to improve interaction quality and lower the barriers to utilizing AI technology. • Exceeding State-of-the-Art Performance: The proposed model surpasses advanced systems like DALL-E 3 and Imagen in intent alignment, image consistency, and dialogue efficiency, achieving top results in LPIPS (0.15) and BLIP (0.59).

OMR-diffusion: Optimizing multi-round enhanced training in diffusion models for improved intent understanding

Abstract

Journal

Organization