CCA: Collaborative Competitive Agents for Image Editing

2024-01-23

arXiv.org

AI Summary

This work addresses the challenging problem of complex controllable image editing under unpaired data conditions. We propose the Collaborative-Competitive Multi-Agent (CCMA) framework, which jointly integrates collaborative and adversarial mechanisms. Collaborative agents perform fine-grained editing via latent-space co-optimization and diffusion-based policy networks; competitive agents enforce a balance between diversity and consistency through adversarial reward modeling and multi-agent reinforcement learning, enabling iterative refinement over multiple rounds. Crucially, CCMA is entirely training-data-free—requiring neither paired image-text annotations nor any supervised training data. Evaluated on multiple image editing benchmarks, it achieves state-of-the-art performance. User studies further demonstrate its superior naturalness and intent fidelity compared to existing methods, empirically validating the effectiveness and practicality of the proposed dual-mechanism design.

Technology Category

• Multiagent Systems: Adversarial Agents
• Computer Vision: Diffusion Models for Vision
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Problem

• Develops a multi-agent generative model for image editing
• Enhances image editing through collaborative competitive agents
• Provides controllable intermediate steps in generation process

Innovation

• Collaborative Competitive Agents enhance editing
• Multiple LLM agents execute complex tasks
• Transparent intermediate steps improve robustness