Edit-R2: A Context-Aware Reinforcement Learning Framework for Multi-Round Image Editing

Edit-R2 is a context-aware reinforcement learning post-training framework for multi-round image editing. It effectively addresses the problems of long-context dilution and state contamination by reconstructing conversational intent and unifying optimization objectives, and it also releases the MICE-Bench evaluation benchmark as a companion. This framework aims to improve the accuracy and stability of multi-round image editing, and promote the evolution of technology towards collaborative interactions that are closer to users' actual needs.

Text-guided image editing technology has made significant progress in recent years, but most methods are limited to single-round scenarios. In practical use, image editing is often an iterative process (e.g., changing the background first then adding elements), which is more practically valuable. However, continuous editing faces two coupled failure modes: long-context dilution and state contamination, which are the core pain points of current models.

1. Long-context dilution: As the number of rounds increases, historical image-text information accumulates, and sparse text constraints are easily overwhelmed (e.g., the early constraint of "keeping facial features" is lost); 2. State contamination: Early editing errors persist or even amplify, affecting subsequent results like a snowball. The two are intertwined, increasing the difficulty of multi-round editing.

Edit-R2 addresses the challenges from three aspects: 1. Conversational intent reconstruction: Before each round, historical constraints are refined into a clear reasoning trajectory to alleviate long-context dilution; 2. Unified optimization objective: Simultaneously optimize discrete text intent generation and continuous latent space image generation to form an end-to-end closed loop; 3. Trajectory filtering mechanism: Suppress editing sequences containing errors during training to solve the state contamination problem.

The research team launched the MICE-Bench evaluation benchmark, which includes three automated metrics: 1. Instruction Following (IF): Evaluate the accuracy of each round's instruction execution; 2. Content Consistency (CC): Check the consistency between the image and historical content; 3. Global Awareness (GA): Measure the ability to grasp the cumulative conversational constraints as a whole, providing a standardized evaluation method for domain research.

Experiments show that Edit-R2 significantly improves multi-round editing capabilities and outperforms baseline methods. Key insight: Explicitly managing session-level constraints is more effective than stacking historical information. In application, it allows users to edit iteratively step by step, transforming the tool from a "passive tool" to an "active collaborator", which is closer to real work processes.

The explicit intent reconstruction idea of Edit-R2 can be extended to sequence decision-making scenarios such as multi-round dialogue and progressive code generation. In the future, how to use multimodal models to handle complex long-term interaction tasks will be an important direction, and Edit-R2 and MICE-Bench have laid a technical foundation for this field.