ByteDance's Edit-R2 Stops AI Image Editors From Forgetting Your Instructions
A new RL post-training framework teaches multimodal models to remember the full history of an editing session, not just the last instruction
Takeaways
- Edit-R2 is a research paper proposing an RL post-training framework for multi-turn image editing on unified multimodal models. [paper]
- It addresses two failure modes: long-context dilution (forgetting early instructions) and state contamination (compounding errors across turns).
- The core mechanism is operative session intent reconstruction — generating an explicit reasoning trace summarizing all session constraints before each edit.
- Training jointly optimizes reasoning in discrete text space and image generation in continuous latent space via flow matching, with a trajectory filter to suppress corrupted rollouts.
- MICE-Bench is introduced as a new large-scale benchmark measuring Instruction Following, Content Consistency, and Global Awareness across full editing sessions.
- No code or model weights are publicly released; this is an academic research contribution with no product release attached.
Ask any image-editing model to add a tree, then change its color, then remove the background , and by the third turn, it has usually forgotten the first two. This is the core failure mode that a team of researchers, including authors affiliated with ByteDance, set out to fix in their paper Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing. The work introduces a reinforcement learning post-training framework designed specifically for the multi-turn editing problem, and it comes with a new benchmark to measure it properly.
The gap nobody was measuring
Instruction-based image editing has captured significant attention as models become capable of handling real-world editing demands, but models trained primarily for single-turn edits often break down in multi-turn editing , the natural interactive setting where a user iteratively refines an image based on the model's own previous outputs. This is not a niche edge case. It is how people actually use these tools: "make the sky orange," then "add a bird," then "make it look like dusk."
For multi-turn editing tasks, all models exhibit a notable decline in background consistency metrics during the second editing round. The degradation compounds with every additional turn, and until now there was no standardized way to measure it systematically across sessions.
Two failure modes, one framework
The paper names two coupled problems that cause multi-turn editing to collapse. The first is long-context dilution: as the conversation history grows, the model's attention gets spread thin across an interleaved sequence of images and text, and earlier constraints , "keep the red ribbon" , become effectively invisible. The second is