Most AI image editors forget the conversation halfway through.
Researchers have introduced Edit-R2, a post-training framework built on top of multimodal foundation models that tackles a specific, underappreciated problem: when a user gives an image editor a sequence of instructions across multiple turns, the model tends to lose track of earlier constraints as the session grows longer. Two failure modes drive this — the paper calls them long-context dilution, where earlier text instructions get buried under a growing pile of interleaved images and prompts, and state contamination, where a bad edit in turn three quietly poisons everything that follows. Edit-R2 addresses both by forcing the model to reconstruct a consolidated "session intent" — an explicit summary of what the user has been trying to accomplish — before each new edit. Training uses a unified objective that covers both text reasoning (discrete space) and image generation (continuous latent space), with a filtering mechanism that throws out corrupted training examples before they can destabilize the model.
Multi-turn editing is the realistic use case that single-turn benchmarks have always glossed over. A user who makes five iterative refinements to a product photo is not running five independent sessions — they expect coherence, and current models largely fail to deliver it. Edit-R2 also ships a new benchmark, MICE-Bench, designed to measure exactly this: instruction following, content consistency, and awareness of accumulated session constraints.
The results are competitive against existing baselines, though "competitive" doing a lot of work there — the field has no strong standard yet for multi-turn image editing, which is part of why the authors had to build their own benchmark. If MICE-Bench catches on, it will matter more than Edit-R2 itself.