A new framework called AnchorEdit aims to stop AI image editors from slowly warping their subjects across repeated editing sessions.
Researchers published AnchorEdit, describing it as the first autoregressive diffusion-based framework built specifically for long-horizon, high-resolution multi-turn image editing. The problem it targets is familiar to anyone who has tried to iteratively refine an AI-generated image: each edit compounds small errors until the subject barely resembles what you started with. Earlier approaches borrowed video-generation techniques that rely on bidirectional attention, which processes everything at once - a design that conflicts with the one-step-at-a-time reality of interactive editing. AnchorEdit uses causal attention instead, the same directional approach that underpins large language models, paired with a memory mechanism that keeps the original subject locked across editing rounds.
The paper reports the system maintains subject fidelity across more than 10 consecutive editing rounds, which would make iterative design workflows substantially more practical for diffusion-based tools. The researchers also released a benchmark designed to stress-test multi-turn editing stability, filling an evaluation gap that has made fair comparisons between competing approaches difficult.
Whether any of this survives contact with a shipping product is a separate question. "State-of-the-art on our own benchmark" has a poor track record of translating into tools real users rely on.