AnchorEdit: Maintaining Temporal Consistency in Multi-turn Image Editing via Causal Memory

2026-06-10

AI Summary

This work addresses the challenges of identity drift and error accumulation in multi-turn image editing, which are exacerbated by existing methods’ reliance on non-causal, bidirectional attention mechanisms that fail to ensure temporal consistency in interactive settings. The paper proposes the first autoregressive diffusion framework tailored for high-resolution, long-horizon multi-turn editing, introducing a novel causal memory mechanism and a self-unrolling strategy to effectively mitigate exposure bias and preserve cross-turn identity consistency. The method employs a three-stage training pipeline—comprising single-turn pretraining, causal fine-tuning, and consistency distillation—and demonstrates significant performance gains over current approaches on a newly introduced high-resolution multi-turn editing benchmark, maintaining high-fidelity identity preservation and precise instruction following even after more than ten editing turns.

Abstract

Multi-turn image editing is essential for iterative design, yet current models often struggle with identity drift and error accumulation over successive steps. While existing research leverages video priors for consistency, their reliance on bidirectional attention is fundamentally misaligned with the causal, sequential nature of interactive editing. In this paper, we propose AnchorEdit, the first autoregressive (AR) diffusion-based framework designed specifically for high-resolution, long-term multi-turn editing. AnchorEdit bridges the gap between video priors and causal inference through a three-stage training curriculum: identity-preserving sing-turn pretraining, causal AR forcing fine-tuning with a novel self-rollout strategy to mitigate exposure bias, and consistency distillation for efficient 4-step generation. During inference, we introduce a memory mechanism to anchor the initial subject identity and ensure stable extrapolation across extended editing trajectories. To evaluate performance, we provide a new high-resolution multi-turn editing benchmark designed to stress-test long-horizon stability. Extensive experiments demonstrate that AnchorEdit achieves state-of-the-art results, maintaining exceptional subject fidelity and instruction following even over 10+ interaction rounds.

Problem

• multi-turn image editing
• temporal consistency
• identity drift
• error accumulation
• causal inference

Innovation

• autoregressive diffusion
• temporal consistency
• causal memory
• multi-turn image editing
• identity preservation