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

Researchers introduce AnchorEdit, an autoregressive diffusion model designed for multi-turn image editing that maintains subject identity and consistency across 10+ sequential editing rounds. The framework uses a causal memory mechanism and three-stage training approach to address identity drift and error accumulation problems in iterative image manipulation tasks.

AnchorEdit represents a meaningful advancement in generative AI by tackling a fundamental challenge in interactive image editing: maintaining visual consistency across multiple sequential operations. Traditional approaches relying on bidirectional attention mechanisms are architecturally misaligned with the sequential, causal nature of user interactions in editing workflows. This research addresses that structural mismatch through an autoregressive framework that processes edits in proper temporal order.

The technical contribution extends beyond simple consistency improvements. The three-stage training curriculum—progressing from identity preservation through causal forcing to consistency distillation—reflects a thoughtful approach to mitigating exposure bias, a known problem in autoregressive models where training and inference distributions diverge. The introduction of a self-rollout strategy during fine-tuning demonstrates sophisticated training methodology.

For developers and content creation teams, this work enables more practical applications of generative AI in design workflows. Long-horizon stability across 10+ editing rounds means users can perform complex iterative refinements without quality degradation. The new high-resolution benchmark provides valuable evaluation infrastructure for the research community.

The efficiency aspect matters for deployment: achieving quality results in just four generation steps makes real-time interactive applications more feasible. However, broader industry impact depends on whether this research translates into accessible tools and whether performance holds across diverse image types and editing scenarios beyond the paper's evaluation.

• →AnchorEdit achieves stable multi-turn image editing over 10+ rounds using causal memory anchoring of initial subject identity
• →The autoregressive framework aligns training architecture with the sequential nature of interactive editing, unlike existing bidirectional attention methods
• →Three-stage curriculum training and self-rollout strategy effectively mitigate exposure bias and improve consistency across extended editing trajectories
• →New high-resolution multi-turn editing benchmark provides standardized evaluation for long-horizon image editing stability
• →Efficient 4-step generation enables practical deployment in real-time interactive design workflows