ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

Paper Overview

• Field: Computer Vision
• Authors: Sara Dorfman, Nir Goren, Gal Barak, Daniel Cohen-Or, Or Patashnik
• Date: 2026-10-15
• arXiv: 2610.22328

Abstract

Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor's original training and multi-turn inference. We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift. A key is to force the editor to operate on its own outputs from previous editing turns during training, bringing the principle of Self-Forcing to image editing.

The main challenge is that self-generated editing trajectories deviate from ground-truth trajectories, making direct supervision unreliable. We address this challenge through a distillation objective: a teacher image editor provides reliable training signals while the model operates on its own previous outputs. We introduce an asymmetric conditioning scheme, where the student performs edits sequentially on its own outputs while the teacher remains anchored to the original clean image and receives the composite editing instruction. Experiments demonstrate that our approach significantly reduces degradation over long multi-turn sequences, while maintaining edit quality and instruction following.

Key Ideas

• Multi-turn editing suffers from accumulating distribution shift: each edit introduces small shifts that compound across turns.
• Self-Forcing for editing: during training, the editor is made to work on its own outputs from previous turns, matching inference conditions.
• Distillation objective solves the unreliable-supervision problem caused by self-generated trajectories deviating from ground truth.
• Asymmetric conditioning: the student edits its own sequential outputs; the teacher stays anchored to the original clean image and receives the composite instruction.
• Results show markedly less degradation over long multi-turn sequences while preserving edit quality and instruction adherence.