Stable and consistent multi-turn image editing with autoregressive diffusion
MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion
Computer Vision and Pattern Recognition
Summary
Editing an image multiple times in a row often causes it to lose quality or change in unintended ways. The authors developed a new training method called MT Forcing to help AI models edit images step-by-step without making mistakes that pile up. Their approach trains the model to keep the image stable and consistent across many editing turns. Tests show it improves the quality and accuracy of multi-step image edits.
What this means in practice
- •For graphic designers: Enable interactive image editing tools that maintain high quality and accuracy across multiple edit steps.$Commercial implications: Supports development of advanced image editing software that offers reliable stepwise modification capabilities for professional designers.
- •For content creators: Improve iterative editing features for generating visual content that stays consistent with user instructions over repeated updates.
Authors
Houyuan Chen, Fanzhou Wang, Xiangyu Fan, Zesong Qiu, Yukun Wei, Songlin Yang, Zhitao Yang, Lvmin Zhang, Pei Yang, Hao Zhao, Lei Yang, Anyi Rao
Abstract
Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects: 1) unstable visual quality, with accumulating artifacts and degraded details; and 2) inconsistent image content, with unintended changes to preserved regions and prior edits. To address these challenges, we propose MT Forcing, a novel training paradigm for multi-turn image editing with unified multimodal models (UMMs) based on autoregressive diffusion. Specifically, we propose Single-Turn Distribution Matching Distillation (Single-Turn DMD), which supervises student self-rollouts with a teacher that applies cumulative edits to the source image in a single turn, providing visual-quality guidance without recursive error accumulation. Additionally, we introduce Teacher-Forcing Guidance, which supervises history-dependent transitions on reference trajectories during warmup and as a regularizer during distillation, encouraging cross-turn content consistency beyond what per-turn distribution matching alone can ensure. Extensive experiments demonstrate that our approach significantly improves the multi-turn image editing capabilities of UMMs, yielding more stable visual quality and stronger cross-turn consistency.