NTH
Diffusion Models

MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion

AuthorsHouyuan Chen, Fanzhou Wang, Xiangyu Fan, Zesong Qiu, Yukun Wei, Songlin Yang, Zhitao Yang, Lvmin Zhang, Pei Yang, Hao Zhao, Lei Yang, Anyi Rao

AffiliationsSenseTime Research · HKUST MMLab · Stanford University · AIR, Tsinghua University

October 28, 2026 2 min read
Watch on YouTube

Resources

View code and benchmarks on Papers with Code Read the original paper on arXiv
The one-line take

MT Forcing trains autoregressive diffusion models to make image edits that stay visually stable and consistent across multiple turns.

Key results

100
MTEdit-Bench sessions

Sessions in the evaluated step-wise benchmark subset.

4.65
Turn-eight instruction-following gain

Point improvement on SenseNova-U1.5.

4.46
Turn-eight content-consistency gain

Point improvement on SenseNova-U1.5.

4.30
Eight-turn drift robustness score

Full method’s mean ablation score on SenseNova-U1.5.

88.4%
Pixel displacement reduction

Eight-turn mean Pixel falls from 6.34 to 0.74.

What the paper found

As image-editing products such as OpenAI’s ChatGPT Images 2.5 raise expectations for interactive editing, a key challenge is keeping images coherent over repeated instructions. MT Forcing addresses this drift in autoregressive diffusion models with two complementary techniques: Single-Turn Distribution Matching Distillation, which trains on the model’s own rollouts while matching each turn to a teacher applying all edits so far directly to the source image, and Teacher-Forcing Guidance, which preserves history-dependent details using consistent reference trajectories. Tested on SenseNova-U1 and SenseNova-U1.5 with the 100-session, eight-turn MTEdit-Bench subset, the method improves later-turn instruction following and content consistency; Qwen3.5-Plus scores these measures. On SenseNova-U1.5 at turn eight, instruction following improves by 4.65 points and content consistency by 4.46 points. In the eight-turn ablation results, the full method reaches a drift robustness score of 4.30, while Pixel displacement falls from 6.34 to 0.74, an 88.4% reduction. The results show why the two training signals are complementary: rollout distillation limits quality degradation, while reference supervision helps retain prior edits.

Original 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.

Read the original paper

More in Diffusion Models

Browse all 73 papers →