MT Forcing: Stabilizing Multi-Turn Image Editing with Single-Turn DMD and Teacher Forcing

Problem: Exposure Bias in Multi-Turn Image Editing

Multi-turn image editing allows users to iteratively modify an image through sequential instructions. However, existing models degrade after a few turns: textures become oversaturated, local details distort, and unmodified regions change unexpectedly. The root cause is exposure bias : during training, models learn from clean reference histories, but at inference they must condition on their own previously generated outputs, which already contain artifacts and deviations. This mismatch causes error accumulation across turns.

Method: MT Forcing

MT Forcing addresses both quality stability and cross-turn consistency through two complementary components:

1. Single-Turn DMD (Distribution Matching Distillation)

Observation: A single editing step typically incurs limited quality loss, and an editing model can execute multiple operations in one instruction. The teacher model takes the original image and a merged instruction representing all edits up to the current turn, producing a high-quality result in one pass. The student model still performs sequential self-rollout, generating each turn conditioned on its own previous output. Single-Turn DMD matches the student's multi-turn output distribution to the teacher's single-turn distribution via a diffusion distillation loss, without requiring pixel-level replication. This teaches the student to handle its own imperfect history while maintaining output quality.

2. Teacher-Forcing Guidance

Single-Turn DMD alone cannot guarantee that fine-grained visual details (e.g., specific texture patterns on a teapot) remain consistent across turns, because the merged instruction does not specify the exact appearance established in earlier turns. Teacher-Forcing Guidance uses a coherent reference trajectory generated by teacher-forcing (feeding ground-truth previous outputs) to supervise the relationship between history and the next output. It consists of two phases: (1) a warm-up phase where the model learns cross-turn content preservation via flow matching on the reference trajectory, and (2) a joint training phase where a low-weight reference supervision term is retained alongside the DMD loss to prevent forgetting this capability. The combined objective is L = L_DMD + λ_TF * L_TF.

3. MT Forcing

MT Forcing integrates both objectives, training a single student model with both distribution matching on self-generated histories and consistency regularization on a reference trajectory.

Experiments

Evaluated on the MTEdit-Bench stepwise editing subset: 100 sessions, 8 turns each . Two base models: SenseNova-U1 and SenseNova-U1.5. Metrics: Instruction Following (IF, max 10), Content Consistency (CC, max 10), Quality Drift Resistance (DRS, max 5), and geometric alignment (mean keypoint displacement in pixels).

Quantitative Results

SenseNova-U1.5 : 8th-turn IF improved from 4.52 to 9.17; CC from 3.22 to 7.68. DRS baseline dropped from 4.09 (turn 1) to 3.03 (turn 8); MT Forcing maintained 4.20–4.35 across all 8 turns. Geometric displacement mean reduced from 6.34 pixels to 0.74 pixels ( 88.4% reduction ).

SenseNova-U1 : 8th-turn IF and CC improved to 7.45 and 6.82 respectively.

Gains emerge from turn 4 onward; first-turn performance slightly lower than baseline.

Qualitative Results

Visual comparison at turns 1, 4, and 8 shows that baseline models develop over-saturated lighting, coarse textures, speckles, and excessive contrast. MT Forcing preserves cleaner textures, coherent scene structure, and previously established edits (e.g., red car body, roof rack, tents, ski boards, signposts).

Resources

Paper: https://arxiv.org/abs/2610.25404

Project page: https://houyuanchen111.github.io/mt-forcing.github.io/

Code: https://github.com/houyuanchen111/MT-Forcing

Model weights: https://huggingface.co/houyuanchen/MT-Forcing