MT Forcing trains multi-turn image editing to limit drift
Original title: MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion
Single-Turn DMD and Teacher-Forcing Guidance target artifact buildup and unintended changes across edit turns.
The paper proposes MT Forcing, a training paradigm for multi-turn image editing with unified multimodal models based on autoregressive diffusion. Existing methods are described as drifting after a few turns through unstable visual quality and inconsistent content. Single-Turn Distribution Matching Distillation supervises student self-rollouts with a teacher that applies cumulative edits in one turn. Teacher-Forcing Guidance regularizes history-dependent transitions for cross-turn consistency. Experiments are reported to improve stability and consistency.
- Drift appears as accumulating artifacts and unintended changes to preserved regions.
- Single-Turn DMD avoids recursive error by matching a single-turn teacher.
- Teacher-Forcing Guidance is used in warmup and as a distillation regularizer.
- The method targets unified multimodal models using autoregressive diffusion.
From the original article
1 min read at huggingface.co
Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users.
By Houyuan Chen, Fanzhou Wang, Xiangyu Fan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.