Multi-turn image editing is meant to let users refine an image through a sequence of natural-language instructions, but current systems tend to show noticeable drift after only a few turns. A new arXiv preprint introduces MT Forcing, an approach that couples autoregressive diffusion with multi-turn editing to improve stability and consistency.

According to the abstract, the method is designed for interactive visual content creation, where users expect each edit to build on the previous one without degrading the image. The authors frame drift as a core obstacle and position MT Forcing as a way to keep edits coherent over longer instruction sequences.

The paper is available as arXiv:2610.25404. The abstract excerpt does not include experimental details or quantitative comparisons, so the reported contribution is the method itself rather than measured outcomes.