2026-11-06 · America/Los_Angeles · 论文 · #2

TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing

Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired…

热度 54.2 / 100；排名与评分保留该期记录。