TailSteer Improves Multi-turn Image Editing Robustness
Original title: TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
TailSteer reduces cumulative drift in multi-turn editing while preserving base model performance.
TailSteer is a final-stage training framework for multi-turn image editing that keeps early denoising dynamics close to the base model and concentrates preservation corrections near the trajectory tail. Corrections are learned from the editor's own on-policy denoising rollouts using unchanged source content as reference. Experiments across FLUX, SenseNova, and Qwen families show substantial reduction in drift over 15-20 editing turns.
- Reduces cumulative drift over 15-20 editing turns
- Largely preserves base-model performance on text-to-image benchmarks
- Learned from on-policy denoising rollouts without curated ground-truth pairs
- Unlocks long-horizon editing robustness in foundation models
From the original article
1 min read at huggingface.co
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.
By Ruikang Li, Zixuan Chen, Jiarui Wu · 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.