TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
The authors report less visual drift over 15–20 editing turns while largely retaining the base models’ generation and editing performance.
TailSteer adds a final training stage that concentrates preservation corrections near the end of denoising, keeping earlier steps close to the base model. It learns from the editor’s own denoising rollouts, using unchanged source content as a reference without curated ground-truth editing pairs. Tests span the FLUX, SenseNova, and Qwen model families. The reported gains come alongside largely preserved results on text-to-image and single-image editing benchmarks.
ArXiv · AI/CL/LG's note
TailSteer adds a final training stage that concentrates preservation corrections near the end of denoising, keeping earlier steps close to the base model. It learns from the editor’s own denoising rollouts, using unchanged source content as a reference without curated ground-truth editing pairs. Tests span the FLUX, SenseNova, and Qwen model families. The reported gains come alongside largely preserved results on text-to-image and single-image editing benchmarks.
ArXiv · AI/CL/LG's note
score 5