TailSteer tames multi-turn drift across FLUX, SenseNova, Qwen

Repeated edits to the same image tend to compound small unintended changes into what the authors call "severe visual drift," and a nine-author team writing on arXiv on November 4 says a final training-stage fix can suppress it without disturbing a base model's pretrained priors. The method, TailSteer, trains an editor on its own denoising rollouts and concentrates corrections "near the trajectory tail, where fine-grained deviations can be suppressed," rather than retraining across curated multi-turn edit pairs or relying on inference-time mechanisms the authors describe as "difficult to scale to foundation models."

Across the FLUX, SenseNova, and Qwen families, TailSteer "substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance" on text-to-image and single-image editing benchmarks, according to the abstract. The paper frames the approach as additive rather than corrective: "long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities."

The abstract publishes no per-benchmark scores and does not announce a code release. "Substantially reduces" and "largely preserving" are the paper's own words; no numeric drift or fidelity figures appear in the public summary, and the quantitative tradeoff between the two remains for the full text or follow-up reporting to settle.