AI
On-Policy Self-Distillation for Multi-Turn Image Editing
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 11:49:19 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 06:18:25 UTC.
Briefings
This study investigates on-policy self-distillation for multi-turn image editing, addressing the degradation of existing models under recursive instruction application. The authors propose MT-OPSD, a framework that trains models on self-generated conditioning states using a clean-conditioned teacher, eliminating the need for multi-turn annotations. Evaluated via LME-Bench with 100 ten-turn sessions, MT-OPSD significantly improves long-horizon editing success and reduces multi-turn collapse while maintaining single-turn quality across three editing backbones.
Key points
- MT-OPSD is an on-policy self-distillation framework that trains models on self-generated conditioning states using a clean-conditioned teacher, eliminating the need for multi-turn annotations.
- MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality across three editing backbones.
Context and limitations
The study evaluates MT-OPSD across three editing backbones using LME-Bench with 100 ten-turn editing sessions; results are limited to this specific benchmark and do not generalize to other editing backbones or multi-turn scenarios.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 11:49:19 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibhkxbg6jsqnvqbtnv6rwew6gxol2allwnv46ojjjwrm2ap4fbj6e