ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning
ACEdit stabilizes multi-turn image editing by distilling a teacher while the student edits its own outputs.
ACEdit is a training scheme for stable multi-turn image editing under accumulating distribution shift. The student is forced to edit its own previous outputs, while a distillation objective uses a teacher editor that remains anchored to the original clean image and a compound instruction. Experiments report substantially less quality drop over long edit sequences without sacrificing instruction adherence.
- Multi-turn image edits accumulate distribution shift and degrade quality.
- ACEdit trains the student on its own previous outputs using Self-Forcing.
- A teacher stays anchored to the original image and a compound instruction.
- The method reduces long-sequence degradation while keeping instruction adherence.
From the original article
1 min read at arxiv.orgImage editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor's original training and multi-turn inference.
By Sara Dorfman · arxiv.org
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.