ACEdit uses asymmetric conditioning to stabilize multi-turn image edits
Original title: ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning
The student edits its own prior outputs while a teacher stays on the clean image with a compound instruction.
Image editors often degrade across turns as small distribution shifts accumulate. ACEdit applies Self-Forcing by training the editor on its own previous outputs, using distillation because those trajectories diverge from ground truth. Asymmetric conditioning keeps the teacher on the original clean image while the student edits sequentially. The authors report less degradation over long sequences while quality and instruction adherence hold.
Unconfirmed: The reported drop in multi-turn degradation comes only from the authors' experiments.
- Direct supervision on self-generated editing trajectories is described as unreliable.
- A teacher editor supplies the training signal during self-forced turns.
- The method targets the gap between single-turn training and multi-turn inference.
- The teacher receives a compound editing instruction rather than turn-by-turn student inputs.
From the original article
1 min read at huggingface.co
Image 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, Nir Goren, Gal Barak · 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.