ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

Key points

• Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns.1
• 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.1
• We introduce an asymmetric conditioning scheme: while the student performs edits sequentially on its own outputs, the teacher remains anchored to the original clean image and receives a compound editing instruction.1
• Experiments show substantially reduced degradation over long multi-turn sequences while maintaining editing quality and instruction adherence.1

Sources (1)

• [1]ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 05:52 PM

  We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift.

  Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns.

Extractive summary: sentences quoted from the sources.