Editing Images into Reusable Layers via Compositional Edit-Prior Distillation

Key points

• High-quality image editing requires precise modification of target regions while preserving the content of unrelated regions.1
• Existing editing models produce flattened outputs that limit precise local manipulation and layer reuse, while decomposition methods typically optimize layer recovery separately from instruction-driven editing.1
• At its core, In-Context RGB-Alpha Token Modeling jointly represents the edited canvas and an ordered sequence of RGB-Alpha layers within a single token sequence, enabling semantic edits and explicit control over individual layer content.1
• To retain the pretrained model's editing capability during adaptation to layered outputs, we propose Compositional Edit-Prior Distillation, which transfers its editing prior to both the edited canvas and the image composed from the predicted layers.1

Sources (1)

• [1]Editing Images into Reusable Layers via Compositional Edit-Prior Distillation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 03:07 PM

  To this end, we introduce LayerEditor, a unified framework that couples instruction-guided image editing with controllable generation of reusable RGBA layers.

  High-quality image editing requires precise modification of target regions while preserving the content of unrelated regions.

Extractive summary: sentences quoted from the sources.