arxiv:2610.24738

Editing Images into Reusable Layers via Compositional Edit-Prior Distillation

Published on Oct 17

· Submitted by

xue wen on Oct 23

jingdong

Upvote

Authors:

Wen Xue ,

Abstract

LayerEditor unifies instruction-guided image editing with reusable RGBA layer generation via joint token modeling and edit-prior distillation, enabling precise local manipulation and layer reuse.

Generated by thinkingmachines/Inkling-Small

High-quality image editing requires precise modification of target regions while preserving the content of unrelated regions. 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. To this end, we introduce LayerEditor, a unified framework that couples instruction-guided image editing with controllable generation of reusable RGBA layers. 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. 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. To support training and evaluation, we construct LayerEditor-100K, a large-scale paired image-editing dataset with ordered RGBA layer annotations, and introduce dedicated benchmarks for layered editing. Extensive experiments demonstrate that LayerEditor maintains strong editing fidelity while generating accurate RGBA layers, supporting both flexible layer-level manipulation and multi-layer decomposition.

View arXiv page View PDF Add to collection

Community

Davidscut

Paper author Paper submitter about 12 hours ago

We introduce LayerEditor, a unified framework that couples instruction-guided image editing with controllable generation of reusable RGBA layers. Given a source image, an editing instruction, and a requested layer count, LayerEditor jointly generates an edited canvas and an ordered RGBA layer stack, enabling explicit control over individual layer content.

In-Context RGB-Alpha Token Modeling. LayerEditor jointly represents the edited canvas and ordered RGB-Alpha layers within a single token sequence, with appearance and opacity modeled as distinct, interacting fields. Shared attention captures dependencies between the canvas and layer stack, while layer-aware rotary coordinates encode field identity and stacking order while preserving spatial correspondence. This formulation supports variable layer counts within a shared architecture.
Compositional Edit-Prior Distillation (C-EPD). To preserve the pretrained model’s editing capability during adaptation to layered outputs, C-EPD queries a frozen editing teacher at intermediate states along the student’s own denoising trajectory. It combines direct-output velocity matching with pixel-space supervision of the image composed from the predicted RGBA layers. Gradients propagate through decoding and alpha compositing, transferring the pretrained editing prior to both the edited canvas and the layered output.
LayerEditor-100K and dedicated benchmarks. We construct LayerEditor-100K, a large-scale paired image-editing dataset containing 100K examples, each with a source image, an editing instruction, an edited target, and an ordered RGBA stack of 2–10 layers. We further introduce Crello-Testset and ScaleEdit-Testset to evaluate editing fidelity, layer quality, and controllability over layer count and user-specified layer content.