Editing Images into Reusable Layers via Compositional Edit-Prior Distillation
Wen Xue, Yong Liu, Xueheng Li, Si Wu, Yong Xu, Yipeng Sun, Yan Li, Simiu Gu
South China University of Technology · Tsinghua University · University of Science and Technology of China · Peng Cheng Laboratory · JD.com
Published October 17, 2026 · on the feed October 23, 2026 · ▲ 196 · arXiv 2610.24738
The abstract
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.
The Brief
2:38A short podcast summary of the paper. Written and voiced by AI.
Transcript
Editing Images into Reusable Layers via Compositional Edit-Prior Distillation. By Wen Xue and seven others, from South China University of Technology, Tsinghua University, and three other institutions. Published on the seventeenth of October. No code repository is linked to it.
High-quality image editing has two obligations at once: modify the target region precisely, and leave the content of every unrelated region alone. Existing editing models meet the first in a way that undercuts reuse. They hand back a flattened output, a single image with everything fused together, which limits precise local manipulation and leaves nothing that can be lifted out and used again as a layer. Decomposition methods do recover layers, but they typically optimize that recovery separately from instruction-driven editing. One family edits without layers. The other produces layers without editing.
LayerEditor is a unified framework that couples the two. Given an instruction, it performs the edit and, in the same pass of work, generates reusable RGBA layers: images that carry colour together with an alpha channel. The edited picture arrives already separated into ordered parts, each of which can be controlled on its own.
At the core sits a component called In-Context RGB-Alpha Token Modeling. It represents the edited canvas and an ordered sequence of RGB-Alpha layers jointly, inside a single token sequence. Holding the canvas and every layer in one sequence is what lets the model make semantic edits while keeping explicit control over what each individual layer contains.
Adapting a pretrained editing model to layered outputs carries a risk: losing the editing capability it started with. Compositional Edit-Prior Distillation is the safeguard. It transfers the pretrained model's editing prior to two targets, the edited canvas and the image composed from the predicted layers. Both the direct result and the reassembled stack are held to what the original editor already knew.
Training and evaluation needed data that did not exist, so the authors construct LayerEditor one hundred K, a large-scale paired image-editing dataset with ordered RGBA layer annotations, and introduce dedicated benchmarks for layered editing. Across extensive experiments, LayerEditor maintains strong editing fidelity while generating accurate RGBA layers. It supports flexible manipulation at the level of individual layers, and it also handles multi-layer decomposition.
What changes here is the form an edit ends in. Instead of one flattened image, the result is a canvas plus an ordered set of layers that can be moved, adjusted or reused, produced by the same model that did the editing, with distillation keeping that model's editing skill intact through the change.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A perfect edit is not much use if you can never pull it apart again. Start with that frustration.
You run design for an online shop. An editing model changes the banner text nicely, but hands back one flat picture with text, products and background fused together.
The road ahead
- 1 · The problem ~7 min Editors hand back one flat picture, and splitting it into layers afterwards gives messy layers because the splitter never knew the edit.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Generate the edited canvas and its ordered RGBA layers in one token sequence, while a frozen copy of the original editor keeps editing quality intact.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Lay canvas and ordered layers in one sequence, stack the predicted layers into an image, and match canvas and stack to the frozen editor.
- 6 · The experiment ~6 min You’ll guess: Adapting an image editor to output separate transparent layers risks hurting its edits. Did the final model keep strong editing fidelity?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.