Paper review · October 24, 2026

Editing images into reusable layers: compositing-based editing prior distillation (LayerEditor)

Editing Images into Reusable Layers via Compositional Edit-Prior Distillation

Generative media Computer vision Advanced Upvotes 73 cs.CV jingdong

It jointly learns instruction-based image editing and RGBA layer generation as a single token sequence, and distills the teacher's editing ability on-policy to obtain accurate layers while preserving editing quality.

Problem

Existing editing models output only a single flattened image with all elements merged, making fine adjustments or element reuse difficult. Splitting the result into layers with a decomposition model after editing optimizes the two stages separately, so the layers end up incomplete or messy. Directly fine-tuning a pretrained editing model to output layers also degrades its original editing ability.

Method

The authors extended FireRed-Image-Edit-1.1 with a rank-256 LoRA. The edited canvas and the RGB and alpha fields of N layers are placed in one sequence and denoised together (In-Context RGB-Alpha Token Modeling). Fields and stacking order are distinguished with layer-aware RoPE coordinates, and alpha is reconstructed with a dedicated decoder. C-EPD queries the frozen original editing model (the teacher) at intermediate states along the student's denoising trajectory. The teacher supervises both the direct output velocity and the image produced by alpha-compositing the predicted layers. For training data, they built LayerEditor-100K, consisting of 100,000 text, object and poster appearance editing pairs with annotations of 2 to 10 layers.

Results

On Crello-Testset (2,994 samples), composite image SSIM is 0.8351, higher than Seedream5-Pro Joint (0.7902), and composite FID is 33.86, lower than FR+QIL-Decomp-FT (57.15).
On Crello-Testset layer metrics, it scores Alpha IoU 0.6074 (Seedream5-Pro Joint 0.5139) and RGBA L1 0.0350 (QIL-Edit-SFT 0.0993), the best among the compared methods.
On ScaleEdit-Testset (1,000 samples), composite SSIM is 0.8839, higher than Seedream5-Pro Joint (0.8365), and Alpha IoU is the highest at 0.6802 (0.6583).
Direct output SSIM on Crello is 0.8530, comparable to the Seedream5-Pro sequential approach (0.8586). This is presented as evidence that editing quality was maintained while adding layer output.
On rubric scores (1–10), it is highest on every Crello item, including 6.14 for editing consistency and 6.83 for layer separation.

Limitations

The data and benchmarks are based on graphic design sources such as Crello and PrismLayersPro, and on ScaleEdit, so these experiments alone cannot show whether it generalizes to complex layer decomposition of natural photos. The number of layers is capped at 10. Rubric scores rely on automatic evaluation. Seedream5-Pro could not be evaluated on some documents due to API constraints.

Why it matters

For work where outputs are revised repeatedly, such as advertising and e-commerce design, edit results can be received directly as independently editable layers. It is also a useful reference for using on-policy distillation to preserve existing abilities when extending a pretrained model to a new output format.

Key terms

RGBA layer: An image layer that carries both color (RGB) and opacity (alpha). Stacking and compositing layers from back to front produces the final image.
On-policy Distillation: A distillation (knowledge distillation) method in which the student queries the teacher for target outputs at the states the student actually passes through. It reduces the gap between the training data distribution and the actual generation distribution.
Alpha Compositing: The standard operation that blends a front layer with the background behind it using each layer's alpha values to produce a single image.
LoRA: A technique for efficiently fine-tuning a model by freezing the original weights and training only small added low-rank matrices.

Original

Authors: Wen Xue, Yong Liu, Xueheng Li, Si Wu, Yong Xu, Yipeng Sun, Yan Li, Simiu Gu · Posted 2026-10-17 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.