Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition (2512.15603v1)

Abstract: Recent visual generative models often struggle with consistency during image editing due to the entangled nature of raster images, where all visual content is fused into a single canvas. In contrast, professional design tools employ layered representations, allowing isolated edits while preserving consistency. Motivated by this, we propose \textbf{Qwen-Image-Layered}, an end-to-end diffusion model that decomposes a single RGB image into multiple semantically disentangled RGBA layers, enabling \textbf{inherent editability}, where each RGBA layer can be independently manipulated without affecting other content. To support variable-length decomposition, we introduce three key components: (1) an RGBA-VAE to unify the latent representations of RGB and RGBA images; (2) a VLD-MMDiT (Variable Layers Decomposition MMDiT) architecture capable of decomposing a variable number of image layers; and (3) a Multi-stage Training strategy to adapt a pretrained image generation model into a multilayer image decomposer. Furthermore, to address the scarcity of high-quality multilayer training images, we build a pipeline to extract and annotate multilayer images from Photoshop documents (PSD). Experiments demonstrate that our method significantly surpasses existing approaches in decomposition quality and establishes a new paradigm for consistent image editing. Our code and models are released on \href{https://github.com/QwenLM/Qwen-Image-Layered}{https://github.com/QwenLM/Qwen-Image-Layered}

• The paper introduces an end-to-end framework that decomposes raster images into semantically disentangled RGBA layers for inherent image editability.
• It leverages a unified RGBA-VAE, a diffusion transformer, and a multi-stage training protocol to ensure high reconstruction fidelity and robust editing.
• Experimental results demonstrate state-of-the-art metrics in decomposition quality, enabling precise, non-destructive edits and scalable image synthesis.

Qwen-Image-Layered: Inherent Editability via Layer Decomposition

Introduction

Qwen-Image-Layered introduces an intrinsically editable image representation through the end-to-end decomposition of raster images into semantically disentangled RGBA layers. Recognizing that established visual generative models are fundamentally constrained by flat, entangled pixel spaces, which propagate inconsistencies under editing operations, this work explicitly transitions to a layered paradigm. By doing so, Qwen-Image-Layered ensures that independent manipulation of each semantic component is feasible without degradation or drift in either spatial or contextual fidelity.

Figure 1: Qwen-Image-Layered decomposes diverse open-domain images into distinct, semantically consistent layers enabling precise, isolated edits.

Methodology

Qwen-Image-Layered formulates the image-to-multi-layer problem as a direct mapping from an input RGB image to a variable-length stack of RGBA layers. Decomposition is supported by three foundational modules:

RGBA-VAE

A unified four-channel VAE directly encodes both RGB and RGBA images into a shared latent space, minimizing representation shift when generating target layers from raster images. This strategy eliminates the disjoint latent distributions that plagued previous dual-VAE decomposers, ensuring encoding consistency and optimizing reconstruction fidelity. The model is trained with a combination of reconstruction, perceptual, and regularization losses for robust generalization across modalities.

Variable Layers Decomposition MMDiT (VLD-MMDiT)

Decomposition into a variable (N) number of semantic layers is architected through VLD-MMDiT, a conditional diffusion transformer. Unlike prior recursive frameworks limited to foreground-background decomposition, VLD-MMDiT concatenates representations of the input image and target layers, enabling direct cross-layer and intra-layer attention. Layer3D RoPE introduces explicit layer index encoding, supporting arbitrary depth with layer-distinguishable features.

Figure 3: VLD-MMDiT concatenates visual and textual conditions, leveraging Layer3D RoPE for permutation-invariant modeling of arbitrary decompositions.

Multi-Stage Training Protocol

A three-phase curriculum is employed:

1. Text-to-RGBA: Adapts the diffusion decoder to the new RGBA latent space by joint training on standard and alpha-augmented samples.
2. Text-to-Multi-RGBA: Facilitates multilayer generation, conditioning the model to simultaneously produce a composite and its decomposed layers.
3. Image-to-Multi-RGBA: Integrates input image guidance, aligning the decomposition procedure with actual RGB-to-multilayer translation.

This progression bridges the modality gap and ensures competitive initialization for the most complex decomposition tasks.

Dataset Construction

Due to the lack of richly annotated multilayer datasets, a pipeline to mine, filter, and annotate Photoshop documents (PSDs) was developed. Non-contributory layers are removed, and spatially non-overlapping groups are merged to standardize complexity and reduce noise. Additionally, the data is annotated with captions generated by a vision-LLM to facilitate text-conditioning tasks.

Figure 2: Distribution of merged layer counts within the processed dataset shows effective reduction in layer complexity for trainability.

Experimental Results

Quantitative Evaluation

Qwen-Image-Layered establishes state-of-the-art metrics for decomposition quality on public benchmarks, including the Crello dataset. It achieves notable gains in both RGB L1 error and Alpha soft IoU, even under flexible layer-merging protocols, substantially surpassing prior decomposition frameworks such as LayerD.

Ablation Study

Component analysis confirms the necessity of the Layer3D RoPE, RGBA-VAE, and multi-stage training. Removing any constituent degrades the soft IoU and increases color and alpha distortion, validating the architectural alignment.

RGBA Reconstruction

RGBA-VAE outperforms specialized reconstruction VAEs on the AIM-500 dataset, with substantial improvement in PSNR, SSIM, rFID, and LPIPS, confirming the reliability of the underlying latent representation for alpha-blended tasks.

Qualitative Analysis: Decomposition and Editing

LayerD and other segmentation-inpainting decomposers yield erroneous masks and texture fill artifacts. In contrast, Qwen-Image-Layered generates highly semantic and visually continuous layers with minimal artifacts. This inherent disentanglement allows for aggressive edits—such as object deletion, resizing, or repositioning—without unintended perturbations elsewhere in the composite.

Figure 4: LayerD exhibits segmentation and inpainting errors, whereas Qwen-Image-Layered produces clean, edit-ready layers.

Figure 5: Qwen-Image-Layered enables precise local edits (e.g., resizing or translating objects) with no collateral change, outperforming diffusion-based editors suffering from spatial drift.

Multi-Layer Image Synthesis

For text-to-multi-layer generation, Qwen-Image-Layered directly synthesizes semantically coherent images. When coupled with an upstream T2I generator, it maintains compositionality and visual quality unattainable by ART and similar systems.

Figure 6: End-to-end text-driven layer generation and T2I+decomposition both deliver superior semantic and visual quality over existing approaches.

Implications and Future Prospects

Qwen-Image-Layered provides a marked advance in the theoretical separation of representation and manipulation for image editing. Its framework eliminates the inherent entanglement of raster pixels, offering a scalable solution to geometric and semantic drift. Practical workflows in design and visual content creation can now harness generative pipelines for precise, repeatable edits that are fully non-destructive and context-preserving.

On the theoretical front, its Layer3D RoPE and sequence-concatenated attention mechanisms may generalize to other permutation-invariant structured data, including 3D vision, video, and multimodal fusion. Expanding the annotated multilayer corpus will further drive research into high-arity, open-domain decomposition.

Conclusion

Qwen-Image-Layered advances the state of the art in image decomposition and inherent editability by directly mapping input images to semantically distinct RGBA layers. Through robust architectural innovation and staged learning, it eliminates primary sources of editing inconsistency, demonstrating both strong empirical performance and wide-ranging application potential in AI-driven creative authoring (2512.15603).