Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition (huggingface.co)

Qwen-Image-Layered has been introduced as a cutting-edge end-to-end diffusion model that decomposes standard RGB images into multiple semantically disentangled RGBA layers. This advancement addresses a common issue in visual generative models, where editing tends to disrupt the consistency of the image due to the fusion of various visual elements. By allowing independent manipulation of each RGBA layer, Qwen-Image-Layered enhances editability and preserves overall image integrity, making it a significant tool for designers and creators who rely on layered editing. The model comprises three innovative components: an RGBA-VAE for unified latent representation of images, a Variable Layers Decomposition MMDiT (VLD-MMDiT) architecture to manage variable layer counts, and a Multi-stage Training strategy to modify pre-existing image generation models for multilayer decomposition. Additionally, it includes a novel pipeline that extracts and annotates multilayer images from Photoshop documents, addressing the issue of limited high-quality training data. Experimental results indicate that Qwen-Image-Layered outperforms existing methods in decomposition quality, marking a transformative step for consistent image editing in the AI/ML space.