On December 22, 2025, Ali unveiled its newly developed open-source image generation model, Qwen-Image-Layered. This model marks a groundbreaking achievement as it is the first to incorporate PS-level layer comprehension and image generation capabilities within a single framework. Leveraging a self-developed, innovative architecture, the model can deconstruct images into multiple layers, facilitating highly precise AI image editing with minimal deviation, or "zero-drift" editing. This breakthrough effectively addresses the challenge of maintaining consistency in AI-generated images and expedites the integration of large models into professional design domains.

Qwen-Image-Layered achieves a more precise "physical perception" of the real world by employing "layering" and "inpainting" techniques. This enables AI to transition from mere "image interpretation" to true "spatial reconstruction". Its proprietary RGBA-VAE encoding integrates an Alpha channel into RGB images, thereby introducing the concept of layers to the model. The innovative VLD-MMDiT architecture, in conjunction with "layer-level 3D positional encoding", empowers AI to automatically envision the background textures of obscured areas, resulting in a profound understanding and generation of layers and spaces.

At present, Qwen-Image-Layered is readily available for download on ModelScope and HuggingFace platforms. Developers and enterprises can access and utilize it for commercial applications free of charge.