Image editing using AI has always been a bit hit and miss, given how AI systems typically recreate the entire image which changes many aspects of the original that the user intended to leave untouched, but a new model from China is looking to change that.
Qwen-Image-Layered, the latest release from Alibaba’s Qwen AI research team, introduces a novel approach to AI-powered image editing by decomposing images into multiple RGBA layers. The model, which has been fully open-sourced, allows users to manipulate individual layers independently without affecting other elements of the composition, addressing one of the most persistent challenges in generative image editing.
A New Paradigm for AI Image Editing
Traditional AI image editing tools face a fundamental problem: when modifying one aspect of an image, the entire composition is typically regenerated, leading to unintended changes throughout the image. This lack of precision has limited the practical utility of AI editing tools for professional workflows where consistency and control are paramount.
Qwen-Image-Layered takes a different approach by physically isolating semantic or structural components into distinct layers, similar to how professional editing software like Photoshop organizes complex compositions. Each layer is represented in RGBA format, providing full transparency control and enabling true native editability. This means users can resize, reposition, or recolor specific elements without triggering a full image regeneration that might alter unrelated content.
Flexible Layer Control and Infinite Decomposition
The model offers prompt-controlled structure, allowing users to specify anywhere from three to ten layers depending on their needs. This flexibility enables both coarse layouts for simple edits and fine-grained detail separation for complex compositions. Users can explicitly direct the model on how to break down an image, making the decomposition process both predictable and tailored to specific editing requirements.
Perhaps most notably, Qwen-Image-Layered supports what the team calls “infinite decomposition,” the ability to continue breaking down layers into sub-layers recursively. This nested approach means users can drill down to any level of detail they require, creating hierarchical structures that mirror the complexity of real-world images.
Bridging Raster and Structured Representations
By reimagining images as composable layers rather than flat raster data, Qwen-Image-Layered bridges a significant gap between traditional image formats and the structured, editable representations that professional workflows demand. The model’s approach enables high-fidelity elementary operations while maintaining consistency across edits, a combination that has proven elusive for most generative AI tools.
The open-source release of Qwen-Image-Layered signals growing sophistication in AI image editing capabilities and could accelerate development of more precise and controllable creative AI tools. For businesses and creators who have been hesitant to adopt AI editing due to its unpredictable nature, this layered approach may represent a more viable path toward integrating generative AI into professional image editing workflows.
Image editing emerging as a top AI use-case
Image editing is quickly turning into one of AI’s most visible use-cases. Last year, the viral Ghibli trend had got ChatGPT so many new users that Sam Altman had posted that their servers were melting from all the demand. Later in the year, Google had released its own Nano Banana image model, which had also gone viral, and brought millions of new users to Gemini. Chinese companies too have some very capable image models of their own. And with Qwen looking to fix one of the most persistent issues with AI image editing — and allowing for precise edits — it appears that AI editing could have yet another viral moment in the coming months.