Diffusers 0.41.0 Adds Qwen-Image 2.1 Pipeline, Deprecates ONNX Support

What changed?

Diffusers 0.41.0 introduces support for the Qwen-Image 2.1 model, enabling unified text-to-image generation and image editing in one pipeline. New capabilities include handling multiple reference images, native RGBA output for transparent images, and support for LoRA training. Single-file loading support was added or extended for Qwen 2.1, Minimax H3, and Krea 2 models, simplifying model management. The release deprecates direct ONNX support, advising users to switch to the Optimum library for ONNX Runtime workflows, and removes previously deprecated APIs, such as the LuminaText2ImgPipeline and Lumina2Text2ImgPipeline aliases. Additional highlights include memory-optimized tensor-parallel checkpoint loading and enhancements in Cosmos 3, Wan 2.2, and LTX-2.5 pipelines.

Why does it matter to an everyday developer?

These changes provide more efficient, capable image generation and editing workflows directly within Diffusers. Developers now have access to a model (Qwen-Image 2.1) that simplifies the pipeline for both image creation and editing, saving development time and unlocking new creative capabilities, such as native transparency handling. The introduction of tensor-parallel checkpoint loading can reduce memory usage on multi-GPU setups, making the library more practical for larger models. Removal of deprecated APIs and ONNX support will require some users to update their code and migrate ONNX workflows to the Optimum library, but this consolidates inference and training logic and improves compatibility.

What can the developer do now?

• Adopt the Qwen-Image 2.1 pipeline for unified image generation and editing, leveraging new features like RGBA output and LoRA training.