Qwen Releases Qwen-Image-2.1… Lightweight 7B Model with Transparency Support

The Alibaba Qwen team released the new image generation model Qwen-Image-2.1 on September 20. Weights are available for download on Hugging Face, ModelScope, and GitHub.

Model Architecture and Performance

• Parameter Reduction: Significantly reduced the number of parameters to 7B compared to the previous version (20B), adopting a 32-layer DiT architecture.
• Performance: According to Qwen's benchmarks, it scored 60.28, slightly exceeding Nano Banana 2.0 (59.82) and GPT-1.5 (59.65), though independent verification has not yet been conducted.
• Resolution: Supports 2K native resolution ranging from 2048x2048 to 2752x1536.

Key Features and Requirements

• Transparency Support: Directly generates PNGs with alpha channels without post-processing.
• Multi-Image Reference: Allows input of up to 10 reference images to synthesize scenes while maintaining object consistency.
• Hardware Requirements: Requires approximately 33GB of disk space, and the decoder can run on consumer-grade hardware equivalent to an RTX 3090 (24GB VRAM). The official code supports offloading certain model parts to system RAM, and quantized versions are also available.

License Change

• Unlike the Apache 2.0 license of the first version, this version applies the Qwen Research License, restricting usage to research purposes only.