Alibaba claims new image model beats Google's Nano Banana 2.0
What's the story
Alibaba's Qwen AI team has launched a new open-weight model for image generation and editing, called Qwen-Image-2.1. The model is said to outperform most closed models on its own benchmark, despite having just seven billion parameters. On the Qwen-Image-Bench evaluation test, Alibaba's new model scored 60.28, beating Google's Nano Banana 2.0 (59.82). However, independent benchmarks are yet to be conducted.
Capabilities
The model can handle up to 10 reference images simultaneously
Qwen-Image-2.1 can generate and edit transparent images (RGBA) natively, allowing users to change text or isolate objects on transparent layers.
The model can handle up to 10 reference images simultaneously for tasks like group portraits, virtual try-ons, or room design.
It also supports local edits with circles, masks, or painted marks as guides.
According to Qwen's team, changes in architecture and KV cache reuse have improved inference speed especially when using multiple reference images at once.
Access
Qwen-Image-2.1 is available on Hugging Face, GitHub, and Model Scope
Qwen-Image-2.1 is available on Hugging Face, GitHub, and Model Scope. There is also a demo available on Hugging Face for users to try out the capabilities of this advanced model.
Qwen-Image-2.1 can also run on consumer laptops with GPUs such as the RTX 3090.
However, it's worth noting that the research license for this tool prohibits commercial use. Businesses interested in using it commercially will have to apply separately to Qwen for an appropriate license.