Qwen-Image-Edit-2511 is now live on Hugging Face and ModelScope. The update focuses on consistency problems that plagued earlier versions, particularly when editing images with multiple people or preserving character identity across edits.
The headline improvements: stronger multi-person consistency for group shots, built-in LoRAs from the community (eliminating the need for users to source and apply them separately), and better geometric reasoning for structural edits. Alibaba's announcement also claims reduced "image drift," the tendency for identity and character details to shift unpredictably during the editing process.
Industrial and product design generation gets attention too. The team positions this for workflows beyond portraits, including engineering mockups and product poster creation. Like its predecessor, 2511 is built on the 20B Qwen-Image foundation model with Qwen2.5-VL handling visual semantic control.
For local users, the model already has diffusers support via a new pipeline parameter. A Lightning LoRA for faster inference dropped alongside the main release, cutting generation to as few as four steps.
The Bottom Line: Qwen-Image-Edit-2511 addresses the consistency complaints that dogged 2509, though performance claims are self-reported and await independent testing.
QUICK FACTS
- Model: Qwen-Image-Edit-2511 (based on 20B Qwen-Image)
- Availability: Hugging Face, ModelScope, Qwen Chat
- Key improvements: Multi-person consistency, built-in LoRAs, identity preservation
- Inference acceleration: 4-step Lightning LoRA available
- License: Apache 2.0
