Qwen-Image-Edit
Image-editing version of Qwen-Image feeds the picture to Qwen2.5-VL for semantics and a VAE encoder for appearance; edits Chinese and English text in place.
Dual-path conditioning enables both low-level appearance edits (add, remove, modify with the rest unchanged) and semantic edits (rotation, style transfer, IP creation), plus bilingual text editing that preserves font, size and style. Built on the 20B Qwen-Image.
- Date
- Tuesday, 19 August 2025
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Blog dated 2025-08-19 in the feed (2025-08-18 UTC; HF repo 2025-08-17). SOTA claims are Alibaba's; benchmarks not opened.
Sources
This record was checked against its sources on 6 October 2026. How we check