Precise bilingual image edits, now state of the art
Qwen has introduced Qwen-Image-Edit, an image editor built on the 20B Qwen-Image model that unifies high-level semantic control with low-level appearance fidelity. By routing inputs through Qwen2.5-VL and a VAE Encoder, it delivers precise bilingual text edits and robust transformations like rotation, style transfer, and region-locked tweaks. Positioned as a foundation model with state-of-the-art results, it targets creators needing reliable, fine-grained edits at scale.
Points clés
- Qwen unveiled Qwen-Image-Edit on August 19, 2025, built atop the 20B Qwen-Image model
- The system feeds images into Qwen2.5-VL (semantic control) and a VAE Encoder (appearance control) simultaneously
- It supports semantic edits like IP creation, style transfer, and object rotation up to 180 degrees while preserving identity
- It enables appearance edits that keep untouched regions identical while adding, removing, or modifying elements
- Precise bilingual (Chinese and English) text editing preserves original font, size, and style, including small poster text
- Public benchmark evaluations report state-of-the-art performance for image editing tasks
- Access is available in Qwen Chat via the “Image Editing” feature
- Demonstrations include MBTI-themed capybara emoji packs, signboard insertion with realistic reflections, and fine object cleanup (e.g., hair strands)
- Additional examples show targeted letter color changes, background swaps, and clothing changes
- A chained workflow corrected Chinese calligraphy (e.g., refining the character “稽” to use “旨” instead of “日”) to complete Lantingji Xu
À retenir
Start simple: try the “Image Editing” mode in Qwen Chat, lock the regions you don’t want touched, and let the model do the pixel-perfect heavy lifting. For tricky cases, go step by step—chained edits are your friend when one-shot miracles refuse to happen (characters and capybaras can be divas). And yes, it edits Chinese and English text while keeping fonts intact—so you can finally fix that poster typo without summoning a designer at 2 a.m., imagine that.
Sources
