Model summary
Qwen image-edit variant that supports LoRAs; community LoRAs/fine-tunes are noted as getting very good (e.g., relighting, next-scene prediction).
Qwen-Image Edit is a 20B-parameter image editing model (MMDiT architecture) with LoRA support, focused on controllable edits to existing images. It performs particularly well on product-centric edits, including multi-angle product placement with strong adherence to reference images and prompts. The model can generally follow editing instructions and maintain overall scene composition, but it struggles with fine-grained visual fidelity in people and style-sensitive tasks. In style transfer, it fails to accurately preserve facial features and does not reliably capture specific artists’ styles, instead producing generic painterly looks with overly bright, cheerful colors. When editing parts of an image (e.g., faces, skin, hair), it tends to over-smooth textures, leading to airbrushed, less realistic skin and hair (e.g., beards). Similarly, for product placement with people, clothing and accessories often look too smooth and synthetic, and object scale can be slightly unrealistic (products rendered too large). Overall, Qwen-Image Edit is well-suited for general product image editing and multi-view product generation with good prompt and reference adherence, but is weaker for high-fidelity portrait edits, realistic textures, and precise, artist-specific style transfer.
