Qwen releases Qwen-Image-2512 text-to-image update
Alibaba's Qwen team has released Qwen-Image-2512, a December update to its text-to-image foundation model that targets the "AI-generated" look in human images and claims the top spot among open-source models.
The Qwen team announced Qwen-Image-2512 as the December update to Qwen-Image, its text-to-image foundation model. The base version came out in August. The new model can be tried in Qwen Chat.
The team lists three main improvements over the August release:
- Human realism: less of the "AI-generated" look, especially in images of people.
- Natural detail: finer rendering of landscapes, animal fur and other natural textures.
- Text rendering: more accurate text, better layout, and better composition when an image combines text and pictures.
On performance, Qwen says it ran more than 10,000 rounds of blind evaluations on AI Arena. It says the results show Qwen-Image-2512 is the strongest open-source model available and is "highly competitive" with closed-source models. The post does not give scores, rankings or the names of the models it was compared against.
The rest of the announcement compares outputs from the two versions using the same prompts. Qwen says the new model renders individual hair strands that the August version blurred together. It also says the model shows age more convincingly, including wrinkles on an elderly couple, and draws background objects like desks and bedding more clearly. In one example, the new model correctly followed a "body leaning slightly forward" instruction that the earlier version missed. Landscape examples point to better water flow, mist, foliage and range of greens.
The part of the post covered here does not give the model's size, licence or download details.