Introducing Qwen Image

Team fal

• 3 min read
Introducing Qwen Image

We are proud to announce fal's inference and training endpoints Alibaba's Qwen Image model.

It is has been an incredible month for generative media. We've been blessed not one, but two unbelievable high quality image models Wan 2.2 14b and Qwen Image.

I went into Wan 2.2 14b in the last blog, but Qwen came out merely a day after is amazing in some unique ways.

Builtin Styles

The incredible prompt following leads to impressive style representation and text rendering.

Qwen Image can generate diverse styles like Thomas Cole (left) and Thangka (right) out of the box

In general, image models are not able to generate precise styles. The difference between a "painting" and "impressionist painting" is often barely noticeable in other models, but Qwen Image has a much more nuanced understanding of art styles.

Qwen Image generations of (from left to right) Impressionism, Romanticism, Rococo, Barque and Art Nouveau styles highlighting its deep understanding of different art styles.

In addition a more diverse and accurate palette of styles, Qwen Image excels at intricate text rendering.

Compared to other popular image generators, Qwen Image can render text that follows the prompt much better in complex scenes.

Qwen Image (right), Wan 2.2 (center) and FLUX.1 [dev] (left) with the prompt 'calligraphy "Zen Garden" sign hangs above carefully raked sand and miniature bonsai trees. Three cultural books sit on a bamboo shelf: "Bonsai Care", "Garden Design", and "Meditation Practice". Stone lanterns and bamboo water features create a peaceful atmosphere, while cherry blossoms drift gently in the breeze.'

Personalization

Using the newly released fal Qwen Image Trainer, we can create high quality personalized generations with a handful of images.

Here I have used seven iPhone headshots to create a LoRA of my face. You can see below that Qwen Image was able to generate an accurate and high quality portrait with the resultant LoRA.

Dataset image (left) and Qwen Image with LoRA generation (right)

Not only does it do a great job with headshots, but it can preserve identity with more complex prompts as well.

Even with small faces Qwen Image preserves the identity using a personalized LoRA

Custom Styles

Although Qwen Image can generate an absurd amount of styles out of the box, sometimes it is necessary to make a custom style LoRA to augment its abilities. This is especially true of obscure styles that don't have an readily available name one can use when prompting.

Here I have trained 1980s luxury interior design LoRA. You can see, Qwen Image does a pretty good job without the LoRA, but the LoRA makes it washed out, adds film grain and cigarette smoke which gives it that authentic 80s feel.

Qwen Image (left) and Qwen Image with custom 80s interior design LoRA (right)

Open source is Cooking

Qwen Image is incredible model for text to image. We at fal can't wait to see what the community comes up with using it. Try out the inference and training with the links below.

Inference: https://fal.ai/models/fal-ai/qwen-image

Training: https://fal.ai/models/fal-ai/qwen-image-trainer