What’s new
- A 20B-parameter open image foundation model: it turned heads at its August 2025 launch for complex text rendering, especially accurate Chinese text.
- Continuously updated with dated releases: this isn't a one-and-done launch — the 2512 (December 2025) update noticeably improved human realism and material detail, cutting down the "obviously AI" look.
- A companion Edit line that keeps pace: Qwen-Image-Edit has iterated all the way to Edit-2511, with better multi-image input and stronger subject consistency along the way.
- Fully open under Apache 2.0: weights and code are downloadable, with free commercial use, fine-tuning and self-hosting allowed — this is the open line Nomi reaches through ModelScope, currently pinned to the Qwen-Image-2512 weights released 2025-12-31.
What it’s good for
- Self-hosted, free-to-commercialize image generation when you don't want to depend on a cloud API key.
- Posters, covers and UI work that need accurate Chinese text rendering.
- Developers and hobbyists who want an open base model to fine-tune or attach LoRAs to.
Prompting tips
- Append quality boosters to your positive prompt: the official examples recommend "Ultra HD, 4K, cinematic composition" in English (or the Chinese equivalent for Chinese prompts).
- Write real negative prompts: the official example lists specifics like "low resolution, low quality, malformed limbs, malformed fingers, oversaturated" — adapting that list beats leaving it blank.
- Use prompt rewriting for edits: Qwen recommends running edit instructions through its Prompt Enhancement Tool first, which noticeably improves editing stability.
- Write out any text you want in the image: accurate text rendering is one of its strongest features, so spell it out rather than leaving it vague.
A modern minimalist poster for a bookstore, featuring a stack of open books and a cup of coffee,
warm yellow tones. Headline text "Midnight Bookstore · Open 24 Hours", subhead
"One coffee, one book, one quiet night." Ultra HD, 4K, cinematic composition.
Negative prompt: low resolution, low quality, blurry text, distorted, oversaturated.
Known limits
- It's a separate product line from the newer Qwen-Image 3.0: 3.0 currently has no public weight information and is only usable through official or reseller APIs, while this open line (including the 2512 build Nomi uses) ships downloadable weights for self-hosting — the two run in parallel rather than one simply superseding the other.
- Qwen acknowledged "performance misalignments" on 2025-08-19 and recommended updating to the latest diffusers commit to get the expected results.
- At 20B parameters it's not light on local hardware, which is why the community keeps shipping acceleration options like the LightX2V distillation project.
Using it in Nomi
On the generation canvas, add a image card, pick “Qwen-Image” in the model picker and choose a mode (Text to image). Add your references, write the prompt and generate. In the storyboard you can pick it per shot. The first time, connect any one of the providers above under Settings → Models.
