Qwen-Image
Qwen-Image by Alibaba's Qwen team: a large image model that is especially good at text inside pictures, with editing versions.
What it is
Qwen-Image (2025) is a 20-billion-parameter model that renders signs, posters and long text in English and Chinese well and follows detailed prompts. Qwen-Image-Edit (and its later versions) edits a picture by instruction, can change text in it and can combine several reference pictures.
Files and nodes
Load Diffusion Model with the Qwen-Image file (fp8 versions are about 20 GB); Load CLIP with qwen_2.5_vl_7b (type qwen_image); Load VAE with qwen_image_vae; EmptySD3LatentImage; ModelSamplingAuraFlow sets the shift in the templates. Editing graphs use TextEncodeQwenImageEdit (or the "Plus" version) with the input picture.
Settings that work
About 1.6 megapixels (1328×1328, 1664×928); 20 to 50 steps, cfg about 2.5 to 4 (with Lightning LoRAs: 4 or 8 steps, cfg 1); euler with simple. Prompt: sentences; put text to render in quotes.
What fits
Only Qwen-Image LoRAs and ControlNets; edit models have their own LoRAs. Its VAE and encoder fit nothing else.
How the app helps
Shown as Qwen-Image on model cards. The lesson Qwen-Image: words inside the picture builds the picture graph; Editing a picture by words builds an editing graph.