Qwen-Image-2512

Qwen-Image-2512

Generates images from text prompts in English or Chinese, with better human realism, natural detail and text in pictures than the August Qwen-Image. For developers and designers using diffusers.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen-Image-2512 is the December update of the Qwen-Image text-to-image model, with about 20.4B parameters. You write a prompt, and it draws a picture, including pictures that contain written text.

Compared with the August release, the card lists three gains. People look less artificial, scenery and animal fur show finer detail, and text inside images comes out clearer with better layout.

Who it suits: It suits developers, designers and creators who want an open image model they can run themselves, and who often need legible text or realistic people in their pictures.

Industry | Who uses it | How they use it | Example
Design | Visual designers | Draft posters and covers with readable text | Make a cover with a title
Marketing | Content teams | Produce pictures from short briefs | Create a product scene
Software | Developers | Add image generation to an app | Call the pipeline from code
Research | ML engineers | Compare open image models | Test prompts in English and Chinese

If you do not have a large GPU or do not want to run a model yourself, you can skip it, since the full weights are about 40.9 GB in BF16.

Where to use it: The model runs through the diffusers library in Python on Windows, macOS or Linux. The card shows a CUDA GPU as the main path, with a CPU fallback.

• Python scripts that use diffusers
• A CUDA GPU for reasonable speed
• Seven set aspect ratios, from square to 16:9

Use case | Who | What they get back
Text to image | Creators | A picture drawn from your prompt
Pictures with text | Designers | Images with clearer lettering and layout
Realistic people | Marketing teams | Portraits with finer faces and settings

How to set it up: You need Python with pip, PyTorch and a recent diffusers, which the card says to install from its source repository on GitHub. Enough disk for the weights is also needed.

• Python with pip
• PyTorch and the latest diffusers
• About 40.9 GB of disk space for BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files, or let diffusers fetch them.
3. Install the latest diffusers as the card describes.

pip install -U huggingface_hub
hf download Qwen/Qwen-Image-2512

How to use it: You load the pipeline, then call it with a prompt, a negative prompt, a width and height, and a step count. The card uses 50 steps and a true_cfg_scale of 4.0.

from diffusers import DiffusionPipeline
import torch

model_name = "Qwen/Qwen-Image-2512"

# Load the pipeline
if torch.cuda.is_available():
    torch_dtype = torch.bfloat16
    device = "cuda"
else:
    torch_dtype = torch.float32
    device = "cpu"

pipe = DiffusionPipeline.from_pretrained(model_name, torch_dtype=torch_dtype).to(device)

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=width,
    height=height,
    num_inference_steps=50,
    true_cfg_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(42)
).images[0]

image.save("example.png")

Aspect ratio | Width and height
1:1 | 1328 by 1328
16:9 | 1664 by 928
9:16 | 928 by 1664
4:3 | 1472 by 1104
3:4 | 1104 by 1472
3:2 | 1584 by 1056
2:3 | 1056 by 1584

Example: You write a prompt for a street scene with a shop sign, pick the 16:9 size and run the pipeline. It returns an image, and you save it as example.png.

For newcomers and experts: If you are new, run the sample with a short prompt at the square size and look at the saved picture. Then try a prompt that asks for a sign with a few words on it.

If you know the field, the pipeline takes a negative prompt and a true CFG scale. The card's example uses a Chinese negative prompt that lists low quality, deformed hands and an AI look.

Load the diffusers pipeline
Write a prompt and a negative prompt
Choose a size and step count
Save the generated image

Good to know: The card says over 10,000 blind rounds on AI Arena put it first among open-source models and competitive with closed ones. That is the maker's claim, so test it on your own prompts.

• License: apache-2.0
• Cost to run: none named, but it needs a strong GPU for comfort
• Data it reaches: files you download, and the prompts you write
• Not for: low-memory machines, since the BF16 weights are about 40.9 GB

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 20.4B
• Weights: about 40.9 GB in BF16; about 10.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 37,650
• Likes: 991
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 26
• Fine-tunes: 43
• Tags: diffusers, safetensors, text-to-image

Links

• Model: https://huggingface.co/Qwen/Qwen-Image-2512
• Files: https://huggingface.co/Qwen/Qwen-Image-2512/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen-Image-2512

Then follow the usage in its model card. Nothing in this kit runs by itself.