Qwen-Image-2512: The Text-to-Image Model That Renders Real Text

Qwen-Image-2512 is a 20 billion parameter text-to-image diffusion model from Alibaba's Qwen team, and its main draw is rendering legible text inside generated images in both English and Chinese, a task where SDXL and Flux checkpoints covered elsewhere on this site typically produce garbled letters. Install it through Diffusers with a git build, run it on CUDA or CPU, and license it under Apache 2.0 with no revenue cap. The model card does not publish a VRAM figure, so test on your own GPU before committing.

Written by Priya Raghunathan, local-AI hardware reviewer. I test diffusion and language models against their published model cards and GitHub repos before recommending install paths, because marketing copy and actual VRAM needs frequently disagree.

Why this model is different from SDXL and Flux

Text-to-image models have historically been bad at typography. Ask SDXL or Flux for a poster with a specific slogan and you usually get a plausible-looking jumble of letters instead of the actual words. Qwen-Image-2512's model card calls out multilingual text rendering, readable typography in both English and Chinese, as a headline feature, and independent blind evaluation on AI Arena (over 10,000 rounds, per the model's release notes) ranked it as the strongest open-source image model in that testing round. December 2025's 2512 update layered on top of that with improved human realism and finer natural textures in landscapes, fur, and water, on top of the text-rendering strength the base Qwen-Image model already had.

Hardware: no VRAM number is published, and 20B is heavy

If full bfloat16 does not fit your card, two options are worth checking rather than guessing at a number: community quantized builds, and the offload path documented in Qwen's own GitHub repo. DiffSynth-Studio, listed in the repo's community support section, documents layer-by-layer offloading that runs Qwen-Image inference within 4GB of VRAM plus FP8 quantization support. Check both projects' own documentation for current compatibility with the 2512 checkpoint specifically before relying on either.

Install with Diffusers

Qwen-Image-2512 needs a git build of Diffusers rather than the stable PyPI release, plus a recent transformers version. Both commands below are copied from the official QwenLM/Qwen-Image GitHub repo and Qwen's Hugging Face model card.

from diffusers import DiffusionPipeline
import torch

model_name = "Qwen/Qwen-Image-2512"

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)

Generate an image

The GitHub repo's own example sets a negative prompt, a specific width and height from a table of supported aspect ratios, 50 inference steps, and a true_cfg_scale of 4.0. Those are the values Qwen tested with, so start there before tuning.

prompt = "A neon sign above a coffee shop reading \"Open 24 Hours\", photorealistic, night scene"
negative_prompt = " "

# width, height pairs Qwen tested: 1:1 (1328,1328), 16:9 (1664,928),
# 9:16 (928,1664), 4:3 (1472,1104), 3:4 (1104,1472), 3:2 (1584,1056), 2:3 (1056,1584)
width, height = 1664, 928

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=device).manual_seed(42),
).images[0]

image.save("output.png")

Licensing: Apache 2.0, no revenue threshold

Qwen-Image-2512 ships under Apache 2.0, confirmed on both the GitHub repo and the Hugging Face model card. That means commercial use is allowed outright, with no revenue cap and no separate enterprise tier to contact, unlike Stability AI's Community License on SD 3.5 Medium and Large, which requires an Enterprise License above $1 million in annual revenue.

FAQ

How much VRAM does Qwen-Image-2512 need?

Qwen has not published a specific VRAM figure for this model. It is a 20 billion parameter diffusion model, larger than SDXL or SD 3.5 Medium, so test it on your own GPU before assuming it fits. If it does not fit at full bfloat16 precision, look at DiffSynth-Studio's offload and FP8 quantization support, linked above.

Is Qwen-Image-2512 free for commercial use?

Yes. It is released under Apache 2.0 with no revenue threshold and no contact-required enterprise tier, unlike some competing image models.

What makes Qwen-Image-2512 different from SDXL or Flux?

Its model card and independent blind testing highlight rendering readable, accurate text inside generated images in both English and Chinese, a task SDXL and Flux checkpoints typically handle poorly. The December 2025 update also improved human realism and natural texture detail over the original Qwen-Image release.

Can Qwen-Image-2512 run on CPU?

Yes, the official code path falls back to float32 on CPU when CUDA is unavailable. It will run noticeably slower than on a GPU given the model's 20B parameter count, so treat CPU as a way to test the pipeline rather than a regular workflow.