QwenLM/Qwen-Image-2.1

Qwen‑Image‑2.1 – A 7 B‑parameter text‑to‑image & editing model

What it does
Qwen‑Image‑2.1 is a diffusion‑based generative model that can create images from a textual prompt, edit existing images, or generate images with a transparent (RGBA) background – all with a single checkpoint. The visual generation core has 7 B parameters (32 DiT layers) and runs efficiently on modern GPUs.

Key capabilities

Feature	Details
Text‑to‑image	Produce 2 K‑resolution images (e.g., 2048×2048) in a variety of aspect ratios.
Image editing	Accept up to 10 reference images; edit via prompts, circles, painted strokes, or masks while preserving subject identity.
Transparent output	Native RGBA generation; useful for stickers, UI assets, etc.
Prompt rewriting	Optional Qwen‑3.5‑VL‑based models expand short prompts into detailed descriptions for higher quality results.
Efficiency tricks	Mixed‑granularity attention, prefix KV‑cache reuse, and support for FP8 quantisation, CUDA‑graph decoding, and tensor‑parallel inference via vLLM‑Omni, SGLang, LightX2V.

How to use it

1. Install the required packages (PyTorch ≥ 2.4, Transformers ≥ 5.17, Diffusers, Accelerate, Pillow).
2. Load the model with the QwenImage21Pipeline from Hugging Face’s diffusers library:

   from diffusers import QwenImage21Pipeline
   pipe = QwenImage21Pipeline.from_pretrained(
       "Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
   ).to("cuda")
   

3. Call pipe with a prompt (and optionally an input image or a list of images). The API mirrors the standard Diffusers pipelines, so existing code for Stable Diffusion works unchanged.
4. For low‑memory GPUs you can enable CPU off‑loading (pipe.enable_model_cpu_offload()).
5. For large‑scale serving, run the model with vLLM‑Omni or SGLang‑Diffusion, which exploit the model’s prefix‑KV cache and provide FP8/TP parallelism.

Ecosystem integration

• Hugging Face / ModelScope – weights and demo spaces are hosted on both platforms.
• ComfyUI – ready‑made workflow JSON files for text‑to‑image and editing.
• Diffusers – Day‑0 support via QwenImage21Pipeline.
• vLLM‑Omni / SGLang / LightX2V – high‑throughput serving and GPU‑optimised inference.
• Prompt‑rewrite models – separate checkpoints (Qwen‑Image‑2.1‑PE‑T2I and ‑PE‑I2I) that can be run locally or via vLLM to automatically expand prompts and suggest aspect ratios.

Hardware notes

• Works on NVIDIA GPUs (CUDA) and, with ROCm, on AMD Radeon GPUs.
• Supports FP8 quantisation and tensor‑parallelism for multi‑GPU clusters.
• FlagOS is mentioned as a way to run the model on a broader range of AI accelerators.

When to pick Qwen‑Image‑2.1

• You need a single model that handles both generation and editing.
• Transparent‑background output is required.
• You prefer a relatively small (7 B) visual backbone that still delivers 2 K quality.
• You plan to serve many requests and want to leverage prefix‑KV caching for speed.

Resources

• Model cards on HuggingFace and ModelScope
• Blog post and demo space for quick visual tests
• Discord & WeChat community for support
• Detailed recipes for vLLM‑Omni, SGLang, LightX2V in the README.