QwenLM/Qwen-Image-2.1

Qwen‑Image‑2.1 – Open‑source Text‑to‑Image & Image‑Editing Model

What it is – A 7 B‑parameter diffusion model (DiT‑style) that can both generate images from a text prompt and edit existing images, including transparent (RGBA) outputs. It is part of the Qwen family and is released under an open‑source licence.

Key strengths

• Compact & fast – Mixed‑granularity attention and prefix KV‑cache reuse give good visual quality with low compute. The Turbo variant runs in just 8 denoising steps.
• Unified generation & editing – Same model handles pure text‑to‑image, transparent‑layer creation, layer‑wise editing, and subject extraction.
• Rich editing interface – Accepts up to 10 reference images, local edit masks (circles, painted strokes, external masks) and preserves identity for people/products.
• Better aesthetics – Improvements in typography, lighting, and fine‑detail rendering.

How to try it

# Install the required libraries
pip install "torch>=2.4.0" "transformers>=5.17" \
    git+https://github.com/huggingface/diffusers accelerate pillow

import torch
from diffusers import QwenImage21Pipeline

# Standard model (high quality, more steps)
pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
).to("cuda")

# Turbo model (8‑step fast generation/editing)
pipe_turbo = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16
).to("cuda")

The README includes a full example that turns a sketch of a motor‑yacht into a photorealistic rendering using the Turbo checkpoint.

Ecosystem support – Day‑0 integrations are provided for:

• 🤗 Diffusers (QwenImage21Pipeline)
• ComfyUI (weights under Comfy-Org/Qwen-Image-2.1 with ready‑made workflow JSONs)
• vLLM‑Omni, SGLang, LightX2V, and OpenVINO for high‑performance inference, KV‑caching, CUDA‑graph execution, FP8 quantisation, and multi‑GPU parallelism.

Hosted services – Alibaba Cloud Model Studio offers paid Pro and Turbo APIs for easy integration without running the model locally.

Typical use cases

• Rapid creation of marketing or educational graphics with custom typography and layout.
• Editing product photos (changing colours, adding/removing objects) while keeping the original identity.
• Generating transparent assets (icons, UI elements) directly from prompts.
• Batch‑style transfer or subject extraction for downstream pipelines.

Where to find more

• Model cards & weights: HuggingFace / ModelScope
• Blog announcement: https://qwen.ai/blog?id=qwen-image-2.1
• Demo Space: https://huggingface.co/spaces/Qwen/Qwen-Image-2.1
• Community: Discord link in the README, plus a WeChat QR code.

---

All details are taken directly from the repository’s README; no additional features are inferred.