Qwen-Image-2.1-Turbo: 8-Step Open Image Model
Alibaba's Qwen team released Qwen-Image-2.1-Turbo, which the model card describes as an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps: it uses the same 7B visual generation architecture and loads directly with QwenImage21Pipeline in Diffusers, ships with its recommended sampling schedule so it works out of the box, generates with CFG=1 by default and uses prefix KV caching to reuse text and reference-image context across denoising steps
Tool Overview
Features, steps and FAQ below
Features
- ✓ Eight steps to an image: the model card says this accelerated checkpoint for text-to-image generation and image editing collapses sampling to 8 denoising steps and includes its recommended sampling schedule, so you do not have to configure a scheduler by hand
- ✓ Same architecture as the base model: the model card says it uses the same 7B visual generation architecture as Qwen-Image-2.1, so it loads directly with QwenImage21Pipeline in Diffusers rather than needing a different pipeline
- ✓ CFG=1 by default, with prefix KV caching: generation uses CFG=1 by default, and prefix KV caching reuses the text and reference-image context across denoising steps to cut repeated work
- ✓ Text-to-image and image editing: the model card says it covers both text-to-image and image editing, including multi-reference composition such as combining several input images into one interior render
- ✓ Documented dependencies and license: the model card says it needs a Diffusers build supporting pipeline-configured sampling sigmas (added in PR #14950), is licensed under the Qwen Research License Agreement, and has weights on both Hugging Face and ModelScope
How to Use
- Install a CUDA-compatible PyTorch build and the dependencies listed on the model card: pip install git+https://github.com/huggingface/diffusers.git plus transformers, accelerate and pillow
- Load Qwen/Qwen-Image-2.1-Turbo with QwenImage21Pipeline.from_pretrained using bfloat16
- Generate from a prompt, or use reference images for editing and multi-image composition
- Use the weights under the Qwen Research License Agreement, reading the license text before any commercial use
FAQ
What is Qwen-Image-2.1-Turbo?
The model card describes it as an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with 8 denoising steps, using the same 7B visual generation architecture.
How is it different from Qwen-Image-2.1?
The model card says the architecture is the same and Turbo is the accelerated checkpoint for 8-step inference, loaded the same way with QwenImage21Pipeline in Diffusers.
How many steps, and do I tune CFG?
The model card says 8 denoising steps, with its recommended sampling schedule included and CFG=1 used by default.
How do I load it?
The model card says to load it with QwenImage21Pipeline.from_pretrained in bfloat16, which requires a Diffusers build that supports pipeline-configured sampling sigmas.
What is the license?
The model card says it is licensed under the Qwen Research License Agreement, with weights published on Hugging Face and ModelScope — read the license text first.