Ready-to-use ComfyUI workflow for Qwen-Image 2.1 Turbo, Qwen's official 8-step model.

What you need

• Turbo model: qwen_image_2.1_turbo_int8_convrot.safetensors (7.26 GB) in models/diffusion_models

• Text encoder: qwen3vl_8b_int8_convrot.safetensors in models/text_encoders

• VAE: qwen_image_2.1_vae_bf16.safetensors in models/vae

• ComfyUI 0.37.0 or newer

All three files are on Hugging Face in Comfy-Org/Qwen-Image-2.1. If you already use Qwen-Image 2.1, you only need the Turbo model.

How to use

1. Drag the workflow into ComfyUI.

2. Write your prompt and pick a size.

3. Run

Settings

Everything is already set: 8 steps, cfg 1, euler, with the sigma schedule Turbo was trained on (in a ManualSigmas node). Leave these as they are. The negative prompt has no effect at cfg 1.

Speed

Time per 1024 x 1024 image, Turbo vs the regular Qwen-Image 2.1 (25 steps):

• RTX 3060 12 GB : 15.0 s (Turbo) vs 34.9 s (Base)

• RTX 5090 32 GB: 2.0 s (Turbo) vs 4.5 s (Base)

See related video that compares the performance.

Licence

Qwen Research Licence, non-commercial, the same as Qwen-Image 2.1.