Qwen-Image ComfyUI Native Workflow Example

Qwen-Image is a 20B parameter MMDiT (Multimodal Diffusion Transformer) model open-sourced under the Apache 2.0 license.

Qwen-Image is the first image generation foundation model released by Alibaba’s Qwen team. It’s a 20B parameter MMDiT (Multimodal Diffusion Transformer) model open-sourced under the Apache 2.0 license. The model has made significant advances in complex text rendering and precise image editing, achieving high-fidelity output for multiple languages including English and Chinese.

Model Highlights:

• Excellent Multilingual Text Rendering: Supports high-precision text generation in multiple languages including English, Chinese, Korean, Japanese, maintaining font details and layout consistency
• Diverse Artistic Styles: From photorealistic scenes to impressionist paintings, from anime aesthetics to minimalist design, fluidly adapting to various creative prompts

Related Links:

• GitHub
• Hugging Face
• ModelScope

Qwen-Image Native Workflow Example

Make sure your ComfyUI is updated.

• Download ComfyUI
• Update Guide

Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.(Desktop version’s update will delay sometime)

If nodes are missing when loading a workflow, possible reasons:

1. Not using the latest ComfyUI version(Nightly version)
2. Using Stable or Desktop version (Latest changes may not be included)
3. Some nodes failed to import at startup

There are three different models used in the workflow attached to this document:

1. Qwen-Image original model fp8_e4m3fn
2. 8-step accelerated version: Qwen-Image original model fp8_e4m3fn with lightx2v 8-step LoRA
3. Distilled version: Qwen-Image distilled model fp8_e4m3fn

VRAM Usage Reference GPU: RTX4090D 24GB

Model Used	VRAM Usage	First Generation	Second Generation
fp8_e4m3fn	86%	≈ 94s	≈ 71s
fp8_e4m3fn with lightx2v 8-step LoRA	86%	≈ 55s	≈ 34s
Distilled fp8_e4m3fn	86%	≈ 69s	≈ 36s

1. Workflow File

After updating ComfyUI, you can find the workflow file in the templates, or drag the workflow below into ComfyUI to load it.

Download Workflow for Qwen-Image Official Model

Distilled version

Download Workflow for Distilled Model

2. Model Download

Available Models in ComfyUI

• Qwen-Image_bf16 (40.9 GB)
• Qwen-Image_fp8 (20.4 GB)
• Distilled versions (non-official, requires only 15 steps)

All models are available at Huggingface and Modelscope

Diffusion model

• qwen_image_fp8_e4m3fn.safetensors

Qwen_image_distill

• qwen_image_distill_full_fp8_e4m3fn.safetensors
• qwen_image_distill_full_bf16.safetensors

• The original author of the distilled version recommends using 15 steps with cfg 1.0.
• According to tests, this distilled version also performs well at 10 steps with cfg 1.0. You can choose either euler or res_multistep based on the type of image you want.

LoRA

• Qwen-Image-Lightning-8steps-V1.0.safetensors

Text encoder

• qwen_2.5_vl_7b_fp8_scaled.safetensors

VAE

qwen_image_vae.safetensors

Model Storage Location

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   ├── qwen_image_fp8_e4m3fn.safetensors
│   │   └── qwen_image_distill_full_fp8_e4m3fn.safetensors ## 蒸馏版
│   ├── 📂 loras/
│   │   └── Qwen-Image-Lightning-8steps-V1.0.safetensors   ## 8步加速 LoRA 模型
│   ├── 📂 vae/
│   │   └── qwen_image_vae.safetensors
│   └── 📂 text_encoders/
│       └── qwen_2.5_vl_7b_fp8_scaled.safetensors

3. Complete the Workflow Step by Step

1. Make sure the Load Diffusion Model node has loaded qwen_image_fp8_e4m3fn.safetensors
2. Make sure the Load CLIP node has loaded qwen_2.5_vl_7b_fp8_scaled.safetensors
3. Make sure the Load VAE node has loaded qwen_image_vae.safetensors
4. Make sure the EmptySD3LatentImage node is set with the correct image dimensions
5. Set your prompt in the CLIP Text Encoder node; currently, it supports at least English, Chinese, Korean, Japanese, Italian, etc.
6. If you want to enable the 8-step acceleration LoRA by lightx2v, select the node and use Ctrl + B to enable it, and modify the Ksampler settings as described in step 8
7. Click the Queue button, or use the shortcut Ctrl(cmd) + Enter to run the workflow
8. For different model versions and workflows, adjust the KSampler parameters accordingly

The distilled model and the 8-step acceleration LoRA by lightx2v do not seem to be compatible for simultaneous use. You can experiment with different combinations to verify if they can be used together.