Qwen-Image 4 Step ComfyUI Workflow & One Click Windows Installer

Qwen Image 4-Step Lightning LoRA has just dropped, enabling you to generate images at much faster speeds than ever before. I'm sharing a custom workflow and a one-click installer that support both text-to-image and image-to-image generation that can be run on low VRAM devices as well. Plus, I've included an extra node for inpainting—this feature is still experimental but shows promising results for certain creative tasks.

About the Qwen Image 20B Model

The new Qwen-Image model is a cutting-edge 20-billion-parameter Multimodal Diffusion Transformer (MMDiT) from Alibaba Cloud. It stands out for its ability to render complex text in images—including multilingual and paragraph-style layouts—with state-of-the-art fidelity. The model supports both image generation and advanced editing, rivaling closed-source giants like GPT-4o in English and leading its class in Chinese performance. It’s open source and designed to work efficiently even on consumer-level hardware, making it accessible to a wide community of creators.

Qwen Image 4-Step Lightning by Lightx2v

With the 4-Step Lightning LoRA (by lightx2v) workflow, you can produce high-quality results with dramatically faster rendering times—often under a minute for high-res images, even on mid-range hardware. The Lightning LoRA achieves this speed by reducing generation steps to just four, with minimal compromise to output quality for quick drafts or workflow iteration. When you need ultimate quality, simply switch back to higher step settings for your final render. This new release empowers both rapid experimentation and production-grade results—all within ComfyUI or your favorite workflow.

Preloaded Models within the Installer (Low VRAM)

• qwen-image-Q3_K_S.gguf (ComfyUI\models\unet)
  Downloaded from: city96/Qwen-Image-gguf

• Qwen2.5-VL-7B-Instruct-Q4_0.gguf (ComfyUI\models\clip)
  Downloaded from: unsloth/Qwen2.5-VL-7B-Instruct-GGUF

• qwen_image_vae.safetensors (ComfyUI\models\vae)
  Downloaded from: Comfy-Org/Qwen-Image_ComfyUI

• Qwen-Image-Lightning-4steps-V1.0.safetensors (ComfyUI\models\loras)
  Downloaded from: lightx2v/Qwen-Image-Lightning

• 2xLexicaRRDBNet_Sharp.pth (ComfyUI\models\upscale_models)
  Downloaded from: Thelocallab/2xLexicaRRDBNet_Sharp

The standard Qwen Image 20B diffusion models (FP16 and FP8) are not packaged with the installer, but you can download them directly from the Comfy Org Hugging Face repository and add them to your ComfyUI/models/diffusion_models folder:
Comfy Org Qwen Image Diffusion Models

Speed:
Generate 768 x 768 resolution images in a minute and 20 seconds on an RTX 4050 with 6GB VRAM. Faster performance is possible on higher-end GPUs.

System Requirements:

• Nvidia RTX 30XX, 40XX, or 50XX series GPU (FP16 support required; GTX 10XX/20XX not tested)

• CUDA-compatible GPU with at least 4–6 GB VRAM

• Windows OS

• At least 40 GB free storage

What’s Included:

• Portable ComfyUI Windows Installer, pre-configured for Qwen Image Text to Image and Image to Image /w Inpainting

• Custom workflow supporting text-to-image and Image to Image Generations

• Automatic downloads for all required nodes and models

   

Usage Notes:
For inpainting, right click on your loaded image in the load image node and click "open mask editor" to create a mask around the objects you want to change. Then your positive prompt to detail the change you want to make in the new image. When inpainting, you may want to play with the denoise value in the Basic scheduler node to maintain some form of consistency until controlnets are released for the model.

Support and More Information

• Community Support: For troubleshooting or to connect with other users, join the Discord server.