What is Qwen-Image?

Qwen-Image is an image generation model developed by the Qwen team of Alibaba, China.

It uses the multimodal LLM Qwen2.5-VL for the text encoder, and its ability to understand prompts is one step ahead compared to models using T5 or Gemma.

Furthermore, unlike Flux.1 dev, the base is not a distilled model, so it is easy to train, and along with its sister model Qwen-Image-Edit (available in Japanese), it is characterized by a rich surrounding ecosystem such as LoRA and Lightning systems.

Recommended Resolution

Qwen-Image recommends around 1.5M to 1.8M pixels.

  • 1:1 ... 1328 × 1328
  • 4:3 ... 1472 × 1104
  • 3:2 ... 1584 × 1056
  • 16:9 ... 1664 × 928

Model Download

📂ComfyUI/
└── 📂models/
    ├── 📂diffusion_models/
    │   └── qwen_image_fp8_e4m3fn.safetensors
    ├── 📂text_encoders/
    │   ├── qwen_2.5_vl_7b_fp8_scaled.safetensors
    │   └── Qwen2.5-VL-7B.gguf    ← Only when using gguf
    ├── 📂unet/
    │   └── qwen-image.gguf       ← Only when using gguf
    └── 📂vae/
        └── qwen_image_vae.safetensors

text2image

qwen_image.json
  • euler or res_multistep is often used for the sampler.

ControlNet (InstantX)

There are several ControlNet models for Qwen-Image, but I will introduce the one based on InstantX provided as ControlNet-Union because it is easy to use.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂controlnet/
        ├── Qwen-Image-InstantX-ControlNet-Union.safetensors
        └── Qwen-Image-InstantX-ControlNet-Inpainting.safetensors

workflow

qwen_image_instantx_controlnet_union.json
qwen_image_instantx_controlnet_inpainting.json

Lightning (High Speed Generation LoRA)

Qwen-Image-Lightning is a LoRA set distilled so that Qwen-Image can be run in 4 / 8 steps.

Since it significantly reduces the number of steps with almost no degradation, it is adopted in quite a few workflows.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂loras/
        ├── Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors
        └── Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors

workflow

qwen_image_lightning_8steps.json
  • Load Lightning LoRA with some nodes such as LoraLoaderModelOnly.
  • Set steps of KSampler to 4 or 8, and CFG to 1.0.