Run Qwen-Image 2.1 locally on an RTX 5090 or RTX PRO 6000

Qwen-Image 2.1 is much smaller than the 20B original, and better at the things that matter — clean small text, native 2K and real alpha channels. One checkpoint does text-to-image and editing. Both cards run it at full precision.

Generator

7B · single-stream DiT

Text encoder

Qwen3-VL 8B

Output

native 2K · RGBA

Licence

Qwen research licence

Pick your GPU

Choose once — every block below follows your pick.

Install ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv .venv && source .venv/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt "huggingface_hub[cli]"

Download the models

Run from the ComfyUI folder. The repo mirrors ComfyUI's layout, so files land straight in models/.

hf download Comfy-Org/Qwen-Image-2.1 \
  diffusion_models/qwen_image_2.1_bf16.safetensors \
  text_encoders/qwen3vl_8b_int8_convrot.safetensors \
  vae/qwen_image_2.1_vae_bf16.safetensors \
  --local-dir models

Launch

python main.py

Open http://127.0.0.1:8188.

Generate

Workflow → Browse Templates → Image. There are three official templates:

• Qwen Image 2.1: Text to Image — start here.
• Qwen Image 2.1: Image Edit — change a photo using two reference images.
• Remove Background: Qwen Image 2.1 — cut-outs with real transparency.

The templates ship with the INT8 files. Set Load Diffusion Model to qwen_image_2.1_bf16, and on the PRO 6000 set Load CLIP to qwen3vl_8b_bf16.

Defaults are 25 steps, CFG 1, euler / simple. To get native 2K, set Resolution Selector to 4.0 MP (2048 × 2048).

Editing: name each image by its tag, <image1> then <image2>, in the order they're connected. The output takes the first image's aspect ratio.

Keep the person in <image1> exactly — face, pose, lighting.
Replace the shirt with the denim shirt from <image2>.
Leave the background untouched.

My results

Same prompt and seed on both cards — filling this in as I test.

Sources

• ComfyUI · Qwen-Image 2.1 native workflows
• Comfy blog · Qwen-Image 2.1 in ComfyUI
• Comfy-Org/Qwen-Image-2.1 files