Qwen-Image 2.1

Generate and edit images with Qwen-Image 2.1 in PhyAI.

Use Qwen-Image 2.1 to create an image from a text prompt or edit an existing image.

Prepare the model

Download the model weights from Qwen/Qwen-Image-2.1 on Hugging Face.

You need Python 3.12 or newer, uv, and NVIDIA GPUs with CUDA support. Run the following commands from the PhyAI repository root. Replace the checkpoint path with the folder where you saved the model:

uv sync
export QWEN_IMAGE_CHECKPOINT=/path/to/Qwen-Image-2.1

Generate an image

This example uses two GPUs and saves a 512 x 512 PNG. Check available GPUs with nvidia-smi, then adjust CUDA_VISIBLE_DEVICES if needed.

CUDA_VISIBLE_DEVICES=0,1 uv run --no-sync python examples/qwen_image_21/run_qwen_image_21.py \
  --checkpoint "$QWEN_IMAGE_CHECKPOINT" \
  --prompt "An orange cat sitting beside a window in warm sunlight" \
  --tp 2 \
  --height 512 \
  --width 512 \
  --steps 40 \
  --seed 42 \
  --out .cache/qwen_image_21.png

Open .cache/qwen_image_21.png to view the result. The script creates the output folder automatically.

Edit an image

Set --image to your input image and describe the change in --prompt:

CUDA_VISIBLE_DEVICES=0,1 uv run --no-sync python examples/qwen_image_21/run_qwen_image_21.py \
  --checkpoint "$QWEN_IMAGE_CHECKPOINT" \
  --image /path/to/input.png \
  --prompt "Change the background to a snowy forest while keeping the subject unchanged" \
  --tp 2 \
  --height 512 \
  --width 512 \
  --steps 40 \
  --seed 42 \
  --out .cache/qwen_image_21_edited.png

Repeat --image to provide multiple reference images, up to 10.

Common options

Option	What it controls
--prompt	The image to create or the edit to make.
--image	An optional reference image for editing.
--height, --width	Output dimensions in pixels. Use multiples of 32.
--steps	Number of generation steps. Default: 40.
--seed	Random seed. Default: 42.
--tp	Number of GPUs to use. Default: 1.
--out	Output image path. Default: .cache/qwen_image_21.png.

For one GPU, use CUDA_VISIBLE_DEVICES=0 and --tp 1. For eight GPUs, use CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 and --tp 8.

To see all available options:

uv run --no-sync python examples/qwen_image_21/run_qwen_image_21.py --help