Qwen3 VL AutoTagger CLI

Standalone CLI version of Qwen3-VL AutoTagger: generate Adobe Stock-style title + keywords and optionally embed XMP metadata into saved images.

ComfyUI Node Version

Need the same tagging pipeline directly inside ComfyUI? Use the node project:

Colab (Production)

Use the production notebook with one-click Colab launch:

• Colab_T4_CLI_Prod.ipynb: run tagging with XMP enabled by default and output download.

Highlights

• Works without ComfyUI
• Processes a single image or a whole folder
• Saves tagged images with XMP metadata when --write-xmp is enabled
• Writes metadata JSONL report (output_dir/metadata.jsonl by default)
• Supports auto-download of Qwen/Qwen3-VL-8B-Instruct on first run
• Optional 4-bit quantization on CUDA

Screenshots

Embedded XMP Metadata

Adobe Stock Result Example 1

Adobe Stock Result Example 2

Quick Start

1. Clone repository:

git clone https://github.com/ekkonwork/qwen3-vl-autotagger-cli
cd qwen3-vl-autotagger-cli

2. Install dependencies:

pip install -r requirements.txt

3. Install exiftool (required only if --write-xmp is enabled):

   • Linux: python install.py (uses apt-get)
   • macOS: brew install exiftool
   • Windows: choco install exiftool

4. Run:

python -m qwen3_vl_autotagger_cli.cli "C:/images" --recursive --output-dir "C:/images_tagged"

CLI Usage

qwen3-vl-autotagger INPUT_PATH [options]

Main options:

• --recursive: recursively scan folders
• --write-xmp / --no-write-xmp: enable/disable XMP embedding
• --require-exiftool / --no-require-exiftool: fail or skip when exiftool is missing
• --output-dir, --output-format, --file-prefix
• --metadata-jsonl: metadata JSONL output path
• --model-id: HF model id (default Qwen/Qwen3-VL-8B-Instruct)
• --auto-download / --no-auto-download
• --local-model-path: local model folder for offline usage
• --load-in-4bit / --no-load-in-4bit
• --min-pixels, --max-pixels, --allow-resize / --no-allow-resize
• --max-keywords, --attempts, --temperature, --top-p, --repetition-penalty

Examples:

# Folder batch, write XMP + JSONL
python -m qwen3_vl_autotagger_cli.cli "./images" --recursive --output-dir "./outputs"

# Single image, metadata only (no saved output image)
python -m qwen3_vl_autotagger_cli.cli "./images/photo.jpg" --no-write-xmp --metadata-jsonl "./report.jsonl"

# Local model only (no download)
python -m qwen3_vl_autotagger_cli.cli "./images" --no-auto-download --local-model-path "D:/models/Qwen3-VL-8B-Instruct"

Output Behavior

• When --write-xmp is enabled, CLI saves tagged images and embeds XMP metadata.
• Saved filenames are auto-incremented (file_prefix_00000, file_prefix_00001, ...) and existing files are not overwritten.
• Metadata JSONL is written for each processed image (input_path, output_path, title, keywords, json).

Model Download Size

By default (--auto-download), the CLI downloads Qwen/Qwen3-VL-8B-Instruct on first run. The download size is about 17.5 GB of weights in total (roughly 16.3 GiB).

Performance

On a Colab T4, a single image typically takes about 60 seconds to auto-tag (varies with resolution and settings).

Troubleshooting (Environment)

• exiftool may be missing or not available in PATH.
• CUDA/driver/VRAM setup can differ between machines.
• bitsandbytes may fail to install or initialize on some systems.
• If you see Qwen3VLForConditionalGeneration is not available, reinstall dependencies:
  • pip uninstall -y transformers qwen-vl-utils
  • pip install -U git+https://github.com/huggingface/transformers
  • pip install -U qwen-vl-utils accelerate bitsandbytes
• Hugging Face (HF) downloads can be unstable due to network/rate limits.

Support

If this project saves you time, you can support development on Boosty:

• Boosty (donate): https://boosty.to/ekkonwork/donate

Crypto Donations (Telegram Wallet)

Scan this QR code in your wallet app to copy the donation address:

• TON: UQAMPvqduXVWyax325-zqk81rTwNG1bRhCvXPyIs7eeIxEVp
• USDT (TON): UQAMPvqduXVWyax325-zqk81rTwNG1bRhCvXPyIs7eeIxEVp
• Memo/Tag: check the Wallet receive screen before sending.

Hire Me

• English: B2 (text-first communication).
• Hiring (full-time/long-term): prefer written communication; for live calls, Russian-speaking teams are preferred.
• Project work: open to worldwide async collaboration.
• Email: ekkonwork@gmail.com
• Telegram: @Mikhail_ML_ComfyUI
• LinkedIn: https://www.linkedin.com/in/mikhail-kuznetsov-14304433b
• Boosty: https://boosty.to/ekkonwork/donate

License

MIT. See LICENSE.