TagScribeR v2 is a modern, GPU-accelerated local image captioning and dataset management suite. Rebuilt from the ground up using PySide6 and powered by Qwen 3-VL (Vision-Language) models(with optional API support), it offers a "Studio" workflow for preparing AI training datasets.

• 🖼️ Gallery Studio: Multi-select visual grid, instant tagging, and batch caption editing.
• 🤖 Qwen 3-VL Captioning: State-of-the-art vision model integration.
  • GPU Accelerated: Supports NVIDIA (CUDA) and AMD (ROCm) on Windows.
  • Real-time Preview: Watch captions appear as they generate.
  • Custom Prompts: Use templates or natural language (e.g., "Describe the lighting in detail").
  • API Mode: Connect to LM Studio, Ollama, or other API services to use other desired models or offload processing to another machine or the cloud.
• ✏️ Batch Editor: Resize, Crop (with focus points), Rotate, and Convert formats in bulk.
• 📂 Dataset Manager: Create, sort, filter, and organize image collections without duplicating files manually.
• ℹ️ Metadata Editor: View and edit EXIF data, specifically targeting Stable Diffusion generation parameters.

• Python 3.10 or 3.11 installed.
• Git installed.

Clone the repository and run the installer:

git clone https://github.com/ArchAngelAries/TagScribeR.git
cd TagScribeR
install.bat

The installer will automatically detect your hardware:

• NVIDIA RTX 20/30/40: Installs Stable CUDA 12.4.
• NVIDIA RTX 50 (Blackwell): Installs Nightly CUDA 12.8 (cu128).
• AMD Radeon: Scans for your architecture (RX 7000, RX 9000, Strix Halo) and installs the correct ROCm Nightly build.

If the auto-installer fails or you need a specific version, activate the venv (.\venv\Scripts\activate) and run the command for your hardware:

Standard (RTX 30/40):

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Bleeding Edge (RTX 50 Series):

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

Find your architecture below. You must run both commands (SDK + Torch).

RX 7000 Series / 780M (gfx110X):

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ "rocm[libraries,devel]"
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ --pre torch torchvision torchaudio

RX 9000 Series (gfx120X):

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ "rocm[libraries,devel]"
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ --pre torch torchvision torchaudio

Strix Halo (gfx1151):

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ "rocm[libraries,devel]"
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchvision torchaudio

Workstation MI300 (gfx94X):

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx94X-dcgpu/ "rocm[libraries,devel]"
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx94X-dcgpu/ --pre torch torchvision torchaudio

⚠️ Important: Do not run pip install torch afterwards, or it will overwrite the AMD version with the generic CPU version.

TagScribeR works natively on Linux with full GPU acceleration.

1. Run the install script:

   chmod +x install.sh
   ./install.sh

2. If the app fails to launch, you may need the Qt XCB library:

   sudo apt-get install libxcb-cursor0

3. Launch:

   source venv/bin/activate
   python main.py

If you have an RTX 5090 and get an error saying RuntimeError: operator torchvision::nms does not exist or no kernel image available:

This means the PyTorch Nightly server has mismatched versions (a common issue on the bleeding edge).

The Fix (Manual Transplant):

1. If you have Fluxgym, Kohya_SS, or OneTrainer running successfully on your machine, go to that application's venv\Lib\site-packages folder.
2. Copy the folders torch, torchvision, and torchaudio.
3. Paste them into TagScribeR\venv\Lib\site-packages, overwriting files.
4. Run this in the TagScribeR terminal to fix dependencies:

   .\venv\Scripts\activate
   pip install torchgen sympy networkx jinja2 fsspec pyyaml

1. Go to the Auto Caption tab.
2. Download a Model: Select a preset (e.g., Qwen 2.5-VL-3B) and click Download.
3. Load Images: Open a folder containing your dataset.
4. Select Images: Click individual images or "Select All".
5. Run: Click "🚀 Caption Selected".

1. Go to the Datasets tab.
2. Create Collection: Click "New" to create a named folder in Dataset Collections.
3. Filter Source: Load a source folder and type tags (e.g., 1girl, outdoors) to find specific images.
4. Add: Select the images and click "➕ Add Selected to Collection". This copies the images and their text files safely.

• Themes: Choose from various Material Design themes (Dark Teal, Dark Amber, Light Blue, etc.).
• Defaults: Set your preferred AI temperature and token limits.

• GUI Framework: PySide6 & qt-material
• AI Backend: HuggingFace Transformers & Qwen-VL
• AMD Support: ROCm for Windows

Created by ArchAngelAries. Code Assisted by Google's Gemini Pro 3.