A web application for extracting image tags and generating prompts using WD14 tagger models. Upload an image and get a list of tags that can be used as prompts for image generation models.

ImageDNA uses the WD14 tagger model to analyze images and extract tags. The tool:

• Extracts tags from images - Analyzes uploaded images using a pre-trained ONNX model to identify tags
• Generates copy-ready prompts - Outputs tags as a comma-separated prompt string ready for use with image generation
• Provides filtering controls - Adjust confidence threshold, exclude specific tags, and customize output formatting
• Offers convenience options - Toggle masterpiece quality tags, switch between underscores and spaces, and consolidate similar tags
• Reads image metadata - Extracts EXIF data and AI generation parameters (positive prompt, negative prompt, sampler settings) from any image

The frontend is built with React and the backend uses Flask with ONNX Runtime for model inference. The WD14 model is automatically downloaded from Hugging Face on first run.

Click the dice icon in the top-left header to switch to the Random Prompt Generator. This tool builds structured prompts from the selected model's tag vocabulary.

• General Tags - Toggle on/off and set how many random general tags to include (1-50). The generator ensures representation from key tag groups (subject count, hair, clothing, background, etc.) before filling the rest randomly.
• Character Tags - Toggle to include a random character tag. When enabled, a Subject Type selector appears to pick the correct subject tag (1girl, 1boy, 1other, or None) to avoid mismatches.
• Consolidate Breasts - Lock a specific breast size instead of letting the generator pick randomly. Choose from flat chest through gigantic breasts.
• Exclude Tags - Comma-separated list of terms to filter out. Exclusions apply live to the current prompt after each term is committed with a comma.
• Structured Ordering - Generated prompts follow conventional AI art prompt order: subject count, solo, nudity, character name, breast size, hair, clothing, background, then everything else.
• Live Controls - Changing the subject type, breast size, or exclude tags instantly updates the current prompt without needing to re-roll.

Click the Generate Prompt button to create a prompt, then use Copy Tags to copy it to clipboard. Click Re-roll to generate a new random prompt.

Click the EXIF tab in the header to switch to the EXIF Extractor. Drop or click to upload any image and the tool will read its embedded metadata.

• Generation Parameters - For AI-generated images (Stable Diffusion, ComfyUI, etc.), the embedded parameters string is split into three sections:
  • Positive - The positive prompt, highlighted in green with its own copy button
  • Negative - The negative prompt, highlighted in red with its own copy button
  • Settings - Sampler, steps, CFG scale, seed, and other generation settings
  • Full Parameters - The raw unmodified string with a "Copy All" button
• Image Text Data - Other PNG text chunks are shown in a collapsible section
• EXIF Metadata - Camera make/model, lens, exposure settings, ISO, date/time, and GPS for photos taken on real cameras
• Drag to replace - With an image already loaded, drag a new image directly onto the preview to replace it instantly

Click the gear icon in the top-right corner of the application to open the Settings menu. The following options are available:

• Model Selection - Choose which tagger model to use for analysis:

  • EVA02 Large v3 (default) - Best accuracy
  • MOAT v2 - Good balance of speed and accuracy
  • SwinV2 v3 - Fast and efficient

  Changing the model while an image is loaded will automatically re-run the tagging with the new model.

• DeviantArt Mode - When enabled, formats tags specifically for DeviantArt submissions:

  • Converts tags to lowercase
  • Removes spaces and replaces hyphens with underscores
  • Provides a separate "Copy DA Tags" button

• DA Tag Limit (visible when DeviantArt Mode is enabled) - A slider to set the maximum number of tags to copy (5-30). DeviantArt has a limit of 30 tags per submission.

You can use the pre-build docker image hosted on Docker Hub if you want to just run the application and don't require the development files

Prerequisites: Docker

docker run -d -p 5000:5000 nuclear314/image-dna:latest

A packaged Windows build (ImageDNA.exe) runs standalone — no Python, Node, or Docker needed. Grab a build from CI (see How to build the Windows standalone app) or a release if one's been published, then just run the exe.

Prerequisites:

• Visual C++ Redistributable 2015–2022 (x64), v14.29+ — onnxruntime requires the VC++ 2019 runtime or later; the linked installer always ships the latest compatible version. Missing it surfaces as ImportError: DLL load failed while importing onnxruntime_pybind11_state when the app starts.
• .NET Desktop Runtime — required by pywebview's pythonnet bridge on Windows. Already preinstalled on Windows 10/11; only relevant on older or minimal Windows installs.
• WebView2 Runtime — renders the app window; preinstalled on Windows 11 and auto-delivered via Windows Update on Windows 10. If missing, install the Evergreen Bootstrapper.

First run requires internet access to download the tagger model from Hugging Face into %APPDATA%\ImageDNA\models; it's cached there afterward, so subsequent launches work offline.

Prerequisites: Node.js 24+ and Python 3.12+

1. Install frontend dependencies:

   npm install

2. Install Python dependencies:

   pip install -r requirements.txt

3. Build the frontend:

   npm run build

4. Start the server:

   python server.py

5. Open http://localhost:5000 in your browser

For development with hot reload, run the Vite dev server (npm run dev) and the Flask server separately, then access the Vite dev server URL.

The Dockerfile uses a multi-stage build to create a production-ready image:

1. Build the image:

   docker build -t imagedna .

2. Run the container:

   docker run -p 5000:5000 imagedna

3. Open http://localhost:5000 in your browser

The container exposes port 5000 and will download the WD14 model from Hugging Face on first startup.

By default the container re-downloads the tagger model every time it's recreated. Point HF_HOME at a named volume so the model survives across docker run recreations:

docker run -p 5000:5000 -e HF_HOME=/cache -v imagedna-cache:/cache imagedna

windows/ contains a PyInstaller-based launcher that bundles an embedded Python server and opens the app in a native window (via pywebview), so end users don't need Python, Node, or Docker installed.

Prerequisites:

• Windows
• Python 3.12+
• Node.js 24+
• A virtual environment at the repo root (.venv) with windows/requirements-windows.txt installed — build.bat activates it automatically if present

cd windows
build.bat

This runs a full build: compiles the frontend, builds the PyInstaller launcher, and sets up the embedded Python server runtime (downloading the Python embeddable package once, cached under windows\cache\).

For fast iteration on just the launcher (windows/main.py or imagedna.spec), skip re-provisioning the embedded server runtime:

build.bat --skip-server

This only works if a full build has already populated release\ImageDNA\server\ — it reuses that directory instead of rebuilding it. Note that --skip-server will not pick up changes to requirements.txt (or other server-side changes), since it skips the step that reinstalls those packages into the embedded runtime — run a full build.bat after touching requirements.txt.

The output lands in release\ImageDNA\. Launch the app with release\ImageDNA\ImageDNA.exe.

First run requires internet access to download the tagger model from Hugging Face into %APPDATA%\ImageDNA\models; it's cached there afterward, so subsequent launches work offline. See How to run locally (Windows) for the runtime prerequisites end users need.