Fern

Fern is a private Windows image and audio studio for running Flux.2 Klein 9B locally with one shared image model and a hardware-specific OpenVINO device. Intel, NVIDIA, and CPU execution all use the same xgirls INT4 OpenVINO release at xgirls/FLUX.2-klein-9B-ov-int4. NVIDIA execution uses the NVIDIA device from the OpenVINO NVIDIA plugin.

Fern also provides audio threads for Qwen3-TTS with OpenVINO. Audio threads keep speech settings and WAV outputs separate from image threads. The Qwen3-TTS workspace supports preset voices, voice design, and reference-audio voice cloning. Set up installs a pinned speech runtime and downloads the selected ready-made OpenVINO export from xgirls on Hugging Face. New installations do not export source checkpoints. Downloads are verified and can resume after interruption; the first take still compiles the downloaded model on the GPU.

The app supports text-to-image and image-guided generation, a local gallery, repeatable seeds, batch generation, and CPU, Intel GPU, or NVIDIA GPU execution. Prompts, reference images, and generated images stay on the local machine.

Use New image thread (Ctrl+N) to start an image workspace, or New audio thread (Ctrl+Shift+N) for speech. Threads in the sidebar keep their own prompt, settings, reference image, and generated images across app restarts. Select a thread to continue it, or open Library to browse all images. Images appear as large thumbnails in the main workspace.

Install

Download Fern-Setup-*.exe from the project's GitHub Releases page and run it. The installer contains the desktop app, Python, OpenVINO, and the OpenVINO NVIDIA plugin when included in the build environment. On first launch, Fern installs its bundled runtime. Optional Klein 4B/9B models can be downloaded from Preferences; Qwen3-TTS models are downloaded from the audio workspace. On hybrid laptops with both Intel integrated graphics and an NVIDIA GPU, Auto-select gives the NVIDIA GPU priority when the NVIDIA driver reports it. If the NVIDIA plugin is not ready, Fern reports that directly instead of silently switching to the Intel GPU. After the model is available, Fern loads it into the selected device memory before enabling generation; the first runtime load can take a little while. During generation, use Stop Generation to cancel at the next inference step without closing the app.

Requirements:

• Windows 10 or 11, 64-bit
• Approximately 12 GB of free disk space for the shared OpenVINO model
• An internet connection for the first model download
• An Intel GPU supported by OpenVINO, an NVIDIA GPU with a current driver and compatible CUDA libraries, or a CPU fallback
• NVIDIA support requires the OpenVINO NVIDIA plugin from openvino_contrib, built against the same OpenVINO runtime as Fern. The upstream plugin currently documents Ubuntu validation; Windows builds must use its Windows/CMake path.

The installer does not require a separate Python, Node.js, Git, or terminal installation.

Local data and cache management

Open Data in the Fern title bar to review disk usage, clear the compiled generation cache, or reset local creative data. A local-data reset removes generated images and WAV takes, thread/gallery metadata, saved prompts, and preferences while preserving downloaded models and bundled runtimes.

Fern automatically prunes generation-cache files that have not been used for 30 days. The cache is also limited dynamically between 1 GiB and 40 GiB based on drive capacity, with a disk-relative free-space reserve of 5–20 GB. Generated images and WAV takes are never deleted automatically.

Development

This repository is an npm workspace. The Electron app lives in apps/desktop; the Python backend and packaging tools live in scripts.

Prerequisites:

• Node.js 22 or newer
• Python 3.12
• PowerShell 7 (Windows PowerShell 5.1 also works for the provided scripts)

npm install
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade pip
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
npm run dev

npm run dev launches the Electron UI directly from source; it does not need a release build or the bundled runtime archive. Development uses the local .venv and opens the UI even when the model has not been downloaded yet. To enable generation in development, download the model into the models/ directory with the bootstrap command below, then relaunch the app.

The development app expects the model at:

models/flux2-klein-9b-circulus-int4

Download it with the same bootstrap used by the installer:

.\.venv\Scripts\python.exe scripts\bootstrap_install.py `
  --model-dir models\flux2-klein-9b-circulus-int4

To build and install the OpenVINO NVIDIA plugin from the linked upstream source:

.\scripts\install-openvino-nvidia-plugin.ps1

Then prepare the shared model for NVIDIA execution:

.\.venv\Scripts\python.exe scripts\bootstrap_install.py `
  --adapter nvidia `
  --models-root models

The NVIDIA adapter calls OVFlux2KleinPipeline.from_pretrained with device="NVIDIA"; Intel uses device="GPU", and CPU uses device="CPU". Fern also keeps an explicit CPU option in the Device menu.

Useful commands:

npm run typecheck
npm run build
npm run check:python
npm run check

Generated models, virtual environments, images, caches, bundled runtimes, and installers are ignored by Git and must not be committed.

Build the one-click installer

Prepare a clean build environment, keeping model experiment dependencies separate:

python -m venv .venv-release
.\.venv-release\Scripts\python.exe -m pip install -r requirements-release.txt
npm run package

The build uses .venv-release when present, otherwise .venv, and rejects retired Qwen Image dependencies or an Optimum Intel checkout that does not match the locked requirements-release.txt. An explicit environment can be selected with scripts/build-installer.ps1 -SourceVenv <folder>. Build the optional NVIDIA plugin against that same environment before packaging if NVIDIA support is required. The build script creates a self-contained runtime and writes the one-click NSIS installer to apps/desktop/release/. Image and speech model weights are not embedded; they download on demand from Hugging Face.

The runtime builder detects the base Python installation used by the selected environment. You can override it when needed:

.\scripts\bundle-python-runtime.ps1 -PythonHome "C:\Path\To\Python312"

Repository layout

apps/desktop/   Electron + React desktop app
scripts/        Flux.2 Klein 9B backend, setup, and packaging tools
.github/        CI and contribution templates

Security and privacy

The API binds to 127.0.0.1 and accepts generation requests only from the local desktop app. See SECURITY.md for reporting vulnerabilities.

Contributing

See CONTRIBUTING.md. By design, changes that add another model family or checkpoint selector are outside this project's scope.

License

MIT. See LICENSE.

The retired Qwen Image 2.1 attempts, self-contained recreation instructions, and removal record are documented in the experiment report.