Videre is an Electron desktop video editor running fully local without auth/cloud dependencies.

• Tauri runtime removed.
• Frontend moved to a React Router app (/app).
• Authentication removed (local single-user mode).
• Project persistence moved to local filesystem JSON storage.
• Media upload/render is local-only via the bundled Remotion/Express server.

• Projects index: local_data/projects.json
• Project state files: local_data/project_state/<project-id>.json
• Imported media: out/<project-id>/
• Rendered outputs: out/

You can override paths with:

• VIDERE_DATA_DIR
• VIDERE_MEDIA_DIR
• TIMELINE_DIR

• Node.js 20+
• pnpm
• ffmpeg available on your PATH

pnpm install
pnpm desktop:dev

This starts:

• React Router dev server on http://127.0.0.1:5173
• Local render/upload server on http://127.0.0.1:8000
• Electron window loading the app

Shortcut:

./scripts/start-dev.sh

• pnpm dev - React Router dev server
• pnpm render:server - Remotion render/upload server
• pnpm desktop:dev - Full desktop dev stack (web + render + Electron)
• pnpm build - React Router production build
• pnpm preview - Serve production build locally
• pnpm typecheck - Type generation + TypeScript checks
• pnpm lint - ESLint checks

The Captions tab uses a local Python runner with transformers + openai/whisper-small.

Use a Python version supported by your local torch build (Python 3.11/3.12 is recommended).

python3.12 -m venv .venv-whisper
source .venv-whisper/bin/activate
pip install -r app/videorender/requirements-whisper.txt

The render server auto-detects .venv-whisper/bin/python first. If you use a different interpreter path, set VIDERE_WHISPER_PYTHON.

• VIDERE_WHISPER_PYTHON (default: python3)
• VIDERE_WHISPER_MODEL (default: openai/whisper-small)
• VIDERE_WHISPER_DEVICE (default: auto)
• VIDERE_WHISPER_FFMPEG_BIN (default: ffmpeg)

The first transcription request downloads the Whisper model weights and may take noticeably longer than subsequent runs.

Images and videos in assets/ can be encoded with google/siglip2-base-patch16-224 for text-based retrieval. Each image gets one embedding; each video is split into 4–5 sections by time, and each section gets an average embedding (from a few sampled frames), so you can retrieve relevant video segments by text.

The model is downloaded automatically the first time you run the build or retrieval script (via transformers’s from_pretrained("google/siglip2-base-patch16-224")). It is cached under your Hugging Face cache directory (e.g. ~/.cache/huggingface/hub/ on Linux/macOS, or %USERPROFILE%\.cache\huggingface\hub\ on Windows).

To pre-download the model into the default cache (e.g. while online) without running the scripts:

pip install huggingface_hub
huggingface-cli download google/siglip2-base-patch16-224

The model is stored in the default Hugging Face cache, so the build and retrieval scripts will use it automatically. Otherwise, no separate download step is needed: run the build or retrieval script once with internet and the model is downloaded and cached for you.

With the virtual environment activated and dependencies installed, run the build script. It loads the model from the cache (or downloads it if missing), encodes every image and every video section in assets/, and writes the embeddings and index:

# Install dependencies (one-time; opencv-python needed for video frame extraction)
pip install torch "transformers>=4.49" pillow numpy opencv-python

# Encode all images and video sections in assets/ and store embeddings
python scripts/build_image_embeddings.py

This creates:

Commit these two files to the repo so retrieval can use them offline.

Nearest-neighbour search over the stored embeddings:

python scripts/retrieve_by_text.py "food"
python scripts/retrieve_by_text.py "a skateboard" -k 3

Query tip: Describe what’s in the image (e.g. "food", "person skating") rather than intents like "i want food". The script wraps your text as "This is a photo of …" for better matching. Use --raw to use your exact query.

Retrieval runs fully offline after the first run (model/tokenizer are loaded from the Hugging Face cache with local_files_only=True). Run build or retrieval once with internet to populate the cache, then it works without network.

• No login/session setup is required.
• Storage/account views now report local disk usage.
• All imported videos/images/audio remain on local disk.

• nexa-caption-lab/
• nexa-video-context-lab/