README
Boostify Dataset Studio
Professional desktop app (Tauri + React + Vite + TailwindCSS) that automatically turns entire music-video libraries into training-ready datasets for NVIDIA Cosmos Predict / Transfer, LoRA, NeMo and future Boostify models.
Point it at a folder or a whole SSD — it scans, indexes, splits, analyzes, captions, scores, de-duplicates and exports. The user never writes metadata by hand.
Stack
- Shell: Tauri v2 (Rust) — native filesystem, SQLite, FFmpeg orchestration
- UI: React 18 + TypeScript + Vite + TailwindCSS (dark, DaVinci/Lightroom feel)
- Index: SQLite (rusqlite, bundled) with SHA-256 dedup — files are never reprocessed
- Media: FFmpeg / FFprobe (scene detection, cutting, thumbnails, perceptual hashing)
Architecture (modular services)
Every stage is an independent module under src-tauri/src/:
| Module | Responsibility |
|---|---|
scanner.rs | Recursive discovery (MP4/MOV/MXF/BRAW/R3D/ProRes/XAVC…), SHA-256 hashing, artist/project inference |
db.rs | SQLite schema + CRUD (videos, clips, datasets, settings) |
probe.rs | FFprobe metadata + FFmpeg scene-cut detection |
splitter.rs | Clip cutting, thumbnails, 8×8 average-hash, brightness/sharpness probes |
ai.rs | Scene analysis + rich captioning + auto-tagging (NVIDIA NIM hook, heuristic fallback) |
pipeline.rs | Orchestrator: split → analyze → caption → score → dedup → approve |
dataset.rs | Dataset tree + train/validation/test.jsonl export (Cosmos/LoRA/NeMo…) |
system.rs | GPU detection (CUDA/Metal/CPU), dependency + storage checks |
watch.rs | Watch mode — auto-index new files as they land |
UI pages: Dashboard, Library, Pipeline, Review (Lightroom-style), Smart Search, Datasets, Settings.
Prerequisites
- Node 18+ and Rust (stable)
- FFmpeg + FFprobe on
PATH(required for splitting/probing). On macOS:brew install ffmpeg. You can also setFFMPEG_PATH/FFPROBE_PATH.
Develop
npm install
npm run app:dev # tauri dev — opens the desktop window
Build installers
npm run app:build # .dmg / .app (macOS), .msi / .exe (Windows)
Plugging in real AI models
The pipeline runs fully offline with local heuristics. To upgrade quality:
- Captions: add an NVIDIA NIM API key in Settings — wire the call in
ai.rs::caption_with_nim(vision-language model, e.g.nvidia/vila). - Detection / pose: run YOLO + MediaPipe as sidecars and feed results into
ai.rs::analyze_scene. - Motion signals: populate
pose/,depth/,edges/,segmentation/,optical_flow/sidecars during export indataset.rs.
Rules honored
- Never reprocess a file already seen (hash + path guard).
- Metadata is always editable in the Review panel.
- Every stage is a separate, swappable service with structured logs.
Contributors
18 commits
Similar repos
Languages
Python
52.1%
JavaScript
16.3%
Rust
15.7%
TypeScript
10.3%
HTML
5.1%
README
Boostify Dataset Studio
Professional desktop app (Tauri + React + Vite + TailwindCSS) that automatically turns entire music-video libraries into training-ready datasets for NVIDIA Cosmos Predict / Transfer, LoRA, NeMo and future Boostify models.
Point it at a folder or a whole SSD — it scans, indexes, splits, analyzes, captions, scores, de-duplicates and exports. The user never writes metadata by hand.
Stack
- Shell: Tauri v2 (Rust) — native filesystem, SQLite, FFmpeg orchestration
- UI: React 18 + TypeScript + Vite + TailwindCSS (dark, DaVinci/Lightroom feel)
- Index: SQLite (rusqlite, bundled) with SHA-256 dedup — files are never reprocessed
- Media: FFmpeg / FFprobe (scene detection, cutting, thumbnails, perceptual hashing)
Architecture (modular services)
Every stage is an independent module under src-tauri/src/:
| Module | Responsibility |
|---|---|
scanner.rs | Recursive discovery (MP4/MOV/MXF/BRAW/R3D/ProRes/XAVC…), SHA-256 hashing, artist/project inference |
db.rs | SQLite schema + CRUD (videos, clips, datasets, settings) |
probe.rs | FFprobe metadata + FFmpeg scene-cut detection |
splitter.rs | Clip cutting, thumbnails, 8×8 average-hash, brightness/sharpness probes |
ai.rs | Scene analysis + rich captioning + auto-tagging (NVIDIA NIM hook, heuristic fallback) |
pipeline.rs | Orchestrator: split → analyze → caption → score → dedup → approve |
dataset.rs | Dataset tree + train/validation/test.jsonl export (Cosmos/LoRA/NeMo…) |
system.rs | GPU detection (CUDA/Metal/CPU), dependency + storage checks |
watch.rs | Watch mode — auto-index new files as they land |
UI pages: Dashboard, Library, Pipeline, Review (Lightroom-style), Smart Search, Datasets, Settings.
Prerequisites
- Node 18+ and Rust (stable)
- FFmpeg + FFprobe on
PATH(required for splitting/probing). On macOS:brew install ffmpeg. You can also setFFMPEG_PATH/FFPROBE_PATH.
Develop
npm install
npm run app:dev # tauri dev — opens the desktop window
Build installers
npm run app:build # .dmg / .app (macOS), .msi / .exe (Windows)
Plugging in real AI models
The pipeline runs fully offline with local heuristics. To upgrade quality:
- Captions: add an NVIDIA NIM API key in Settings — wire the call in
ai.rs::caption_with_nim(vision-language model, e.g.nvidia/vila). - Detection / pose: run YOLO + MediaPipe as sidecars and feed results into
ai.rs::analyze_scene. - Motion signals: populate
pose/,depth/,edges/,segmentation/,optical_flow/sidecars during export indataset.rs.
Rules honored
- Never reprocess a file already seen (hash + path guard).
- Metadata is always editable in the Review panel.
- Every stage is a separate, swappable service with structured logs.
Similar repos
Contributors
18 commits
Languages
Python
52.1%
JavaScript
16.3%
Rust
15.7%
TypeScript
10.3%
HTML
5.1%