---

ENGLISH

What's New in v4.0

• 📝 Subtitle detection rewritten — works at any resolution and aspect ratio (16:9, 9:16, 480p…), detects single-line and top-positioned subtitles, and scans the full frame for text lines instead of a fixed bottom band
• ✂️ Smarter subtitle removal — frames are cropped from the correct side (top or bottom) and the crop shrinks around overlay/subtitle zones instead of falling back to an unconstrained crop
• 🖥️ System monitor fixed — GPU utilization and VRAM now read device-wide via nvidia-smi (with a torch.cuda.mem_get_info fallback), so real usage shows up even while training runs in a subprocess or torch is CPU-only
• 🔁 --no-resume actually works — the CLI flag is now wired through to the processor
• 🐛 Bug fixes — duplicate translation keys caused the wrong tooltip on the detection-confidence spinbox; duplicate _spinbox_style method removed
• 🧹 Codebase cleanup — 7 unused legacy modules removed, dead imports cleaned across 16 files, caches and stale files purged

What's New in v3.0

• 🧩 Caption Studio — one page for both auto-generation (WD14) and manual editing, with live Danbooru tag autocomplete while you type
• 🎯 Quality presets — pick High Accuracy / Balanced / High Speed; the best WD14 model, confidence, and tag count are chosen for you
• 🔖 Trigger + Suffix — prepend a LoRA keyword and append quality tokens in the same UI
• 🖼️ Florence-2 — optional natural-language captions for full-sentence alt-text
• ✂️ SAM2 + PySceneDetect — subject-aware masking and cut-level deduplication
• 🎨 Dark / Light theme + font scaling + differentiated Start/Pause/Skip/Stop buttons

What Is It?

LoRA-Harvester extracts high-quality frames from videos, sorts them by character using face recognition, and optionally generates captions — producing a clean, ready-to-train LoRA dataset in minutes instead of hours.

Main workflow:

Video(s) → Extract frames → AI detects person → Smart crop → Quality filter → Caption → Dataset

Character Sorter workflow:

Image folder → Face detection → Match / cluster by identity → Sort into named folders

---

Features

---

Installation

# 1. Clone the repository
git clone https://github.com/AllastorV/LoRA-Harvester.git
cd LoRA-Harvester

# 2. Create a virtual environment
python -m venv venv
source venv/bin/activate        # Linux / Mac
venv\Scripts\activate           # Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. Launch
python main.py

GPU support: Install PyTorch with CUDA before installing other requirements. See https://pytorch.org/get-started/locally/

---

Launching

---

Usage

GUI Mode

python main.py

1. Step 1 — Drop video file(s) or select a folder
2. Step 2 — Configure settings (format, interval, model, quality, captions)
3. Step 3 — Click Start and monitor the log

CLI Mode

# Basic
python scripts/cli.py video.mp4

# Common options
python scripts/cli.py video.mp4 -f 1:1 -i 15 -c 0.7 --quality --ensemble --turbo

# With captions
python scripts/cli.py video.mp4 --caption --caption-mode combined --trigger "mychar"

# Character sort
python scripts/character_sort.py /images/input --references /refs --max-characters 2

---

Settings Reference

Video Extractor

config.yaml — Advanced

quality:
  blur_threshold: 80.0        # Min sharpness (Laplacian variance). Higher = stricter
  brightness_min: 35          # Darkest allowed frame (0–255)
  brightness_max: 225         # Brightest allowed frame (0–255)
  duplicate_threshold: 0.90   # Similarity cutoff (0–1). Higher = keep more
  min_contrast: 20            # Minimum contrast level

captioning:
  wd14:
    model: "SmilingWolf/wd-convnext-tagger-v3"   # Auto-selected by quality preset in GUI
  florence2:
    model: "microsoft/Florence-2-base"   # "-base" (fast) or "-large" (accurate)
    task: "<MORE_DETAILED_CAPTION>"
  tags:
    min_confidence: 0.35      # Min tag confidence (0–1)
    max_tags: 30              # Max tags per caption
    trigger_word: ""          # Prepended to every caption
    caption_suffix: ""        # Appended at end (e.g. "masterpiece, best quality")

Character Sorter

Output structure:

_sorted/
├── character_name/   ← matched to reference images
├── character_01/     ← auto-clustered unknown group
├── other/            ← overflow when max_characters limit reached
├── unknown/          ← could not form a cluster
├── no_face/          ← no face detected
└── multi_face/       ← multiple faces, no clear match

---

Quick Presets

---

Troubleshooting

---

Project Structure

LoRA-Harvester/
├── main.py                       # GUI entry point
├── run.bat                       # Windows launcher (CMD closes after start)
├── run_silent.vbs                # Silent launcher (no window)
├── config.yaml                   # Advanced configuration
├── requirements.txt
├── assets/
│   └── icon.png                  # Application icon
├── src/
│   ├── core/
│   │   ├── unified_processor.py   # Main video processing engine
│   │   ├── enhanced_processor.py  # Checkpoint/resume processing
│   │   ├── text_detector.py       # Subtitle & overlay detection
│   │   ├── cropper.py             # Smart aspect-ratio crop
│   │   ├── character_recognizer.py
│   │   ├── advanced_captioner.py  # WD14 / Danbooru tagger
│   │   ├── florence2_captioner.py # Natural-language captions
│   │   └── tag_autocomplete.py    # Danbooru tag loader for autocomplete
│   └── ui/
│       ├── main_window.py
│       ├── caption_studio_page.py # Generate + Edit tabs, Danbooru autocomplete
│       ├── character_sort_page.py
│       ├── tag_frequency_page.py
│       ├── translations.py
│       └── theme.py
├── scripts/
│   ├── cli.py                    # Command-line interface
│   ├── character_sort.py         # Character sorter CLI
│   ├── check_gpu.py              # GPU diagnostics
│   ├── install.bat
│   ├── install_gpu.bat
│   └── run_batch.bat
└── docs/
    ├── CHANGELOG.md
    ├── QUICKSTART.md
    ├── OPTIMIZATION.md
    ├── ENSEMBLE.md
    └── SECURITY.md

---

TURKCE

v4.0'da Yenilikler

• 📝 Altyazı tespiti yeniden yazıldı — her çözünürlük ve en-boy oranında çalışır (16:9, 9:16, 480p…), tek satırlık ve üst konumlu altyazıları da bulur, sabit alt bant yerine tüm kareyi metin satırları için tarar
• ✂️ Daha akıllı altyazı temizleme — kare doğru taraftan (üst veya alt) kırpılır; kırpma alanı overlay/altyazı bölgesinin etrafında daraltılır
• 🖥️ Sistem monitörü düzeltildi — GPU kullanımı ve VRAM artık nvidia-smi ile cihaz genelinde okunur (torch.cuda.mem_get_info yedeği ile); eğitim alt süreçte çalışırken veya torch CPU sürümüyken bile gerçek değerler görünür
• 🔁 --no-resume artık çalışıyor — CLI bayrağı işlemciye bağlandı
• 🐛 Hata düzeltmeleri — çift çeviri anahtarları yüzünden tespit güveni kutusunda yanlış ipucu görünüyordu; çift _spinbox_style metodu kaldırıldı
• 🧹 Kod temizliği — kullanılmayan 7 eski modül kaldırıldı, 16 dosyada ölü import temizlendi, cache ve bayat dosyalar silindi

v3.0'da Yenilikler

• 🧩 Altyazi Studyosu — otomatik uretim (WD14) ve manuel duzenleme tek sayfada; yazarken canli Danbooru etiket oneri
• 🎯 Kalite onayarlari — Yuksek Dogruluk / Dengeli / Yuksek Hiz sec, model ve degerler otomatik ayarlanir
• 🔖 Tetikleyici + Sonek — hem LoRA anahtar kelimesi bastan hem kalite tokenlari sondan ayni ekranda
• 🖼️ Florence-2 — istege bagli, tam cumleli dogal dil aciklamalar
• ✂️ SAM2 + PySceneDetect — nesne odakli maskeleme ve kesim-seviyesinde tekrar azaltma
• 🎨 Karanlik / Aydinlik tema + yazi tipi olcekleme + ayri Baslat/Duraklat/Atla/Durdur butonlari

Ne Ise Yarar?

LoRA-Harvester, videolardan yuksek kaliteli kareler cikarir, yuz tanima ile bunlari karaktere gore siralar ve istege bagli olarak caption uretir. Saatlerce suren manuel islemi dakikalar icinde tamamlayarak egitime hazir bir LoRA dataseti olusturur.

Ana akis:

Video(lar) → Kare cikar → AI kisi tespiti → Akilli kirp → Kalite filtresi → Caption → Dataset

Karakter Siralayici akisi:

Gorsel klasoru → Yuz tespiti → Kimlige gore eslestir/kumele → Isimli klasorlere sirala

---

Ozellikler

---

Kurulum

# 1. Klonla
git clone https://github.com/AllastorV/LoRA-Harvester.git
cd LoRA-Harvester

# 2. Sanal ortam olustur
python -m venv venv
source venv/bin/activate        # Linux / Mac
venv\Scripts\activate           # Windows

# 3. Bagımliliklari kur
pip install -r requirements.txt

# 4. Basla
python main.py

GPU destegi: Diger gereksinimleri kurmadan once PyTorch'u CUDA ile kur. Bkz. https://pytorch.org/get-started/locally/

---

Baslатma Yontemleri

---

Kullanim

Arayuz (GUI) Modu

python main.py

1. Adim 1 — Video dosya(lari) sürükle-birak ya da klasor sec
2. Adim 2 — Ayarlari yapilandir (format, aralik, model, kalite, caption)
3. Adim 3 — Baslat'a tikla, logu izle

Komut Satiri (CLI) Modu

# Temel kullanim
python scripts/cli.py video.mp4

# Yaygin secenekler
python scripts/cli.py video.mp4 -f 1:1 -i 15 -c 0.7 --quality --ensemble --turbo

# Caption ile
python scripts/cli.py video.mp4 --caption --caption-mode combined --trigger "karakterim"

# Karakter siralayici
python scripts/character_sort.py /gorseller/giris --references /referanslar --max-characters 2

---

Ayar Referansi

Video Cikarici

Karakter Siralayici

---

Hizli Onayarlar

---

Sorun Giderme

---

Proje Yapisi

LoRA-Harvester/
├── main.py                       # GUI giris noktasi
├── run.bat                       # Windows baslayicisi (CMD kapanir)
├── run_silent.vbs                # Sessiz baslayici (pencere yok)
├── config.yaml                   # Gelismis yapilandirma
├── requirements.txt
├── assets/
│   └── icon.png                  # Uygulama ikonu
├── src/
│   ├── core/
│   │   ├── unified_processor.py
│   │   ├── enhanced_processor.py   # Checkpoint/devam ettirme islemcisi
│   │   ├── text_detector.py        # Altyazi ve overlay tespiti
│   │   ├── cropper.py              # Akilli en-boy oranli kirpma
│   │   ├── character_recognizer.py
│   │   ├── advanced_captioner.py   # WD14 / Danbooru etiketleyici
│   │   ├── florence2_captioner.py  # Dogal dil aciklamalar
│   │   └── tag_autocomplete.py     # Otomatik tamamlama icin etiket yukleyici
│   └── ui/
│       ├── main_window.py
│       ├── caption_studio_page.py  # Olustur + Duzenle sekmeleri
│       ├── character_sort_page.py
│       ├── tag_frequency_page.py
│       ├── translations.py
│       └── theme.py
├── scripts/
│   ├── cli.py                    # Komut satiri arayuzu
│   ├── character_sort.py         # Karakter siralayici CLI
│   ├── check_gpu.py              # GPU tani araci
│   ├── install.bat
│   ├── install_gpu.bat
│   └── run_batch.bat
└── docs/
    ├── CHANGELOG.md
    ├── QUICKSTART.md
    ├── OPTIMIZATION.md
    ├── ENSEMBLE.md
    └── SECURITY.md

---