Vision Studio

A unified, dark-themed PyQt desktop app with four top-level tabs:

• Video → Images extracts frames from video files.
• OCR processes image batches with PaddleOCR, EasyOCR, or MinerU2.5-Pro.
• Dialogue Tools cleans OCR dialogue and turns frame-by-frame OCR dumps into readable transcripts.
• Overall Run chains video extraction, OCR, and dialogue cleaning into one guided workflow.

---

Features

Video → Images

• Five extraction modes: total count, 1 image/second, all frames, 4 images/second, and custom fps.
• Live resolution, fps, duration, and frame-count metadata.
• Output estimates before extraction.
• JPG, PNG, and WebP output with quality controls.
• Background extraction with progress and cancellation.
• Stable output names such as clip_000012.jpg.

OCR

• Drag & drop or pick multiple images (png / jpg / jpeg / bmp / webp / tif / tiff).
• Threaded PaddleOCR worker — UI stays responsive, Cancel supported.
• Three selectable OCR backends:
  • PaddleOCR for fast general-purpose OCR.
  • EasyOCR as a lightweight optional alternative.
  • MinerU2.5-Pro for multilingual, layout-aware document parsing, including text, titles, tables, equations, and code blocks.
• Model download/loading bar for PaddleOCR, EasyOCR, and MinerU, with percentage, downloaded size, and cancellation whenever the backend exposes byte totals.
• 16 language models out of the box (English, Chinese, Japanese, Korean, French, German, Arabic, Hindi, …).
• Optional rotation / angle classifier.
• Per-file preview, full text view, and JSON view with bounding boxes + confidence scores.
• Export all results to TXT or JSON in one click.
• Export All… — pick a folder, get both ocr_results_<timestamp>.txt and ocr_results_<timestamp>.json written at the same time.
• Auto-export on completion — flip the checkbox, pick a folder once, and every finished OCR run is saved automatically. The choice is remembered.
• Modern Fusion dark theme via pure QSS — no external style files needed.

Dialogue Tools

• Clean regular OCR text by removing exact and fuzzy repetition.
• Keep more complete readings and join wrapped or hyphenated dialogue.
• Normalize dialogue written as SPEAKER: text or a speaker on its own line.
• Flag possibly incomplete text without inventing missing words.
• Consolidate progressive typewriter-style readings from frame-by-frame OCR.
• Filter transcript entries by confidence and optionally include narration, source frame ranges, and custom speaker aliases.
• Preview the finished result and optionally save a detailed JSON review report.

Overall Run

• Runs the full sequence automatically: video → temp_img_ocr → OCR → cleaned text.
• Collects the extraction, backend, language, and cleaner settings in one place.
• Shows a mandatory settings warning and complete summary immediately before execution.
• Displays extraction progress, per-image OCR progress, and OCR model-download progress.
• Supports safe cancellation between and during processing stages.
• Saves temporary frames to temp_img_ocr inside the selected output folder.
• Saves <video>_ocr_raw.txt, <video>_cleaned.txt, and a JSON cleaning report in the selected output folder.

---

Install

# 1. Create a fresh env (recommended)
python -m venv .venv
source .venv/bin/activate            # Windows: .venv\Scripts\activate

# 2. Install the shared UI, image, and video dependencies
pip install PyQt6 Pillow numpy opencv-python

# 3. Install PaddlePaddle (CPU build is enough for most cases)
#    Pick the wheel for your platform from:
#    https://www.paddlepaddle.org.cn/install/quick
pip install paddlepaddle

# 4. Install PaddleOCR
pip install paddleocr

On first run PaddleOCR will auto-download the detection / recognition models for the language you pick. They live in ~/.paddleocr/, so this only happens once.

Optional backends

EasyOCR:

pip install easyocr

MinerU2.5-Pro (local Transformers backend):

pip install -U "mineru-vl-utils[transformers]"

Select MinerU2.5-Pro in the Backend menu. Language detection is automatic. The first run downloads opendatalab/MinerU2.5-Pro-2604-1.2B from Hugging Face. The app uses the simple Transformers backend, which runs on the device selected automatically by PyTorch/Accelerate; a CUDA GPU is strongly recommended for practical speed.

Run

python paddle_ocr_studio.py

---

OCR output formats

TXT

Plain, copy-pasteable text. Section per file, with metadata:

==============================================================================
[1/2] receipt.png
Path:    /tmp/receipt.png
Elapsed: 1.234s
Lines:   12
------------------------------------------------------------------------------
COFFEE SHOP
2x Latte     8.50
1x Croissant 3.20
...

JSON

Full structured dump with bounding boxes and confidence:

{
  "generated_at": "2026-07-03T12:34:56",
  "app": "Vision Studio",
  "count": 2,
  "results": [
    {
      "path": "/tmp/receipt.png",
      "name": "receipt.png",
      "elapsed_sec": 1.234,
      "line_count": 12,
      "lines": [
        {
          "text": "COFFEE SHOP",
          "confidence": 0.9876,
          "box": [[12.0, 30.0], [220.0, 30.0], [220.0, 60.0], [12.0, 60.0]],
          "type": "text"
        }
      ],
      "error": null
    }
  ]
}

---

Project layout

paddle_ocr_studio.py     # the entire application (single file)
ocr_dialogue_cleaner.py  # dialogue cleanup engine
video_ocr_dialogue_extractor.py  # frame transcript engine
requirements.txt
README.md

The PyQt tabbed UI and its workers live in paddle_ocr_studio.py. The dialogue algorithms remain in focused standard-library modules so they can also be used from the command line.

---

Caution

I Have a 10Gb video card so im not sure if its works with lesser vram

---

License

GNU General Public License v2.0