DOCS/CONTENT & AUDIO CRAFT/Image Scraping & Vision

Image Scraping & Vision

Multi-source retrieval, OCR watermark rejection, framing tags, and library grammar.

Stages 05, 06, and 07 plan visual demand, scrape high-resolution imagery, reject watermarked frames with OCR, and catalog sorted assets into a reusable library.

Multi-Source Scraping

The download engine queries multiple sources configured in config.yaml:

•Real-World & History: Google Images, Bing, DuckDuckGo, Wikimedia Commons.
•Anime & Fiction: Safebooru, Wallhaven, and dedicated anime imageboards.

OCR Watermark & Quality Gates

Every candidate image is evaluated through automated local checks before entering the library:

•OCR Text & Watermark Filter: Tesseract scans the frame. Any watermark, logo, or burned-in subtitles exceeding 5% coverage trigger instant rejection.
•Resolution Floor: Images smaller than 640×640px or with low Laplacian sharpness variance are discarded.
•Perceptual Hash Deduplication: pHash comparison discards duplicate or near-identical images within the same series.

Framing & Naming Grammar

Stage 07 sorts images into images/<character>/ with strict camera tagging:

Image Naming Convention

1<entity>-<descriptor>-<shot_type>-<focal_x>-<focal_y>.<ext>

3Shot Types:

4- ECU: Extreme Close Up (face, eyes, intense emotion)

5- MCU: Medium Close Up (head and shoulders)

6- MS: Medium Shot (waist up)

7- WS: Wide Shot (full body or environment scenery)

9Focal Coordinates:

10- c-t: Center-Top (standard for portraits)

11- c-m: Center-Middle (landscapes, centered subjects)

12- l-m: Left-Middle / r-m: Right-Middle

Was this "Image Scraping & Vision" page helpful?

CMU Arpabet Phonemes

Accurate phonetic overrides for anime lore, fantasy names, and uncommon terms.

Audio Mastering & -14 LUFS

Voice synthesis, BGM sidechain ducking, SFX placement, and broadcast loudness.