manga_bubbles_detect

Русская версия

YOLOv8 model for detecting speech bubbles in manga pages. Trained to locate bubble bounding boxes (single class location-of-bubbles).

Validation mAP50 ≈ 0.977

Part of Manga Translate — a full manga translation pipeline (bubble detection → OCR → inpaint → LLM → render).

Usage

from ultralytics import YOLO

model = YOLO("bubbles_detect.pt")
results = model.predict("page.jpg", conf=0.25, iou=0.5, imgsz=1024)

Or use it automatically via the Manga Translate app — the model is downloaded from here on first run.

Training data

Trained on a combined dataset published at PSImera/manga_bubbles_detect:

• manga.v4i (Roboflow) — 1304/189/103 train/valid/test pages
• 1079 additional pages from DLS Manga Translator, manually corrected in CVAT

Training scripts and reports: training/