Service Configuration: Text Block Detection
What detection does
Section titled “What detection does”The Text Block Detection section in Service Profile configures dedicated text-block detection models. Current vision-language models can often indicate roughly where text is, but are not reliable at locating precise coordinates. This task therefore uses models specifically adapted for OCR, object detection, or region segmentation.
Text block detection finds text regions in comic images and creates text blocks. These blocks determine what later Extract Text and translation steps process.
It only locates text regions; it is not the same as text extraction. Whether a detector also returns source text depends on the service.
Supported services
Section titled “Supported services”| Service | API required | Returns source text during detection | Best for |
|---|---|---|---|
| Local text block detector | No | No | Free, offline box detection |
| PaddleOCR VL 1.6 | Yes | Yes | Detection and OCR in one request |
| GLM-OCR | Yes | Yes | Slightly weaker at extracting vertical source text |
Local text block detector
Section titled “Local text block detector”Open Settings → Service Profile → Text Block Detection and select Download. The app creates the local profile after the download completes. You can also import a model ZIP offline; the ZIP is available on the Neko Translator website.
The local detector currently returns text-block positions only. For a complete workflow, configure a service that provides Extract Text, such as a vision-capable Generic AI model or the local text extractor. In most manga translation scenarios, pairing the local text-block detector with the local text extractor is sufficient and is currently the recommended setup.
The local text-block detector can automatically distinguish text inside bubbles from other text.
Downloads and imports may involve mirror selection, network errors, missing files, or hash verification failures. Only choose to continue after a verification failure if you trust the model source and understand the risk.
PaddleOCR VL 1.6 and GLM-OCR
Section titled “PaddleOCR VL 1.6 and GLM-OCR”Online OCR services require an API key. After creating a configuration, select the corresponding model in batch processing or the editor.
Because these services return source text during text block detection, a separate Extract Text step is usually unnecessary.
Both models are open source, and their maintainers provide source code and tutorials for deploying them locally. Refer to their official documentation for the exact steps.
Review detection results
Section titled “Review detection results”Always check:
- whether any speech bubbles or text outside bubbles were missed;
- whether multiple independent text blocks were merged by mistake;
- whether neighboring text blocks contain or overlap each other;
- whether the position and direction of each text block are correct;
- whether narration and sound effects need separate handling.