saengha/qwen3-vl-2b-finetuned-korean-game-ui-ocr
The saengha/qwen3-vl-2b-finetuned-korean-game-ui-ocr model is a 2 billion parameter vision-language model, fine-tuned from Qwen3-VL-2B-Instruct, specifically for Korean game image recognition. It excels at extracting and segmenting Korean text from complex game UIs, distinguishing between dialogue, UI elements, and other text. This model significantly improves Korean OCR accuracy for in-game fonts and reduces hallucination issues common in base models, making it ideal for automated Korean game UI analysis.
Loading preview...
Overview
This model is a specialized fine-tuned version of Qwen3-VL-2B-Instruct, developed by saengha, focusing exclusively on Korean Game Image Recognition. It leverages Knowledge Distillation with diverse game images to enhance its performance beyond the base model.
Key Capabilities
- Complex UI Segmentation: Accurately distinguishes and segments text within dialogue boxes, system UI elements, and other text areas in game screenshots.
- Reduced Hallucination: Effectively eliminates the repetitive text generation (hallucination) issues often observed in the base Qwen3-VL-2B-Instruct model.
- Improved Korean OCR Accuracy: Demonstrates significantly enhanced recognition of Korean text, including special in-game fonts and smaller text sizes, within complex game UIs.
Use Cases
This model is specifically designed for tasks involving the extraction and parsing of Korean text from game screenshots. It is highly effective for:
- Automated text extraction from Korean game interfaces.
- Analyzing and categorizing in-game dialogue and UI elements.
- Supporting game localization efforts by accurately identifying and extracting text for translation.
Note: This model is optimized for game images and Korean text. Its performance may vary for general image recognition or non-Korean natural language tasks.