Update README.md

README.md

@@ -37,8 +37,17 @@ We've made a first attempt to filter out duplicates (e.g., same appearance descr
 }
 ```
 
-At times, the appearance descriptions may be lengthy, spanning multiple paragraphs and introducing several appearances for a single character.
 
-However, this can result in mismatches with the associated image, as only the first figure from the Wiki page is captured.
+**Caution for Image-Text Mismatch**
+
+- **Too much text**
+
+At times, the appearance descriptions may be lengthy, spanning multiple paragraphs and introducing several appearances for a single character. However, this can result in mismatches with the associated image, as only the first figure from the Wiki page is captured.
 
 Please take care when using this data for image captioning tasks to avoid potential hallucination issues. To remedy this, we recommend performing some processing on the data. Using a powerful vision-language model to filter and extract only the most relevant appearance description can help create a more precise match between each image and its caption.
+
+- **Incomplete Image**
+
+Sometimes, the character figure in the image is only partially visible (e.g., a headshot or half-body), which may not fully align with the details in the "appearance" field that describe a full-body view. This can introduce another type of mismatch.
+
+To address this, consider using a vision (-language) model to identify and categorize the character’s visible features (e.g., face-only, half-body, or full-body) and filter or modify descriptions accordingly.