Update README.md

README.md

@@ -39,7 +39,7 @@ Website: [https://behavior-in-the-wild.github.io/behavior-llava.html](https://be
 
 By modeling these downstream receiver behaviors, training on BLIFT improves **content understanding** of VLMs, showing significant improvements across 46 tasks in image, video, text, and audio understanding.
 
-<img src="./bllava-fig_2.pdf" alt="bllava-fig" width="1000"/>
+<img src="./bllava-fig_2.png" alt="bllava-fig" width="1000"/>
 
 ---
 
@@ -75,7 +75,7 @@ BLIFT combines high-quality behavioral data from two sources:
 - Metadata: Likes, views, top comments, replay graphs
 - Filtering: English language, minimum 10k views, NSFW, duplicates
 
-<img src="./filtering-final.pdf" alt="filtering" width="1000"/>
+<img src="./filtering-final.png" alt="filtering" width="1000"/>
 
 ---
 
@@ -89,7 +89,7 @@ Using BLIFT to train **Behavior-LLaVA** (a fine-tuned LLaMA-Vid), the model outp
 - 26 benchmark datasets
 - Across image, video, audio, and text modalities
 
-<img src="./radar_chart (1).pdf" alt="results" width="1000"/>
+<img src="./radar_chart (1).png" alt="results" width="1000"/>
 
 
 ---