Update README.md

README.md

@@ -62,7 +62,7 @@ The full pipeline consists of four components:
 ### Motion Planning
 
 <p align="center">
-<img src="docs/prediction_min_ade.svg" alt="EPE" width="500px">
+<img src="docs/prediction_min_ade.png" alt="EPE" width="500px">
 </p>
 
 GARFIELD infers accurate motion from very sparse information and, with only four conditioning points, outperforms Motion-I2V using sixteen.
@@ -70,7 +70,7 @@ GARFIELD infers accurate motion from very sparse information and, with only four
 ### Direct Density Decoding
 
 <p align="center">
-<img src="docs/energy_score_vs_latency_new.svg" alt="Discrete Energy Scores" width="500px">
+<img src="docs/energy_score_vs_latency_new.png" alt="Discrete Energy Scores" width="500px">
 </p>
 
 The density decoder achieves superior discrete energy scores while estimating motion densities orders of magnitude faster than Monte-Carlo sampling.
@@ -79,7 +79,7 @@ The density decoder achieves superior discrete energy scores while estimating mo
 ### Entropy-informed Conditioning
 
 <p align="center">
-<img src="docs/entropy_vs_epe_cond_new_data.svg" alt="Entropy-based cond" width="500px">
+<img src="docs/entropy_vs_epe_cond_new_data.png" alt="Entropy-based cond" width="500px">
 </p>
 
 Selecting additional constraints by entropy collapses uncertainty without ground-truth errors and reaches comparable performance with fewer conditioning points.