Update README.md

README.md

@@ -146,7 +146,7 @@ The workflow operates in **two main steps**:
 
 ### Prompt Adherence Evaluation
 
-* Quantitative prompt adherence was evaluated on a test set of 50 images with varying complexity of actions.
+* Quantitative prompt adherence was evaluated on a test set of 30 images with varying complexity of actions.
 
 * For **simple actions** (e.g., sitting, standing, swimming):
 
@@ -162,11 +162,11 @@ The workflow operates in **two main steps**:
 
 * Character consistency was evaluated manually by comparing generated characters with the reference images in terms of **hair color**, **ethnic appearance**, and **facial structure**.
 
-* We used count-based consistency across 200 generated samples:
+* We used count-based consistency across 20 generated samples for each sections:
 
-  * **Hair color match rate:** 91% (182/200)
-  * **European appearance match:** 88% accuracy (88/100 reference-preserved)
-  * **Asian appearance match:** 85% accuracy (85/100 reference-preserved)
+  * **Hair color match rate:** 90% (18/20)
+  * **European appearance match:** 100% accuracy (20/20 reference-preserved)
+  * **Asian appearance match:** 95% accuracy (19/20 reference-preserved)
 
 * Consistency is high within supported classes (Asian and European), but we observe a drop-off for characters with underrepresented traits or less common species due to data limitations.