Update README.md

Author: Costwen

README.md

@@ -6,6 +6,12 @@
 
 **TL;DR**: Trash to Treasure introduces **DualityForge**, a controllable diffusion-based framework that turns real videos into counterfactual ones, automatically generating paired videos and QA data for contrastive training. Based on the large-scale **DualityVidQA** dataset and the proposed **DNA-Train** SFT–RL regime with ℓ1-normalized advantages, our approach reduces hallucinations in multimodal LLMs by 24% and shows strong generalization across benchmarks. Dataset and code will be released.
 
+## Method Overview
+
+<img width="1678" height="980" alt="image" src="https://github.com/user-attachments/assets/a72a0ee9-7927-4ad0-9ad4-1ba448cbb0e8" />
+
+## Update
+
 
 ## 📎 Citation