Update README.md

README.md

@@ -1,3 +1,31 @@
+
+# SAGE: Structure-Aware Geometric Regularization (ECCV-26)
+
+**Paper:** The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models  
+**Venue:** ECCV 2026  
+**Authors:** Adeel Yousaf, Soumik Ghosh, James Beetham, Amrit Singh Bedi, Mubarak Shah  
+**Institution:** University of Central Florida  
+**Project Page:** [https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/](https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/)
+
 ---
-license: cc-by-nc-4.0
+
+## Overview
+
+We show that existing T2I safety alignment methods create an **illusion of high utility** — they appear to have high-utility preservation under coarse metrics (FID, CLIPScore) but suffer significant drops in fine-grained semantic fidelity (TIFA). We trace this to **semantic collapse** in the text encoder embedding space.
+
+**SAGE** is a geometry-aware safety alignment method that preserves embedding spread and local similarity structure during fine-tuning, achieving only a **−1.2% TIFA drop** vs. **−6.2% for DES** while maintaining strong safety (Avg. ASR 1.2%).
+
 ---
+
+## Model
+
+This is the fine-tuned **text encoder** of Stable Diffusion v1.4. The UNet remains frozen.
+
+```python
+import torch
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
+checkpoint = torch.load("SAGE.pt", map_location="cpu")
+pipe.text_encoder.load_state_dict(checkpoint)
+pipe = pipe.to("cuda")