Implement NSFW filter using Hugging Face Transformers

- Add NSFW content detection using Falconsai/nsfw_image_detection model
- Filter only product images for inappropriate content (porn/explicit)
- Add transformers and torch dependencies to requirements.txt
- Implement fail-safe design with comprehensive logging
- Block content with >70% confidence threshold
- Provide clear error messages to users

app.py

@@ -18,6 +18,7 @@ import threading
 
 import gradio as gr
 from supabase import create_client, Client
+from transformers import pipeline
 
 # -----------------------------------------------------------------------------
 # Environment & Supabase setup
@@ -76,6 +77,20 @@ except Exception as exc:  # noqa: BLE001
     print(f"[startup] Bucket check/create raised {exc!r}. Continuing…")
 
 
+# -----------------------------------------------------------------------------
+# NSFW Filter Setup
+# -----------------------------------------------------------------------------
+
+# Initialize NSFW classifier at startup
+print("[startup] Loading NSFW classifier...")
+try:
+    nsfw_classifier = pipeline("image-classification", model="Falconsai/nsfw_image_detection")
+    print("[startup] NSFW classifier loaded successfully")
+except Exception as exc:
+    print(f"[startup] Failed to load NSFW classifier: {exc!r}")
+    nsfw_classifier = None
+
+
 # -----------------------------------------------------------------------------
 # Rate Limiting
 # -----------------------------------------------------------------------------
@@ -273,6 +288,37 @@ def _public_storage_url(path: str) -> str:
     return f"{SUPABASE_URL}/storage/v1/object/public/{path.lstrip('/')}"
 
 
+def is_nsfw_content(img: Image.Image) -> bool:
+    """Check if image contains NSFW content using Hugging Face transformer."""
+    if nsfw_classifier is None:
+        print("[NSFW] Classifier not available, skipping check")
+        return False
+    
+    try:
+        # Run classification
+        results = nsfw_classifier(img)
+        print(f"[NSFW] Classification results: {results}")
+        
+        # Check for explicit content
+        for result in results:
+            label = result['label'].lower()
+            score = result['score']
+            print(f"[NSFW] Label: {label}, Score: {score:.3f}")
+            
+            # Flag as NSFW if "porn" or explicit content detected with high confidence
+            if label in ['porn', 'nsfw', 'explicit'] and score > 0.7:
+                print(f"[NSFW] BLOCKED - {label} detected with {score:.3f} confidence")
+                return True
+                
+        print("[NSFW] Content approved")
+        return False
+        
+    except Exception as exc:
+        print(f"[NSFW] Error during classification: {exc!r}")
+        # Fail open - don't block if classifier has issues
+        return False
+
+
 # -----------------------------------------------------------------------------
 # Main generate function
 # -----------------------------------------------------------------------------
@@ -292,6 +338,11 @@ def generate(
     if not check_rate_limit(client_ip):
         raise gr.Error("Rate limit exceeded. Please try again later (30 requests per hour limit).")
 
+    # NSFW content filtering - only check product image
+    print(f"[NSFW] Checking product image for inappropriate content...")
+    if is_nsfw_content(garment_img):
+        raise gr.Error("Product image contains inappropriate content. Please use a different image.")
+
     # 1. Persist both images to Supabase storage
     job_id = str(uuid.uuid4())
     folder = f"user_uploads/gradio/{job_id}"

requirements.txt

@@ -2,4 +2,6 @@ gradio>=4.29.0
 supabase>=2.0.0
 python-dotenv>=1.0.0
 requests>=2.31.0
-Pillow 
+Pillow
+transformers>=4.30.0
+torch>=2.0.0