diff --git a/webcam_demo.py b/webcam_demo.py
new file mode 100644
index 00000000..27d4d34c
--- /dev/null
+++ b/webcam_demo.py
@@ -0,0 +1,163 @@
+import argparse
+import cv2
+import time
+import tempfile
+import os
+import sys
+from termcolor import colored
+
+# Ensure we can find the llava module if running from the root of the repo
+# This is classic python path hacking to make sure imports work without package installation
+current_dir = os.path.dirname(os.path.abspath(__file__))
+if current_dir not in sys.path:
+    sys.path.append(current_dir)
+
+import llava
+from llava import conversation as clib
+from llava.media import Image
+
+def capture_frame(camera_index=0):
+    """
+    Opens the webcam, allows it to warm up, and grabs a single frame.
+    Returns the frame (numpy array) or None if something went wrong.
+    """
+    print(f"[*] Connecting to webcam (ID: {camera_index})...")
+    cap = cv2.VideoCapture(camera_index)
+
+    if not cap.isOpened():
+        print("[!] Error: Could not open the webcam. check connection?")
+        return None
+
+    # Cameras often need a moment to adjust to light levels (auto-exposure)
+    # 15 frames is usually enough to settle the sensor
+    print("[*] Warming up camera sensor...")
+    for _ in range(15):
+        cap.read()
+        time.sleep(0.05)
+
+    ret, frame = cap.read()
+    cap.release()
+
+    if not ret:
+        print("[!] Error: Failed to grab a frame from the camera.")
+        return None
+    
+    return frame
+
+def processing_loop(model, args):
+    """
+    Main application loop: Capture -> Save -> Inference -> Cleanup
+    """
+    # Set up the conversation template
+    # 'auto' usually does a good job, but we default to vicuna if unsure
+    conv_template = clib.conv_templates.get(args.conv_mode)
+    if not conv_template and args.conv_mode != 'auto':
+        print(f"[!] Warning: Conversation mode '{args.conv_mode}' not found, defaulting to 'vicuna_v1'")
+        args.conv_mode = 'vicuna_v1'
+
+    # We set the global conversation just in case, though we usually manage it per loop
+    if args.conv_mode != 'auto':
+        clib.default_conversation = clib.conv_templates[args.conv_mode].copy()
+
+    while True:
+        try:
+            # 1. Snap a picture
+            print("\n" + "-"*40)
+            print("Say cheese! 📸 Capturing image...")
+            frame = capture_frame()
+            
+            if frame is None:
+                print("[!] Aborting loop due to camera error.")
+                break
+
+            # VILA expects a file path for the image, so we save it temporarily.
+            # cv2 writes images as BGR, but that's fine for standard file formats like jpg.
+            with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp:
+                temp_filename = tmp.name
+                cv2.imwrite(temp_filename, frame)
+            
+            print(f"[*] Image saved to temporary file: {temp_filename}")
+
+            # 2. Build the prompt
+            prompt_parts = []
+            
+            # The model expects our custom Image object wrapper
+            media_obj = Image(temp_filename)
+            prompt_parts.append(media_obj)
+            prompt_parts.append(args.text)
+
+            # 3. Run Inference
+            print("[*] Model is thinking...")
+            start_time = time.time()
+            
+            # generate_content handles the tokenization and model forward pass
+            response = model.generate_content(prompt_parts)
+            
+            end_time = time.time()
+            duration = end_time - start_time
+
+            # 4. Display result
+            print("\n" + "="*10 + " VILA RESPONSE " + "="*10)
+            print(colored(response, "green", attrs=["bold"]))
+            print("="*35)
+            print(f"[*] Inference took {duration:.2f} seconds")
+
+            # Cleanup the temp file so we don't fill up the drive
+            if os.path.exists(temp_filename):
+                os.remove(temp_filename)
+
+            # Loop control
+            if not args.run_loop:
+                break
+            
+            print(f"\nWaiting {args.loop_delay} seconds before next capture... (Press Ctrl+C to stop)")
+            time.sleep(args.loop_delay)
+
+        except KeyboardInterrupt:
+            print("\n[*] Stopping demo. Goodbye!")
+            # Clean up usage if we exit mid-loop
+            if 'temp_filename' in locals() and os.path.exists(temp_filename):
+                os.remove(temp_filename)
+            break
+        except Exception as e:
+            print(f"\n[!] Unexpected error: {e}")
+            break
+
+def main():
+    parser = argparse.ArgumentParser(description="Run VILA inference live from your webcam.")
+    
+    parser.add_argument("--model-path", "-m", type=str, required=True, 
+                        help="Path to the model directory (e.g. VILA1.5-3B)")
+    parser.add_argument("--lora-path", "-l", type=str, default=None, 
+                        help="Optional path to a LoRA checkpoint")
+    parser.add_argument("--conv-mode", "-c", type=str, default="auto", 
+                        help="Conversation mode/template (default: auto)")
+    parser.add_argument("--text", type=str, default="Describe what you see in the image.", 
+                        help="The prompt to send to the VILA model")
+    parser.add_argument("--run-loop", action="store_true", 
+                        help="Run continuously in a loop")
+    parser.add_argument("--loop-delay", type=float, default=3.0, 
+                        help="Wait time (seconds) between captures in loop mode")
+
+    args = parser.parse_args()
+
+    # Load the model
+    print(f"[*] Loading model from: {args.model_path}")
+    print("[*] This might take a minute, especially for larger models...")
+    
+    try:
+        # Load the unified VILA model
+        if args.lora_path:
+            model = llava.load(args.lora_path, model_base=args.model_path)
+        else:
+            model = llava.load(args.model_path, model_base=None)
+        print("[*] Model loaded successfully!")
+    except Exception as e:
+        print(f"[!] Failed to load model. Is the path correct?\nError: {e}")
+        return
+
+    # Start the application loop
+    processing_loop(model, args)
+
+if __name__ == "__main__":
+    main()