make it possible to edit system prompt

app.py

@@ -20,9 +20,7 @@ MAX_NEW_TOKENS = 2048
 TEMPERATURE = 1.0
 MIN_P = 0.1
 SYSTEM_PROMPT = """You are a helpful assistant for privacy analysis of images. Please always answer in English. Please obey the users instructions and follow the provided format."""
-DEFAULT_PROMPT = """
-
-You are an expert at pixel perfect image analysis and in privacy. Your task is to find all private data in the image and report its position, as well as explanations as to why it is private data. Private data is all data that relates to a unique person and can be used to identify them.
+DEFAULT_PROMPT = """You are an expert at pixel perfect image analysis and in privacy. Your task is to find all private data in the image and report its position, as well as explanations as to why it is private data. Private data is all data that relates to a unique person and can be used to identify them.
 
 First write down your thoughts within a <think> block.
 Please go through all objects in the image and consider whether they are private data or not.
@@ -52,9 +50,11 @@ Here is the image to analyse, start your analysis directly after:
 """
 
 
-def build_messages(image, history: Optional[list[dict[str, Any]]] = None, prompt: Optional[str] = None):
+def build_messages(image, history: Optional[list[dict[str, Any]]] = None, prompt: Optional[str] = None, system_prompt_text: Optional[str] = None):
     if not prompt:
         prompt = DEFAULT_PROMPT
+    if not system_prompt_text:
+        system_prompt_text = SYSTEM_PROMPT # Fallback if not provided
 
     if history:
         return [
@@ -68,7 +68,7 @@ def build_messages(image, history: Optional[list[dict[str, Any]]] = None, prompt
             "content": [
                 {
                     "type": "text",
-                    "text": SYSTEM_PROMPT,
+                    "text": system_prompt_text, # Use the passed system prompt
                 }
             ],
         },
@@ -118,7 +118,7 @@ def anonymise_image(input_image_np: np.ndarray, boxes: list[BoundingBox]):
         raise gr.Error(f"Failed to initialize Blurnonymizer. Check logs. Error: {e}")
 
 
-def run_model_inference(input_image_pil: Image.Image, prompt_text: str):
+def run_model_inference(input_image_pil: Image.Image, prompt_text: str, system_prompt_text: str):
     """
     Runs model inference on the input image and prompt.
     """
@@ -127,7 +127,9 @@ def run_model_inference(input_image_pil: Image.Image, prompt_text: str):
     print("Running model inference...")
     messages = build_messages(
         input_image_pil,
-        prompt=prompt_text)
+        prompt=prompt_text,
+        system_prompt_text=system_prompt_text # Pass system prompt here
+    )
 
     input_text = tokenizer.apply_chat_template(
         messages, tokenize=False, add_generation_prompt=True
@@ -176,7 +178,7 @@ def run_model_inference(input_image_pil: Image.Image, prompt_text: str):
 
 
 @spaces.GPU(duration=90)  # Request GPU for this function, allow up to 120 seconds
-def analyze_image(input_image_pil: Image.Image, prompt_text: str):
+def analyze_image(input_image_pil: Image.Image, prompt_text: str, system_prompt_text: str):
     """
     Analyzes the input image using the VLM, visualizes findings, and anonymizes.
     """
@@ -184,9 +186,11 @@ def analyze_image(input_image_pil: Image.Image, prompt_text: str):
         raise gr.Error("Please upload an image.")
     if not prompt_text:
         raise gr.Error("Please provide a prompt.")
+    if not system_prompt_text:
+        raise gr.Error("Please provide a system prompt.") # Added check
 
     try:
-        raw_model_output, image_height, image_width = run_model_inference(input_image_pil, prompt_text)
+        raw_model_output, image_height, image_width = run_model_inference(input_image_pil, prompt_text, system_prompt_text) # Pass system prompt
     except Exception as e:
         print(f"Error during model inference: {e}")
         print(traceback.format_exc())
@@ -196,6 +200,7 @@ def analyze_image(input_image_pil: Image.Image, prompt_text: str):
 
     return raw_model_output, visualized_image_np, anonymized_image_np
 
+
 @spaces.GPU(duration=90)
 def perform_anonymisation(input_image_pil: Image.Image, raw_model_output: str) -> tuple[np.ndarray, np.ndarray]:
     original_image_np = np.array(input_image_pil)
@@ -276,6 +281,7 @@ with gr.Blocks() as demo:
     with gr.Row():
         with gr.Column(scale=1):
             input_image = gr.Image(type="pil", label="Upload Image")
+            system_prompt_input = gr.Textbox(label="System Prompt", value=SYSTEM_PROMPT, lines=5, interactive=True) # New system prompt input
             prompt_textbox = gr.Textbox(
                 label="Analysis Prompt", value=DEFAULT_PROMPT, lines=4
             )
@@ -295,7 +301,7 @@ with gr.Blocks() as demo:
 
     analyze_button.click(
         fn=analyze_image,
-        inputs=[input_image, prompt_textbox],
+        inputs=[input_image, prompt_textbox, system_prompt_input], # Add system_prompt_input here
         outputs=[raw_output, output_visualized, output_anonymized],
     )
 

utils.py

@@ -361,7 +361,7 @@ def visualize_boxes_annotated(image: np.ndarray | Image.Image, boxes: list[Bound
         ax.add_patch(rect)
 
         # Add label text above the box
-        ax.text(x_min, y_min-5, label, color=color, fontsize=5,
+        ax.text(x_min, y_min-5, label, color=color, fontsize=4,
                 bbox=dict(facecolor='white', alpha=0.7, edgecolor='none'))
 
     # Instead of displaying, save to numpy array