Update app.py

app.py

@@ -4,19 +4,26 @@ import torch
 from PIL import Image
 from transformers import AutoProcessor, AutoModelForCausalLM
 
-
-
 subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
 florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
 
-def generate_caption(image):
+# Modify the generate_caption function to accept a user_prompt
+def generate_caption(image, user_prompt_text=""):
     if not isinstance(image, Image.Image):
         image = Image.fromarray(image)
     
-    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
+    # Florence-2 uses specific task prompts. 
+    # Combining a user prompt directly might not work as expected for its internal tasks.
+    # However, we can try to append it or use it to filter/refine the output later.
+    
+    # For Florence-2's internal mechanism, we still use its task prompt.
+    # The user_prompt_text will be used *after* Florence-2 generates its raw description.
+    florence_task_prompt = "<MORE_DETAILED_CAPTION>"
+    
+    inputs = florence_processor(text=florence_task_prompt, images=image, return_tensors="pt").to(device)
     generated_ids = florence_model.generate(
         input_ids=inputs["input_ids"],
         pixel_values=inputs["pixel_values"],
@@ -26,20 +33,39 @@ def generate_caption(image):
         num_beams=3,
     )
     generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
-    parsed_answer = florence_processor.post_process_generation(
+    raw_caption = florence_processor.post_process_generation(
         generated_text,
-        task="<MORE_DETAILED_CAPTION>",
+        task=florence_task_prompt,
         image_size=(image.width, image.height)
-    )
-    prompt =  parsed_answer["<MORE_DETAILED_CAPTION>"]
-    print("\n\nGeneration completed!:"+ prompt)
-    return prompt
+    )[florence_task_prompt]
     
-io = gr.Interface(generate_caption,
-                  inputs=[gr.Image(label="Input Image")],
-                  outputs = [gr.Textbox(label="Output Prompt", lines=3, show_copy_button = True),
-                            ],
-                  theme="Yntec/HaleyCH_Theme_Orange",
-                  description="⚠ Sorry for the inconvenience. The space are currently running on the CPU, which might affect performance. We appreciate your understanding."
-                 )
+    print(f"\n\nRaw Florence-2 Generation: {raw_caption}")
+
+    # --- POST-PROCESSING STEP ---
+    # Now, we use a separate LLM (like the one I'm operating as) to refine the raw_caption
+    # based on the user_prompt_text. This is crucial because Florence-2 itself
+    # isn't designed for arbitrary stylistic prompting like "focus on clothes, age, footwear".
+    
+    # In a real deployed app, you'd integrate another API call here (e.g., to OpenAI, Gemini, etc.)
+    # For this example, I'll simulate it by returning both.
+    
+    # You would replace this with an actual call to another LLM to refine 'raw_caption'
+    # using 'user_prompt_text' as a guide.
+    refined_prompt_output = f"Original Caption: {raw_caption}\n\nRefinement Request: {user_prompt_text}\n\n(Note: A secondary AI would process this for your desired output.)"
+    
+    return raw_caption, refined_prompt_output # Return both for demonstration
+
+io = gr.Interface(
+    generate_caption,
+    inputs=[
+        gr.Image(label="Input Image"),
+        gr.Textbox(label="Refinement Prompt (e.g., 'focus on clothes, age, hair color, footwear')", lines=2, placeholder="Optional: Describe specific focus areas for refinement.") # New text input
+    ],
+    outputs=[
+        gr.Textbox(label="Raw Florence-2 Caption", lines=3, show_copy_button=True),
+        gr.Textbox(label="Refined Output (Requires Secondary AI)", lines=5, show_copy_button=True) # Output for the refined prompt
+    ],
+    theme="Yntec/HaleyCH_Theme_Orange",
+    description="⚠ Sorry for the inconvenience. The space are currently running on the CPU, which might affect performance. We appreciate your understanding."
+)
 io.launch(debug=True)