Update app.py

app.py

@@ -3,7 +3,7 @@ import gradio as gr
 from PIL import Image
 from transformers import BlipProcessor, BlipForConditionalGeneration, pipeline
 
-# Load BLIP for image captioning (slow processor avoids torchvision dependency)
+# Initialize BLIP for image captioning (slow mode avoids torchvision dependency)
 blip_processor = BlipProcessor.from_pretrained(
     "Salesforce/blip-image-captioning-base",
     use_fast=False
@@ -12,33 +12,18 @@ blip_model = BlipForConditionalGeneration.from_pretrained(
     "Salesforce/blip-image-captioning-base"
 )
 
-# Hugging Face pipelines (all using Flan-T5-small for speed, temperature=1.0)
-category_generator = pipeline(
+# Flan-T5-small pipelines (temperature=1 for diversity, max_new_tokens increased for depth)
+gen_pipeline = lambda model_name, tokens: pipeline(
     "text2text-generation",
-    model="google/flan-t5-small",
-    tokenizer="google/flan-t5-small",
-    max_new_tokens=50,
-    do_sample=True,
-    temperature=1.0
-)
-
-analysis_generator = pipeline(
-    "text2text-generation",
-    model="google/flan-t5-small",
-    tokenizer="google/flan-t5-small",
-    max_new_tokens=500,
-    do_sample=True,
-    temperature=1.0
-)
-
-suggestion_generator = pipeline(
-    "text2text-generation",
-    model="google/flan-t5-small",
-    tokenizer="google/flan-t5-small",
-    max_new_tokens=500,
+    model=model_name,
+    tokenizer=model_name,
+    max_new_tokens=tokens,
     do_sample=True,
     temperature=1.0
 )
+category_generator = gen_pipeline("google/flan-t5-small", 100)
+analysis_generator = gen_pipeline("google/flan-t5-small", 500)
+suggestion_generator = gen_pipeline("google/flan-t5-small", 500)
 
 # Example URLs for gallery
 def get_recommendations():
@@ -55,42 +40,47 @@ def get_recommendations():
         "https://i.imgur.com/Xj92Cjv.jpeg",
     ]
 
-# Generate BLIP caption from image
+# Step 1: BLIP caption from image
 def generate_caption(image):
     inputs = blip_processor(images=image, return_tensors="pt")
     outputs = blip_model.generate(**inputs)
     return blip_processor.decode(outputs[0], skip_special_tokens=True)
 
-# Generate concise category via Flan
+# Step 2: Flan interprets caption into a category label
 def generate_category(caption):
     prompt = (
         f"Caption: {caption}\n"
-        "Provide a concise category label for this ad."
+        "Provide a concise category label for this ad (e.g. 'Food Ad', 'Fitness Promotion')."
     )
     raw = category_generator(prompt)[0]["generated_text"].strip()
     return raw.splitlines()[0]
 
-# Produce 5-sentence analysis via Flan
+# Step 3: Flan produces a five-sentence analysis of the caption
 def generate_analysis(caption):
     prompt = (
         f"Caption: {caption}\n"
-        "Write exactly five sentences explaining what this ad conveys and its key message."
+        "Write exactly five sentences explaining what the ad conveys, its core message, and its emotional impact."
     )
     raw = analysis_generator(prompt)[0]["generated_text"].strip()
     sentences = re.split(r'(?<=[.!?])\s+', raw)
     return " ".join(sentences[:5])
 
-# Suggest 5 bullet-point improvements via Flan
+# Step 4: Flan suggests five bullet-point improvements
 def generate_suggestions(caption):
     prompt = (
         f"Caption: {caption}\n"
-        "Suggest five distinct improvement points for this ad, formatted as a bullet list starting each line with '- '."
+        "Suggest five distinct improvements for this ad as a bullet list. "
+        "Each line must start with '- ' and describe one actionable change."
     )
     raw = suggestion_generator(prompt)[0]["generated_text"].strip()
-    lines = [l for l in raw.splitlines() if l.strip().startswith('-')]
-    return "\n".join(lines[:5]) if lines else "\n".join(raw.splitlines()[:5])
-
-# Full pipeline combining all steps
+    lines = [line for line in raw.splitlines() if line.strip().startswith('- ')]
+    # ensure exactly five bullets
+    if len(lines) < 5:
+        fallback = [line for line in raw.splitlines() if line.strip()]
+        lines = ['- ' + fallback[i] if not fallback[i].startswith('- ') else fallback[i] for i in range(min(5, len(fallback)))]
+    return "\n".join(lines[:5])
+
+# Full workflow
 def process(image):
     caption = generate_caption(image)
     category = generate_category(caption)
@@ -99,11 +89,11 @@ def process(image):
     recs = get_recommendations()
     return category, analysis, suggestions, recs
 
-# UI Layout using Gradio
+# Gradio UI
 with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as demo:
     gr.Markdown("## 📢 Smart Ad Analyzer")
     gr.Markdown(
-        "Upload an image ad to see: an Ad Category label, a five-sentence analysis, five bullet-point improvements, and example ads."
+        "Upload an image ad to see: a category, five-sentence analysis, five bullet-point improvements, and example ads."
     )
 
     with gr.Row():