import torch from PIL import Image import gradio as gr import easyocr import open_clip import cv2 import numpy as np import re from concurrent.futures import ThreadPoolExecutor, as_completed from sklearn.cluster import KMeans from transformers import ( BlipProcessor, BlipForConditionalGeneration, BlipForQuestionAnswering ) # === Device setup === device = torch.device("cuda" if torch.cuda.is_available() else "cpu") torch.backends.cudnn.benchmark = True print(f"Using device: {device}") # === Model initialization === # BLIP for captions blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large") blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large") blip_model.to(device).eval() # BLIP for VQA (optional) try: blip_vqa_processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base") blip_vqa_model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base") blip_vqa_model.to(device).eval() VQA_AVAILABLE = True except Exception as e: print(f"VQA model not available: {e}") VQA_AVAILABLE = False if device.type == "cuda": blip_model = blip_model.half() if VQA_AVAILABLE: blip_vqa_model = blip_vqa_model.half() # CLIP for zero-shot classification clip_model, _, clip_preprocess = open_clip.create_model_and_transforms("ViT-L-14", pretrained="openai") clip_model.to(device).eval() if device.type == "cuda": clip_model = clip_model.half() # OCR reader ocr_reader = easyocr.Reader(["en", "es", "fr", "de", "it"], gpu=(device.type == "cuda")) # === Label sets (unchanged) === COMPREHENSIVE_PRODUCT_LABELS = [ "trucker hat", "baseball cap", "beanie", "cap", "hat", "sunglasses", "watch", "jewelry", "necklace", "bracelet", "ring", "earrings", "bag", "backpack", "purse", "wallet", "belt", "scarf", "gloves", "shoes", "sneakers", "boots", "sandals", "heels", "smartphone", "laptop", "tablet", "headphones", "earbuds", "speaker", "camera", "gaming console", "controller", "keyboard", "mouse", "monitor", "TV", "remote", "charger", "cable", "smartwatch", "fitness tracker", "drone", "robot", "board game", "card game", "puzzle", "action figure", "doll", "toy car", "LEGO", "building blocks", "stuffed animal", "teddy bear", "toy robot", "game controller", "dice", "playing cards", "collectible card", "miniature figure", "model kit", "perfume bottle", "cologne", "lipstick", "mascara", "foundation", "powder compact", "nail polish", "skincare bottle", "lotion", "cream jar", "soap", "shampoo", "conditioner", "toothbrush", "razor", "makeup brush", "mirror", "hair dryer", "water bottle", "soda can", "beer bottle", "wine bottle", "coffee mug", "tea cup", "food package", "cereal box", "snack bag", "candy bar", "chocolate", "cookie package", "book", "magazine", "comic book", "DVD", "Blu-ray", "CD", "vinyl record", "notebook", "journal", "planner", "calendar", "poster", "artwork", "photo frame", "basketball", "football", "soccer ball", "tennis ball", "baseball", "golf ball", "dumbbells", "yoga mat", "water bottle", "sports equipment", "helmet", "gloves", "coffee maker", "blender", "toaster", "microwave", "pan", "pot", "plate", "bowl", "cup", "glass", "utensils", "knife", "cutting board", "storage container", "paint brush", "marker", "pen", "pencil", "crayon", "paint tube", "canvas", "sketchbook", "ruler", "scissors", "glue", "tape", "stickers", "box", "package", "container", "jar", "bottle", "can", "tube", "pouch", "bag" ] COMPREHENSIVE_BRAND_LABELS = [ "Apple", "Samsung", "Google", "Microsoft", "Sony", "Nintendo", "PlayStation", "Xbox", "HP", "Dell", "Lenovo", "ASUS", "Acer", "Canon", "Nikon", "GoPro", "DJI", "Nike", "Adidas", "Puma", "Under Armour", "Reebok", "Converse", "Vans", "New Balance", "Ray-Ban", "Oakley", "Rolex", "Casio", "Fossil", "Michael Kors", "Coach", "Louis Vuitton", "L'Oreal", "Maybelline", "MAC", "Sephora", "Clinique", "Estee Lauder", "Chanel", "Dior", "YSL", "Urban Decay", "Too Faced", "NARS", "Fenty Beauty", "Hasbro", "Mattel", "LEGO", "Ravensburger", "Fantasy Flight Games", "Asmodee", "Days of Wonder", "CMON", "Wizards of the Coast", "Pokémon", "Yu-Gi-Oh", "Coca-Cola", "Pepsi", "Starbucks", "Red Bull", "Monster", "Gatorade", "Powerade", "Nestle", "Hershey's", "Mars", "Ferrero", "Lindt", "Godiva", "Crayola", "Sharpie", "Prismacolor", "Winsor & Newton", "Molotow", "Posca", "Staedtler", "Faber-Castell", "Sakura", "Tombow", "Pilot", "Uni-ball", "unknown", "generic", "no brand visible" ] CHARACTER_LABELS = [ "person", "man", "woman", "child", "baby", "elderly person", "athlete", "professional", "cartoon character", "anime character", "superhero", "mascot", "celebrity", "model", "fantasy character", "sci-fi character", "historical figure", "fictional character" ] DETAIL_LABELS = [ "text", "logo", "brand name", "price tag", "barcode", "QR code", "warning label", "ingredients list", "nutritional information", "instructions", "serial number", "model number", "size information", "color variants", "pattern", "texture", "metallic finish", "matte finish", "glossy finish", "transparent", "translucent" ] # === Basic color palette for mapping === BASE_COLORS = { 'red': (255, 0, 0), 'green': (0, 128, 0), 'blue': (0, 0, 255), 'yellow': (255, 255, 0), 'orange': (255, 165, 0), 'purple': (128, 0, 128), 'pink': (255, 192, 203), 'brown': (165, 42, 42), 'gray': (128, 128, 128), 'black': (0, 0, 0), 'white': (255, 255, 255), } def get_basic_color_name(rgb_tuple): """Map any RGB to nearest basic color name.""" best, best_name = float('inf'), 'unknown' for name, ref in BASE_COLORS.items(): dist = sum((a - b) ** 2 for a, b in zip(rgb_tuple, ref)) ** 0.5 if dist < best: best, best_name = dist, name return best_name def extract_dominant_colors(image: Image.Image, n_colors=3): """Use k-means on downsampled pixels, then map to basic colors.""" arr = np.array(image.resize((150, 150))).reshape(-1, 3) # drop pure black/white mask = ~((arr == [255,255,255]).all(axis=1) | (arr == [0,0,0]).all(axis=1)) arr = arr[mask] if mask.sum() > 0 else arr if len(arr) < n_colors: return [] kmeans = KMeans(n_clusters=n_colors, random_state=42, n_init=10).fit(arr) colors = [] for center in kmeans.cluster_centers_: rgb = tuple(map(int, center)) colors.append(get_basic_color_name(rgb)) # dedupe and preserve order seen = set() result = [] for c in colors: if c not in seen: seen.add(c) result.append(c) return result # === Zero-shot classification setup === def precompute_embeddings(): global _prod_emb, _brand_emb, _char_emb, _detail_emb prompts_prod = [f"a photo of a {l}" for l in COMPREHENSIVE_PRODUCT_LABELS] prompts_brand = [f"a {b} product" for b in COMPREHENSIVE_BRAND_LABELS] prompts_char = [f"a photo of a {c}" for c in CHARACTER_LABELS] prompts_det = [f"an image showing {d}" for d in DETAIL_LABELS] with torch.inference_mode(): _prod_emb = clip_model.encode_text(open_clip.tokenize(prompts_prod).to(device)) _prod_emb /= _prod_emb.norm(dim=-1, keepdim=True) _brand_emb = clip_model.encode_text(open_clip.tokenize(prompts_brand).to(device)) _brand_emb /= _brand_emb.norm(dim=-1, keepdim=True) _char_emb = clip_model.encode_text(open_clip.tokenize(prompts_char).to(device)) _char_emb /= _char_emb.norm(dim=-1, keepdim=True) _det_emb = clip_model.encode_text(open_clip.tokenize(prompts_det).to(device)) _det_emb /= _det_emb.norm(dim=-1, keepdim=True) def enhanced_zero_shot_classification(image: Image.Image): with torch.inference_mode(): img_t = clip_preprocess(image).unsqueeze(0).to(device) img_e = clip_model.encode_image(img_t) img_e /= img_e.norm(dim=-1, keepdim=True) results = {} for labels, emb, key, topk in [ (COMPREHENSIVE_PRODUCT_LABELS, _prod_emb, 'products', 3), (COMPREHENSIVE_BRAND_LABELS, _brand_emb, 'brands', 3), (CHARACTER_LABELS, _char_emb, 'characters', 3), (DETAIL_LABELS, _det_emb, 'details', 5), ]: logits = (100.0 * img_e @ emb.T) probs = torch.softmax(logits, dim=-1) top = torch.topk(probs, topk, dim=-1) results[key] = [ (labels[idx], float(prob)) for prob, idx in zip(top.values[0], top.indices[0]) ] return results # === BLIP captioning with enhanced prompts === DETAILED_PROMPTS = [ "Provide a detailed, comprehensive description of this image including all visual elements, characters, objects, text, colors, setting, and artistic style.", "Describe this image in great detail, mentioning the title, characters' appearance, clothing, expressions, objects they're holding, background elements, and any text visible.", "Write a thorough description of this cover art or product image, including character details, clothing, facial expressions, objects, background setting, colors, and any visible text or branding.", "Analyze this image completely: describe the main subject, their appearance and clothing, any objects or accessories, the background setting, color scheme, artistic style, and all visible text." ] def get_enhanced_caption(image: Image.Image): captions = [] # unconditional baseline try: inputs = blip_processor(image, return_tensors="pt").to(device) out = blip_model.generate( **inputs, max_new_tokens=150, num_beams=8, do_sample=True, temperature=0.8, top_p=0.9, no_repeat_ngram_size=2, repetition_penalty=1.1 ) captions.append(blip_processor.decode(out[0], skip_special_tokens=True).strip()) except Exception as e: print(f"Baseline caption error: {e}") # detailed prompts for prompt in DETAILED_PROMPTS: try: inputs = blip_processor(image, text=prompt, return_tensors="pt").to(device) out = blip_model.generate( **inputs, max_new_tokens=200, num_beams=8, do_sample=True, temperature=0.7, top_p=0.85, no_repeat_ngram_size=3, repetition_penalty=1.15 ) captions.append(blip_processor.decode(out[0], skip_special_tokens=True).strip()) except Exception as e: print(f"Prompt caption error: {e}") if not captions: return "Unable to generate a caption." # pick the most content-rich (longest) caption best = max(captions, key=lambda s: len(s.split())) best = re.sub(r'^(a|an|the)\s+', '', best, flags=re.IGNORECASE).strip().capitalize() return best if best.endswith('.') else best + '.' # === OCR === def extract_comprehensive_ocr(image: Image.Image): try: arr = np.array(image) gray = cv2.cvtColor(arr, cv2.COLOR_RGB2GRAY) preproc = [arr] clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8)).apply(gray) preproc.append(cv2.cvtColor(clahe, cv2.COLOR_GRAY2RGB)) _, th = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU) preproc.append(cv2.cvtColor(th, cv2.COLOR_GRAY2RGB)) texts = [] for p in preproc: for _, txt, conf in ocr_reader.readtext(p, detail=1, paragraph=True): if conf > 0.3 and len(txt.strip()) > 1: texts.append((txt.strip(), conf)) # dedupe seen = set() unique = [] for txt, conf in sorted(texts, key=lambda x: x[1], reverse=True): key = re.sub(r'\W+', '', txt.lower()) if key not in seen: seen.add(key) unique.append(txt) return " | ".join(unique[:5]) except Exception as e: print(f"OCR error: {e}") return "" # === Comprehensive pipeline === def comprehensive_image_analysis(image: Image.Image): # resize for efficiency if max(image.size) > 1024: image = image.copy() image.thumbnail((1024,1024), Image.Resampling.LANCZOS) return { 'caption': get_enhanced_caption(image), 'classification': enhanced_zero_shot_classification(image), 'ocr_text': extract_comprehensive_ocr(image), 'colors': extract_dominant_colors(image) } # === Build narrative alt-text === def build_narrative_alt_text(analysis): main_desc = analysis['caption'] parts = [] # OCR text ocr = analysis['ocr_text'] if ocr: parts.append(f'Visible text: "{ocr}"') # Dominant colors cols = analysis['colors'] if cols: if len(cols) == 1: parts.append(f"The primary color is {cols[0]}") else: col_str = ", ".join(cols[:-1]) + f", and {cols[-1]}" parts.append(f"The color palette features {col_str}") # Brand (only if high confidence) brands = analysis['classification']['brands'] if brands and brands[0][1] > 0.6: parts.append(f"Published by {brands[0][0]}") # Combine final = main_desc for p in parts: final += " " + p + "." final = re.sub(r'\s+', ' ', final).strip() if not final.endswith('.'): final += '.' return final # === Initialize embeddings === print("Precomputing embeddings...") precompute_embeddings() print("Embeddings ready!") # === Gradio interface === def handle_single_image(file): try: img = Image.open(file.name).convert("RGB") except Exception as e: return f"Failed to open image: {e}" try: analysis = comprehensive_image_analysis(img) return build_narrative_alt_text(analysis) except Exception as e: return f"Analysis failed: {e}" def preview_images_grid(files): previews = [] for f in (files or [])[:12]: try: img = Image.open(f.name).convert("RGB") img.thumbnail((200,200), Image.Resampling.LANCZOS) previews.append(img) except: continue return previews def process_images(files): if not files: return "No images uploaded." results = [] with ThreadPoolExecutor(max_workers=1) as exe: futures = {exe.submit(handle_single_image, f): i for i, f in enumerate(files[:8])} for fut in as_completed(futures): idx = futures[fut] try: results.append(f"**Image {idx+1}:**\n{fut.result()}\n") except Exception as e: results.append(f"**Image {idx+1}:** Failed - {e}\n") return "\n".join(sorted(results, key=lambda s: int(re.search(r"Image (\d+):", s).group(1)))) with gr.Blocks( title="🔍 Optimized Narrative Alt-Text Generator", theme=gr.themes.Soft(), css=""" .gradio-container { max-width:1400px!important; } .description-output { font-family:'Georgia',serif; line-height:1.6; font-size:14px; } .upload-section { border:2px dashed #ccc; border-radius:10px; padding:20px; } """ ) as demo: gr.Markdown(""" # 🖼️ Optimized Alt-Text Generator - Improved basic color mapping - Brand only if high confidence (≥60%) - Removed unreliable VQA-brand/material - K-means → basic colors for clearer palettes """) with gr.Row(): with gr.Column(scale=1, elem_classes=["upload-section"]): inp = gr.File(label="Upload Images", file_count="multiple", file_types=["image"], height=200) btn = gr.Button("Generate Alt-Text", variant="primary", size="lg") with gr.Column(scale=2): gal = gr.Gallery(label="Preview", columns=4, rows=2, height=350, object_fit="contain") out = gr.Textbox(label="Generated Alt-Text", lines=20, show_copy_button=True, elem_classes=["description-output"], placeholder="...") inp.change(preview_images_grid, inp, gal, show_progress="hidden") btn.click(process_images, inp, out, show_progress=True) if __name__ == "__main__": demo.launch(server_name="0.0.0.0", server_port=7860, share=False)