Aarnb/Image_Title_as_Nandalal

Use this to generate Title of the image in Nandalal Bose style

import torch
from transformers import BlipProcessor, BlipForConditionalGeneration
from PIL import Image
from huggingface_hub import login

# Optional: Login if your repo is Private. If Public, you can skip this.
# HF_TOKEN = "hf_..."
# login(token=HF_TOKEN)

MODEL_REPO_ID = "Aarnb/Image_Title_as_Nandalal"
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

# LOAD MODEL & PROCESSOR
print(f"Loading model from {MODEL_REPO_ID}...")

# Load the processor (handles image resizing and normalization)
processor = BlipProcessor.from_pretrained(MODEL_REPO_ID)

# Load the model (the weights you trained)
model = BlipForConditionalGeneration.from_pretrained(MODEL_REPO_ID).to(device)
model.eval() # Set to evaluation mode

print("Model loaded successfully!")

# PREDICTION FUNCTION
def predict_description(image_path):
    """
    Takes an image path and returns the generated description.
    """
    try:
        raw_image = Image.open(image_path).convert('RGB')

        # Note: We do NOT pass text here, because we want the model to generate it.
        inputs = processor(images=raw_image, return_tensors="pt").to(device)

        # max_new_tokens: Controls how long the generated description is
        outputs = model.generate(
            **inputs,
            max_new_tokens=50,
            min_length=5,
            num_beams=5, # Using beam search for better quality
            early_stopping=True
        )

        description = processor.decode(outputs[0], skip_special_tokens=True)

        return description

    except Exception as e:
        return f"Error: {e}"


# Replace this with the path to the image you want to test
img_path = "img26.jpg"


print("-" * 30)
print(f"Analyzing: {img_path}")
result = predict_description(img_path)
print(f"Predicted Description: {result}")
print("-" * 30)