from transformers import BlipProcessor, BlipForConditionalGeneration
import torch

# Use GPU if available, otherwise fallback to CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load processor and model
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(device)

@torch.no_grad()
def generate_caption(image):
    inputs = processor(images=image, return_tensors="pt").to(device)
    output = model.generate(**inputs, max_new_tokens=50)
    caption = processor.tokenizer.decode(output[0], skip_special_tokens=True)
    return caption