import torch
from transformers import BlipForConditionalGeneration
from transformers import BlipProcessor

processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base", torch_dtype=torch.float32).to("cpu")

def caption_image(raw_image):
    inputs = processor(raw_image.convert('RGB'), return_tensors="pt").to("cpu", torch.float32)
    out = model.generate(**inputs, max_new_tokens=100)
    return processor.decode(out[0], skip_special_tokens=True)