from transformers import Blip2ForConditionalGeneration, Blip2Processor
from PIL import Image
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

processor = Blip2Processor.from_pretrained("kkatiz/THAI-BLIP-2")
model = Blip2ForConditionalGeneration.from_pretrained("kkatiz/THAI-BLIP-2", device_map=device, torch_dtype=torch.bfloat16)

img = Image.open("Your image...")
inputs = processor(images=img, return_tensors="pt").to(device, torch.bfloat16)

# Adjust your `max_length`
generated_ids = model.generate(**inputs, max_length=20)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_text)