from transformers import pipeline
from PIL import Image

# 깊이 추정 파이프라인
pipe = pipeline("depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf")

image = Image.open("image.jpg")
result = pipe(image)

depth_map = result["depth"]  # PIL 이미지
depth_map.save("depth_output.png")

import torch
import cv2

# MiDaS 모델 로드
model = torch.hub.load("intel-isl/MiDaS", "DPT_Large")
midas_transforms = torch.hub.load("intel-isl/MiDaS", "transforms")
transform = midas_transforms.dpt_transform

model.eval()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

# 추론
image = cv2.imread("image.jpg")
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
input_tensor = transform(image_rgb).to(device)

with torch.no_grad():
    prediction = model(input_tensor)
    prediction = torch.nn.functional.interpolate(
        prediction.unsqueeze(1),
        size=image.shape[:2],
        mode="bicubic",
    ).squeeze()

depth_map = prediction.cpu().numpy()