diff --git a/.gitattributes b/.gitattributes
deleted file mode 100644
index 3a24117..0000000
--- a/.gitattributes
+++ /dev/null
@@ -1,2 +0,0 @@
-*.jpg filter=lfs diff=lfs merge=lfs -text
-*.png filter=lfs diff=lfs merge=lfs -text
diff --git a/.github/workflows/cpu-ci.yml b/.github/workflows/cpu-ci.yml
index 672767e..b2c07fd 100644
--- a/.github/workflows/cpu-ci.yml
+++ b/.github/workflows/cpu-ci.yml
@@ -19,17 +19,17 @@ jobs:
     runs-on: ${{ matrix.os }}
 
     steps:
-      - name: Checkout Repository
+      - name: Checkout repository
         uses: actions/checkout@v4
         with:
           lfs: true
 
-      - name: Install Prerequisites for Linux
+      - name: Install prerequisites for Linux
         if: ${{ matrix.os == 'ubuntu-latest' || matrix.os == 'ubuntu-24.04-arm' }}
         run:
           sudo apt update && sudo apt install -y libturbojpeg exiftool ffmpeg libheif-dev poppler-utils
 
-      - name: Install Prerequisites for MacOS
+      - name: Install prerequisites for MacOS
         if: ${{ matrix.os == 'macos-13' || matrix.os == 'macos-latest' }}
         run:
           brew install libjpeg exiftool ffmpeg libheif poppler
@@ -40,18 +40,18 @@ jobs:
           python-version: ${{ matrix.python-version }}
           check-latest: true
 
-      - name: Install Dependencies
+      - name: Install dependencies
         run: |
           python -m pip install -U pip wheel "numpy>=2" "cython>=3.0.12" "setuptools>=69"
           python setup.py build_ext --inplace
           python -m pip install .
 
-      - name: Lint with Pylint
+      - name: Lint with pylint
         run: |
           python -m pip install pylint
           python -m pylint pyface --rcfile=.github/workflows/.pylintrc
 
-      - name: Run Tests with Pytest
+      - name: Run tests with pytest
         run: |
           mkdir -p tests/coverage
           python -m pip install pytest pytest-cov typeguard
diff --git a/.github/workflows/cuda-ci.yml b/.github/workflows/cuda-ci.yml
index c51d6c3..eb403fd 100644
--- a/.github/workflows/cuda-ci.yml
+++ b/.github/workflows/cuda-ci.yml
@@ -7,15 +7,55 @@ on:
       - "doc/**"
       - "**.md"
 
+env:
+  LOCAL_REGISTRY: localhost:5000
+  DOCKERFILE: docker/pr.dockerfile
+
 jobs:
+  get_runner_and_uid:
+    name: Get Runner
+    runs-on: [self-hosted, unicorn]
+    steps:
+      - name: Get UID and GID
+        id: uid_gid
+        run: |
+          echo "uid_gid=$(id -u):$(id -g)" >> "$GITHUB_OUTPUT"
+    outputs:
+      runner: ${{ runner.name }}
+      uid: ${{ steps.uid_gid.outputs.uid_gid }}
+
+  build_docker_image:
+    name: Build Docker Image
+    needs: [get_runner_and_uid]
+    runs-on: ${{ needs.get_runner_and_uid.outputs.runner }}
+
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v4
+
+      - name: Build Docker Image
+        id: docker_build
+        uses: docker/build-push-action@v3
+        with:
+          file: ${{ env.DOCKERFILE }}
+          no-cache: false
+          push: true
+          tags: ${{ env.LOCAL_REGISTRY }}/pyface-container:ci
+
+    outputs:
+      image: ${{ env.LOCAL_REGISTRY }}/pyface-container:ci
+
   ci:
     name: CI
+    needs: [get_runner_and_uid, build_docker_image]
+    runs-on: ${{ needs.get_runner_and_uid.outputs.runner }}
     strategy:
-      fail-fast: false
       matrix:
-        python-version: ["3.10", "3.12"]
-
-    runs-on: ["self-hosted", "unicorn"]
+        python-version:
+          - "3.10"
+    container:
+      image: ${{ needs.build_docker_image.outputs.image }}
+      options: --user ${{ needs.get_runner_and_uid.outputs.uid }} --gpus all
 
     steps:
       - name: Checkout repository
diff --git a/demo/face_depth.py b/demo/face_depth.py
index 791de6f..d2d4ca7 100644
--- a/demo/face_depth.py
+++ b/demo/face_depth.py
@@ -3,10 +3,10 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
-def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
+def main(img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg")):
     face_detect = pf.build_face_detection()
     face_depth = pf.build_face_depth()
 
@@ -16,7 +16,7 @@ def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
     results = face_depth([img], [face_box], return_depth=True)
     plotted = face_depth.draw_results(img, [face_box], results)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_depth1.png")
     cb.imwrite(results[0]["depth_img"], out_folder / "face_depth2.png")
diff --git a/demo/face_detection.py b/demo/face_detection.py
index 33e55ea..48ad861 100644
--- a/demo/face_detection.py
+++ b/demo/face_detection.py
@@ -5,11 +5,11 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     score_th: float = 0.5,
     inp_size: Tuple[int, int] = (480, 640),
 ):
@@ -22,7 +22,7 @@ def main(
     proposals_list = face_detection([img] * 7)
     plotted = face_detection.draw_proposals(img, proposals_list[0])
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_detection.png")
 
diff --git a/demo/face_gender.py b/demo/face_gender.py
index 11d88f4..0fb4ad6 100644
--- a/demo/face_gender.py
+++ b/demo/face_gender.py
@@ -3,10 +3,10 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
-def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
+def main(img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg")):
     face_detect = pf.build_face_detection()
     face_gender = pf.build_gender_detection()
 
@@ -16,7 +16,7 @@ def main(img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg")):
     results = face_gender([img], [face_box])
     plotted = face_gender.draw_results(img, [face_box], results)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_gender.png")
 
diff --git a/demo/face_landmark.py b/demo/face_landmark.py
index 2790a50..77ba853 100644
--- a/demo/face_landmark.py
+++ b/demo/face_landmark.py
@@ -5,11 +5,11 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     score_th: float = 0.5,
     inp_size: Tuple[int, int] = (480, 640),
 ):
@@ -25,7 +25,7 @@ def main(
     result = face_landmark([img], [box])[0]
     plotted = face_landmark.draw_result(img, box, result, plot_details=True)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(plotted, out_folder / "face_landmark.jpg")
 
diff --git a/demo/face_normalization.py b/demo/face_normalization.py
index 78da1b6..5c2ede8 100644
--- a/demo/face_normalization.py
+++ b/demo/face_normalization.py
@@ -5,15 +5,15 @@
 
 import pyface.components as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
     scale: float = 1,
     dst_size: Tuple[int, int] = (224, 224),
 ):
-    face_detection = pf.SCRFD()
+    face_detection = pf.build_face_detection()
     face_normaliation = pf.FaceNormalize(
         dst_size=dst_size,
         interpolation=cb.INTER.BILINEAR,
@@ -22,13 +22,9 @@ def main(
     img = cb.imread(img_path)
     proposals = face_detection([img])[0]
     norm_img = face_normaliation([img], [cb.Keypoints(proposals["lmk5pts"][0])])[0]
-    norm_img = cb.draw_keypoints(
-        norm_img,
-        face_normaliation.destination_pts,
-        scale=1,
-    )
+    norm_img = cb.draw_keypoints(norm_img, face_normaliation.destination_pts, scale=1)
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     cb.imwrite(norm_img, out_folder / f"face_normalize_s={scale}.png")
 
diff --git a/demo/face_recognition.py b/demo/face_recognition.py
index edf2ad3..64dd653 100644
--- a/demo/face_recognition.py
+++ b/demo/face_recognition.py
@@ -3,12 +3,12 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    src_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
-    tgt_path: str = str(cur_folder / "data" / "face_bank" / "EmmaWatson.jpg"),
+    src_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
+    tgt_path: str = str(RESOURCE_DIR / "face_bank" / "EmmaWatson.jpg"),
 ):
     face_detection = pf.build_face_detection()
     face_recognition = pf.build_face_recognition()
@@ -17,7 +17,7 @@ def main(
     tgt_img = cb.imread(tgt_path)
     lmks = [p["lmk5pts"][0] for p in face_detection([src_img, tgt_img])]
 
-    out_folder = cur_folder / "output"
+    out_folder = cb.get_curdir(__file__) / "output"
     out_folder.mkdir(exist_ok=True, parents=True)
     results = face_recognition([src_img, tgt_img], lmks)
     src_emb = results[0]["embeddings"]
diff --git a/demo/face_service.py b/demo/face_service.py
index 01a9a72..0e83300 100644
--- a/demo/face_service.py
+++ b/demo/face_service.py
@@ -3,12 +3,12 @@
 
 import pyface as pf
 
-cur_folder = cb.get_curdir(__file__)
+RESOURCE_DIR = cb.get_curdir(__file__).parent / "tests" / "resources"
 
 
 def main(
-    img_path: str = str(cur_folder / "data" / "EmmaWatson1.jpg"),
-    face_bank: str = str(cur_folder / "data" / "face_bank"),
+    img_path: str = str(RESOURCE_DIR / "EmmaWatson1.jpg"),
+    face_bank: str = str(RESOURCE_DIR / "face_bank"),
 ):
     face_service = pf.FaceService(
         enable_gender=True,
@@ -19,7 +19,9 @@ def main(
     )
     img = cb.imread(img_path)
     faces_on_img = face_service([img], do_1n=True)[0]
-    cb.imwrite(faces_on_img.gen_info_img(), str(cur_folder / "output" / "face_service.jpg"))
+    out_folder = cb.get_curdir(__file__) / "output"
+    out_folder.mkdir(exist_ok=True, parents=True)
+    cb.imwrite(faces_on_img.gen_info_img(), str(out_folder / "face_service.jpg"))
 
 
 if __name__ == "__main__":
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 15c6a8a..72cc3a7 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -1,5 +1,5 @@
 # syntax=docker/dockerfile:experimental
-FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu22.04 as builder
+FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04 as builder
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
     DEBIAN_FRONTEND=noninteractive \
diff --git a/docker/pr.dockerfile b/docker/pr.dockerfile
new file mode 100644
index 0000000..dc60eb1
--- /dev/null
+++ b/docker/pr.dockerfile
@@ -0,0 +1,17 @@
+# syntax=docker/dockerfile:experimental
+FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04 as builder
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    DEBIAN_FRONTEND=noninteractive \
+    TZ=Asia/Taipei
+
+RUN apt-get update -y && apt-get upgrade -y && \
+    apt-get install -y --no-install-recommends \
+    tzdata wget git git-lfs libturbojpeg exiftool ffmpeg poppler-utils libpng-dev \
+    libtiff5-dev libjpeg8-dev libopenjp2-7-dev zlib1g-dev gcc \
+    libfreetype6-dev liblcms2-dev libwebp-dev tcl8.6-dev tk8.6-dev python3-tk \
+    python3-pip libharfbuzz-dev libfribidi-dev libxcb1-dev libfftw3-dev gosu \
+    libpq-dev python3-dev build-essential && \
+    ln -sf /usr/share/zoneinfo/$TZ /etc/localtime && \
+    dpkg-reconfigure -f noninteractive tzdata && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
\ No newline at end of file
diff --git a/pyface/components/face_depth/tddfav2.py b/pyface/components/face_depth/tddfav2.py
index 34b6fa5..2dfbde6 100644
--- a/pyface/components/face_depth/tddfav2.py
+++ b/pyface/components/face_depth/tddfav2.py
@@ -7,7 +7,7 @@
 
 from ..enums import FacePose
 from ..utils import append_to_batch, detach_from_batch, download_model_and_return_model_fpath
-from .Sim3DR import sim3dr_cython
+from .Sim3DR import sim3dr_cython  # pylint: disable=E0611
 
 
 def rasterize(
@@ -168,9 +168,7 @@ def _prepare_bfm_npz(
     return u_base, w_shp_base, w_exp_base
 
 
-def _parse_tffda_param(
-    param: np.ndarray,
-) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+def _parse_tffda_param(param: np.ndarray) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
     """
     matrix pose form
     param: shape = (trans_dim + shape_dim + exp_dim) i.e., 62 = 12 + 40 + 10
@@ -261,16 +259,12 @@ def initialize(self) -> None:
         dummy_inputs = {k: np.zeros(input_dims[k], dtype=v["dtype"]) for k, v in self.bfm_engine.input_infos.items()}
         self.bfm_engine(**dummy_inputs)
 
-    def _transform(
-        self,
-        img: np.ndarray,
-        box: cb.Box,
-    ) -> Tuple[np.ndarray, Tuple[float, float], np.ndarray]:
+    def _transform(self, img: np.ndarray, box: cb.Box) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
         inp_w, inp_h = self.engine.input_infos["input"]["shape"][2:]
 
         expanded_box = box.square().scale(fx=0.98, fy=0.98)
+        scale = np.array([expanded_box.width / inp_w, expanded_box.height / inp_h])
         shift = expanded_box.left_top
-        scale = expanded_box.width / inp_w, expanded_box.height / inp_h
         blob = cb.imcropbox(img, expanded_box)
         blob = cb.imresize(blob, (inp_w, inp_h))
         blob = blob.transpose(2, 0, 1)[None]
@@ -280,7 +274,7 @@ def preprocess(
         self,
         imgs: List[np.ndarray],
         boxes: List[cb.Box],
-    ):
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
         if len(imgs) != len(boxes):
             raise ValueError(f"imgs and boxes should have same length, but got {len(imgs)} and {len(boxes)}")
 
@@ -294,11 +288,7 @@ def preprocess(
             shifts.append(shift)
         return blobs, scales, shifts
 
-    def _reconstruct_vertices(
-        self,
-        param: np.ndarray,
-        dense_flag: bool = False,
-    ) -> np.ndarray:
+    def _reconstruct_vertices(self, param: np.ndarray, dense_flag: bool = False) -> np.ndarray:
         def _similar_transform(pts3d, scales):
             x_s, y_s = scales[..., 1], scales[..., 0]
             z_s = (x_s + y_s) / 2
@@ -374,15 +364,22 @@ def __call__(
         blobs, scales, shifts = self.preprocess(imgs, boxes)
         preds = {k: [] for k in self.engine.output_infos.keys()}
         for batch in cb.make_batch(blobs, self.batch_size):
+            current_batch_size = len(batch)
             inputs = {
                 name: np.concatenate(append_to_batch(batch, self.batch_size)).astype(v["dtype"])  # E1123
                 for name, v in self.engine.input_infos.items()
             }
             tmp_preds = self.engine(**inputs)
             for k, v in tmp_preds.items():
-                preds[k].append(detach_from_batch(v, len(batch)))
+                preds[k].append(detach_from_batch(v, current_batch_size))
         preds = {k: np.concatenate(v, 0) for k, v in preds.items()}
-        params = np.concatenate((preds["params"], scales, shifts), -1)
+        scales = np.stack(scales, 0)  # n x 2
+        shifts = np.stack(shifts, 0)  # n x 2
+        # Ensure shapes are compatible for concatenation
+        n = preds["params"].shape[0]
+        if scales.shape[0] != n or shifts.shape[0] != n:
+            raise ValueError(f"Shape mismatch: preds['params'] has {n} rows, scales has {scales.shape[0]}, shifts has {shifts.shape[0]}")
+        params = np.concatenate((preds["params"], scales, shifts), axis=-1)
         lmk3d68pts = self._gen_3d_landmarks(params)
         pose_degrees = self._get_pose_degrees(params)
 
diff --git a/pyface/components/face_detection/scrfd.py b/pyface/components/face_detection/scrfd.py
index 1eb51f7..ef24dfd 100644
--- a/pyface/components/face_detection/scrfd.py
+++ b/pyface/components/face_detection/scrfd.py
@@ -193,10 +193,11 @@ def __call__(self, imgs: List[np.ndarray]) -> List[dict]:
         preds = {k: [] for k in self.engine.output_infos.keys()}
         b = self.metadata["InputSize"][0]
         for batch in cb.make_batch(blobs, b):
+            current_batch_size = len(batch)
             inputs = {name: np.concatenate(append_to_batch(batch, b)) for name, _ in self.engine.input_infos.items()}
             tmp_preds = self.engine(**inputs)
             for k, v in tmp_preds.items():
-                preds[k].append(detach_from_batch(v, len(batch)))
+                preds[k].append(detach_from_batch(v, current_batch_size))
         preds = {k: np.concatenate(v, 0) for k, v in preds.items()}
         proposals_list = self.postprocess(preds, scales)
         return proposals_list
diff --git a/pyface/components/face_gender.py b/pyface/components/face_gender.py
index aae7d2b..703bcd8 100644
--- a/pyface/components/face_gender.py
+++ b/pyface/components/face_gender.py
@@ -113,10 +113,11 @@ def __call__(self, imgs: List[np.ndarray], boxes: List[cb.Box]) -> List[dict]:
         preds = {k: [] for k in self.engine.output_infos.keys()}
         b = self.metadata["InputSize"][0]
         for batch in cb.make_batch(blobs, b):
+            current_batch_size = len(batch)
             inputs = {name: np.concatenate(append_to_batch(batch, b)) for name, _ in self.engine.input_infos.items()}
             tmp_preds = self.engine(**inputs)
             for k, v in tmp_preds.items():
-                preds[k].append(detach_from_batch(v, len(batch)))
+                preds[k].append(detach_from_batch(v, current_batch_size))
         preds = {k: np.concatenate(v, 0) for k, v in preds.items()}
         preds = self.postprocess(preds)
         return preds
diff --git a/pyface/components/face_recognition/arcface.py b/pyface/components/face_recognition/arcface.py
index b500444..2b090f6 100644
--- a/pyface/components/face_recognition/arcface.py
+++ b/pyface/components/face_recognition/arcface.py
@@ -85,6 +85,7 @@ def __call__(
         preds = {k: [] for k in self.engine.output_infos.keys()}
         b = self.metadata["InputSize"][0]
         for batch in cb.make_batch(blobs, b):
+            current_batch_size = len(batch)
             inputs = {name: np.concatenate(append_to_batch(batch, b)) for name, _ in self.engine.input_infos.items()}
             tmp_preds = self.engine(**inputs)
             if self.enable_flip:
@@ -93,7 +94,7 @@ def __call__(
                 for k, v in tmp_preds.items():
                     tmp_preds[k] = np.concatenate([v, tmp_flip_preds[k]], 1)
             for k, v in tmp_preds.items():
-                preds[k].append(detach_from_batch(v, len(batch)))
+                preds[k].append(detach_from_batch(v, current_batch_size))
         preds = {k: np.concatenate(v, 0) for k, v in preds.items()}
         emgeddings = l2_norm(preds["encode"])
         return [
diff --git a/pyface/utils.py b/pyface/utils.py
index da1ac4b..3597f37 100644
--- a/pyface/utils.py
+++ b/pyface/utils.py
@@ -1,6 +1,3 @@
-import capybara as cb
-import onnxruntime as ort
-
 from .components.face_depth.tddfav2 import TDDFAV2
 from .components.face_detection.scrfd import SCRFD
 from .components.face_gender import GenderDetector
@@ -9,7 +6,6 @@
 
 __all__ = [
     "download_models",
-    "get_ort_backend",
 ]
 
 
@@ -19,22 +15,3 @@ def download_models():
     GenderDetector.download_models()
     CoordinateReg.download_models()
     ArcFace.download_models()
-
-
-def cuda_avaliable():
-    providers = ort.get_available_providers()
-    return "CUDAExecutionProvider" in providers
-
-
-def coreml_avaliable():
-    providers = ort.get_available_providers()
-    return "CoreMLExecutionProvider" in providers
-
-
-def get_ort_backend():
-    if cuda_avaliable():
-        return cb.Backend.cuda
-    elif coreml_avaliable():
-        return cb.Backend.coreml
-    else:
-        return cb.Backend.cpu
diff --git a/tests/resources/EmmaWatson1.jpg b/tests/resources/EmmaWatson1.jpg
new file mode 100644
index 0000000..40919d4
Binary files /dev/null and b/tests/resources/EmmaWatson1.jpg differ
diff --git a/tests/resources/EmmaWatson1.png b/tests/resources/EmmaWatson1.png
deleted file mode 100644
index 9022558..0000000
--- a/tests/resources/EmmaWatson1.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7a06751436d341d1d4389b01db504d35ea477e7a10d80b38951b6d7d8194d650
-size 1257800
diff --git a/tests/resources/JohnnyDepp1.jpg b/tests/resources/JohnnyDepp1.jpg
new file mode 100644
index 0000000..b3b823a
Binary files /dev/null and b/tests/resources/JohnnyDepp1.jpg differ
diff --git a/tests/resources/JohnnyDepp1.png b/tests/resources/JohnnyDepp1.png
deleted file mode 100644
index eea2d5c..0000000
--- a/tests/resources/JohnnyDepp1.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:dd88f4714180e7ef110c2f06c043ac4e2cfd5bd1b4301d2820b114df9a8664b2
-size 1667229
diff --git a/tests/resources/answer/EmmaWatson1_depth.npz b/tests/resources/answer/EmmaWatson1_depth.npz
new file mode 100644
index 0000000..516ca29
Binary files /dev/null and b/tests/resources/answer/EmmaWatson1_depth.npz differ
diff --git a/tests/resources/answer/EmmaWatson1_detection.npz b/tests/resources/answer/EmmaWatson1_detection.npz
new file mode 100644
index 0000000..c9fc45a
Binary files /dev/null and b/tests/resources/answer/EmmaWatson1_detection.npz differ
diff --git a/tests/resources/answer/JohnnyDepp1_depth.npz b/tests/resources/answer/JohnnyDepp1_depth.npz
new file mode 100644
index 0000000..344d751
Binary files /dev/null and b/tests/resources/answer/JohnnyDepp1_depth.npz differ
diff --git a/tests/resources/answer/JohnnyDepp1_detection.npz b/tests/resources/answer/JohnnyDepp1_detection.npz
new file mode 100644
index 0000000..d43145e
Binary files /dev/null and b/tests/resources/answer/JohnnyDepp1_detection.npz differ
diff --git a/tests/resources/face_bank/EmmaWatson.jpg b/tests/resources/face_bank/EmmaWatson.jpg
new file mode 100644
index 0000000..da4d54d
Binary files /dev/null and b/tests/resources/face_bank/EmmaWatson.jpg differ
diff --git a/tests/resources/face_bank/EmmaWatson.png b/tests/resources/face_bank/EmmaWatson.png
deleted file mode 100644
index 9168717..0000000
--- a/tests/resources/face_bank/EmmaWatson.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1abd290e74139db3e22ac9dea2a942f27e626d076dd4375020d4849f5a98125b
-size 1943539
diff --git a/tests/resources/face_bank/JohnnyDepp.jpg b/tests/resources/face_bank/JohnnyDepp.jpg
new file mode 100644
index 0000000..e3a3fa0
Binary files /dev/null and b/tests/resources/face_bank/JohnnyDepp.jpg differ
diff --git a/tests/resources/face_bank/JohnnyDepp.png b/tests/resources/face_bank/JohnnyDepp.png
deleted file mode 100644
index dfc3ecd..0000000
--- a/tests/resources/face_bank/JohnnyDepp.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:898f56977f7a50f6e38e90b94cf67d37567b9b845dc36bf4141fe98ad0fbfe76
-size 1250454
diff --git a/tests/resources/xl_face.jpg b/tests/resources/xl_face.jpg
new file mode 100644
index 0000000..6e394e2
Binary files /dev/null and b/tests/resources/xl_face.jpg differ
diff --git a/tests/resources/xl_face.png b/tests/resources/xl_face.png
deleted file mode 100644
index 7dfda05..0000000
--- a/tests/resources/xl_face.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:25c2c283659ec9175a924ced630bfe0d99d45fde735fe91079709d8c653a493c
-size 450456
diff --git a/tests/resources/xs_faces.jpg b/tests/resources/xs_faces.jpg
new file mode 100644
index 0000000..4eeee3c
Binary files /dev/null and b/tests/resources/xs_faces.jpg differ
diff --git a/tests/resources/xs_faces.png b/tests/resources/xs_faces.png
deleted file mode 100644
index a669d64..0000000
--- a/tests/resources/xs_faces.png
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:97a93bfd39e8b7d1e8ee4906931a08a0b5b1ac6c4089313123983f66ff3069b0
-size 4645655
diff --git a/tests/test_face_depth.py b/tests/test_face_depth.py
new file mode 100644
index 0000000..0df28dc
--- /dev/null
+++ b/tests/test_face_depth.py
@@ -0,0 +1,72 @@
+import capybara as cb
+import numpy as np
+import pytest
+
+from pyface import build_face_depth
+from tests.tools import assert_allclose
+
+RESOURCE_DIR = cb.get_curdir(__file__) / "resources"
+ANSWER_DIR = RESOURCE_DIR / "answer"
+ANSWER_DIR.mkdir(exist_ok=True, parents=True)
+
+TEST_DATA = [
+    {
+        "img_fpath": RESOURCE_DIR / "EmmaWatson1.jpg",
+        "detect": ANSWER_DIR / "EmmaWatson1_detection.npz",
+        "expected": ANSWER_DIR / "EmmaWatson1_depth.npz",
+    },
+    {
+        "img_fpath": RESOURCE_DIR / "JohnnyDepp1.jpg",
+        "detect": ANSWER_DIR / "JohnnyDepp1_detection.npz",
+        "expected": ANSWER_DIR / "JohnnyDepp1_depth.npz",
+    },
+]
+
+
+@pytest.mark.parametrize("data", TEST_DATA)
+def test_face_depth(data):
+    m = build_face_depth(backend="cpu")
+    img = cb.imread(data["img_fpath"])
+    expected = np.load(data["expected"], allow_pickle=True)
+    boxes = np.load(data["detect"], allow_pickle=True)["boxes"]
+    results = m(imgs=[img], boxes=cb.Boxes(boxes), return_depth=True)[0]
+
+    assert_allclose(results["param"], expected["param"])
+    assert_allclose(results["lmk3d68pt"], expected["lmk3d68pt"])
+    assert_allclose(results["pose_degree"], expected["pose_degree"])
+    assert_allclose(results["depth_img"], expected["depth_img"])
+
+
+@pytest.mark.parametrize("data", TEST_DATA)
+def test_face_depth_batch(data):
+    m = build_face_depth(batch_size=4)
+    img = cb.imread(data["img_fpath"])
+    expected = np.load(data["expected"], allow_pickle=True)
+    boxes = np.load(data["detect"], allow_pickle=True)["boxes"]
+    results = m(imgs=[img] * 6, boxes=cb.Boxes(np.tile(boxes, (6, 1))), return_depth=True)
+
+    for result in results:
+        assert_allclose(result["param"], expected["param"])
+        assert_allclose(result["lmk3d68pt"], expected["lmk3d68pt"])
+        assert_allclose(result["pose_degree"], expected["pose_degree"])
+        assert_allclose(result["depth_img"], expected["depth_img"])
+
+
+def gen_target():
+    m = build_face_depth(backend="cpu")
+    for data in TEST_DATA:
+        img = cb.imread(data["img_fpath"])
+        boxes = np.load(data["detect"], allow_pickle=True)["boxes"]
+        results = m(imgs=[img], boxes=cb.Boxes(boxes), return_depth=True)[0]
+        np.savez_compressed(
+            data["expected"],
+            param=results["param"],
+            lmk3d68pt=results["lmk3d68pt"],
+            pose_degree=results["pose_degree"],
+            depth_img=results["depth_img"],
+            allow_pickle=False,
+        )
+
+
+if __name__ == "__main__":
+    gen_target()
diff --git a/tests/test_face_detection.py b/tests/test_face_detection.py
index 1b808e6..86c67a5 100644
--- a/tests/test_face_detection.py
+++ b/tests/test_face_detection.py
@@ -1,81 +1,87 @@
 import capybara as cb
-import cv2
 import numpy as np
 import pytest
 
 import pyface as pf
+from tests.tools import assert_allclose
 
 RESOURCE_DIR = cb.get_curdir(__file__) / "resources"
+ANSWER_DIR = RESOURCE_DIR / "answer"
+ANSWER_DIR.mkdir(exist_ok=True, parents=True)
 
 TEST_DATA = [
     {
-        "img_fpath": RESOURCE_DIR / "EmmaWatson1.png",
-        "expected": {
-            "boxes": np.array([[412.52731323, 194.85551453, 751.04302979, 642.71121216]], dtype="float32"),
-            "lmk5pts": np.array(
-                [
-                    [
-                        [449.80097198, 366.0531826],
-                        [581.79126072, 367.359972],
-                        [465.10058594, 458.07265472],
-                        [460.2443161, 527.72718811],
-                        [570.59969902, 529.41928101],
-                    ]
-                ],
-                dtype="float32",
-            ),
-            "scores": np.array([[0.80300564]], dtype="float32"),
-        },
+        "img_fpath": RESOURCE_DIR / "EmmaWatson1.jpg",
+        "expected": ANSWER_DIR / "EmmaWatson1_detection.npz",
     },
     {
-        "img_fpath": RESOURCE_DIR / "JohnnyDepp1.png",
-        "expected": {
-            "boxes": np.array([[434.57501197, 188.672789, 691.22249818, 507.53144979]], dtype="float32"),
-            "lmk5pts": np.array(
-                [
-                    [
-                        [478.99239435, 284.38753269],
-                        [585.12645214, 287.86553004],
-                        [500.60212723, 351.98840269],
-                        [480.54181178, 422.01026003],
-                        [553.02625371, 426.00003743],
-                    ]
-                ],
-                dtype="float32",
-            ),
-            "scores": np.array([[0.7277902]], dtype="float32"),
-        },
+        "img_fpath": RESOURCE_DIR / "JohnnyDepp1.jpg",
+        "expected": ANSWER_DIR / "JohnnyDepp1_detection.npz",
     },
 ]
 
 
 @pytest.mark.parametrize("data", TEST_DATA)
-def test_build_face_detection(data):
-    model = pf.build_face_detection(
+def test_face_detection(data):
+    m = pf.build_face_detection(
         name="scrfd",
         model_name="scrfd_10g_gnkps_fp32",
-        backend=pf.get_ort_backend(),
+        backend=cb.get_recommended_backend(),
     )
-    img = cv2.imread(data["img_fpath"])
-    faces_on_img = model(imgs=[img])[0]
+    img = cb.imread(data["img_fpath"])
+    expected = np.load(data["expected"])
+    faces_on_img = m(imgs=[img])[0]
 
     assert faces_on_img["infos"]["num_proposals"]
-    np.testing.assert_allclose(faces_on_img["boxes"], data["expected"]["boxes"], rtol=1e-4)
-    np.testing.assert_allclose(faces_on_img["lmk5pts"], data["expected"]["lmk5pts"], rtol=1e-4)
-    np.testing.assert_allclose(faces_on_img["scores"], data["expected"]["scores"], rtol=1e-4)
+    assert_allclose(faces_on_img["boxes"], expected["boxes"])
+    assert_allclose(faces_on_img["lmk5pts"], expected["lmk5pts"])
+    assert_allclose(faces_on_img["scores"], expected["scores"])
 
 
 @pytest.mark.parametrize("data", TEST_DATA)
-def test_build_face_detection_cpu(data):
-    model = pf.build_face_detection(
+def test_face_detection_cpu(data):
+    m = pf.build_face_detection(name="scrfd", model_name="scrfd_10g_gnkps_fp32", backend="cpu")
+    img = cb.imread(data["img_fpath"])
+    expected = np.load(data["expected"], allow_pickle=True)
+    faces_on_img = m(imgs=[img] * 8)[0]
+
+    assert faces_on_img["infos"]["num_proposals"]
+    assert_allclose(faces_on_img["boxes"], expected["boxes"])
+    assert_allclose(faces_on_img["lmk5pts"], expected["lmk5pts"])
+    assert_allclose(faces_on_img["scores"], expected["scores"])
+
+
+@pytest.mark.parametrize("data", TEST_DATA)
+def test_face_detection_batch(data):
+    m = pf.build_face_detection(
         name="scrfd",
         model_name="scrfd_10g_gnkps_fp32",
-        backend="cpu",
+        batch_size=4,
+        backend=cb.get_recommended_backend(),
     )
-    img = cv2.imread(data["img_fpath"])
-    faces_on_img = model(imgs=[img])[0]
+    img = cb.imread(data["img_fpath"])
+    expected = np.load(data["expected"], allow_pickle=True)
+    faces_list = m(imgs=[img] * 5)
+    for faces in faces_list:
+        assert faces["infos"]["num_proposals"]
+        assert_allclose(faces["boxes"], expected["boxes"])
+        assert_allclose(faces["lmk5pts"], expected["lmk5pts"])
+        assert_allclose(faces["scores"], expected["scores"])
 
-    assert faces_on_img["infos"]["num_proposals"]
-    np.testing.assert_allclose(faces_on_img["boxes"], data["expected"]["boxes"], rtol=1e-4)
-    np.testing.assert_allclose(faces_on_img["lmk5pts"], data["expected"]["lmk5pts"], rtol=1e-4)
-    np.testing.assert_allclose(faces_on_img["scores"], data["expected"]["scores"], rtol=1e-4)
+
+def gen_target():
+    m = pf.build_face_detection(name="scrfd", model_name="scrfd_10g_gnkps_fp32", backend="cpu")
+    for data in TEST_DATA:
+        img = cb.imread(data["img_fpath"])
+        faces_on_img = m(imgs=[img] * 8)[0]
+        np.savez_compressed(
+            data["expected"],
+            boxes=faces_on_img["boxes"],
+            lmk5pts=faces_on_img["lmk5pts"],
+            scores=faces_on_img["scores"],
+            allow_pickle=False,
+        )
+
+
+if __name__ == "__main__":
+    gen_target()
diff --git a/tests/tools.py b/tests/tools.py
new file mode 100644
index 0000000..2e19f26
--- /dev/null
+++ b/tests/tools.py
@@ -0,0 +1,5 @@
+from functools import partial
+
+import numpy as np
+
+assert_allclose = partial(np.testing.assert_allclose, rtol=1e-05, atol=np.inf, equal_nan=False)