diff --git a/src/utils/__init__.py b/src/utils/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/utils/animal_landmark_runner.py b/src/utils/animal_landmark_runner.py
new file mode 100644
index 0000000..963acbd
--- /dev/null
+++ b/src/utils/animal_landmark_runner.py
@@ -0,0 +1,138 @@
+# coding: utf-8
+
+"""
+face detectoin and alignment using XPose
+"""
+
+import os
+import pickle
+import torch
+import numpy as np
+from PIL import Image
+from torchvision.ops import nms
+
+from .timer import Timer
+from .rprint import rlog as log
+from .helper import clean_state_dict
+
+from .dependencies.XPose import transforms as T
+from .dependencies.XPose.models import build_model
+from .dependencies.XPose.predefined_keypoints import *
+from .dependencies.XPose.util import box_ops
+from .dependencies.XPose.util.config import Config
+
+
+class XPoseRunner(object):
+    def __init__(self, model_config_path, model_checkpoint_path, embeddings_cache_path=None, cpu_only=False, **kwargs):
+        self.device_id = kwargs.get("device_id", 0)
+        self.flag_use_half_precision = kwargs.get("flag_use_half_precision", True)
+        self.device = f"cuda:{self.device_id}" if not cpu_only else "cpu"
+        self.model = self.load_animal_model(model_config_path, model_checkpoint_path, self.device)
+        self.timer = Timer()
+        # Load cached embeddings if available
+        try:
+            with open(f'{embeddings_cache_path}_9.pkl', 'rb') as f:
+                self.ins_text_embeddings_9, self.kpt_text_embeddings_9 = pickle.load(f)
+            with open(f'{embeddings_cache_path}_68.pkl', 'rb') as f:
+                self.ins_text_embeddings_68, self.kpt_text_embeddings_68 = pickle.load(f)
+            print("Loaded cached embeddings from file.")
+        except Exception:
+            raise ValueError("Could not load clip embeddings from file, please check your file path.")
+
+    def load_animal_model(self, model_config_path, model_checkpoint_path, device):
+        args = Config.fromfile(model_config_path)
+        args.device = device
+        model = build_model(args)
+        checkpoint = torch.load(model_checkpoint_path, map_location=lambda storage, loc: storage)
+        load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
+        model.eval()
+        return model
+
+    def load_image(self, input_image):
+        image_pil = input_image.convert("RGB")
+        transform = T.Compose([
+            T.RandomResize([800], max_size=1333),  # NOTE: fixed size to 800
+            T.ToTensor(),
+            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ])
+        image, _ = transform(image_pil, None)
+        return image_pil, image
+
+    def get_unipose_output(self, image, instance_text_prompt, keypoint_text_prompt, box_threshold, IoU_threshold):
+        instance_list = instance_text_prompt.split(',')
+        
+        if len(keypoint_text_prompt) == 9:
+            # torch.Size([1, 512]) torch.Size([9, 512])
+            ins_text_embeddings, kpt_text_embeddings = self.ins_text_embeddings_9, self.kpt_text_embeddings_9
+        elif len(keypoint_text_prompt) ==68:
+            # torch.Size([1, 512]) torch.Size([68, 512])
+            ins_text_embeddings, kpt_text_embeddings = self.ins_text_embeddings_68, self.kpt_text_embeddings_68
+        else:
+            raise ValueError("Invalid number of keypoint embeddings.")
+        target = {
+            "instance_text_prompt": instance_list,
+            "keypoint_text_prompt": keypoint_text_prompt,
+            "object_embeddings_text": ins_text_embeddings.float(),
+            "kpts_embeddings_text": torch.cat((kpt_text_embeddings.float(), torch.zeros(100 - kpt_text_embeddings.shape[0], 512, device=self.device)), dim=0),
+            "kpt_vis_text": torch.cat((torch.ones(kpt_text_embeddings.shape[0], device=self.device), torch.zeros(100 - kpt_text_embeddings.shape[0], device=self.device)), dim=0)
+        }
+
+        self.model = self.model.to(self.device)
+        image = image.to(self.device)
+
+        with torch.no_grad():
+            with torch.autocast(device_type=self.device[:4], dtype=torch.float16, enabled=self.flag_use_half_precision):
+                outputs = self.model(image[None], [target])
+
+        logits = outputs["pred_logits"].sigmoid()[0]
+        boxes = outputs["pred_boxes"][0]
+        keypoints = outputs["pred_keypoints"][0][:, :2 * len(keypoint_text_prompt)]
+
+        logits_filt = logits.cpu().clone()
+        boxes_filt = boxes.cpu().clone()
+        keypoints_filt = keypoints.cpu().clone()
+        filt_mask = logits_filt.max(dim=1)[0] > box_threshold
+        logits_filt = logits_filt[filt_mask]
+        boxes_filt = boxes_filt[filt_mask]
+        keypoints_filt = keypoints_filt[filt_mask]
+
+        keep_indices = nms(box_ops.box_cxcywh_to_xyxy(boxes_filt), logits_filt.max(dim=1)[0], iou_threshold=IoU_threshold)
+
+        filtered_boxes = boxes_filt[keep_indices]
+        filtered_keypoints = keypoints_filt[keep_indices]
+
+        return filtered_boxes, filtered_keypoints
+
+    def run(self, input_image, instance_text_prompt, keypoint_text_example, box_threshold, IoU_threshold):
+        if keypoint_text_example in globals():
+            keypoint_dict = globals()[keypoint_text_example]
+        elif instance_text_prompt in globals():
+            keypoint_dict = globals()[instance_text_prompt]
+        else:
+            keypoint_dict = globals()["animal"]
+
+        keypoint_text_prompt = keypoint_dict.get("keypoints")
+        keypoint_skeleton = keypoint_dict.get("skeleton")
+
+        image_pil, image = self.load_image(input_image)
+        boxes_filt, keypoints_filt = self.get_unipose_output(image, instance_text_prompt, keypoint_text_prompt, box_threshold, IoU_threshold)
+
+        size = image_pil.size
+        H, W = size[1], size[0]
+        keypoints_filt = keypoints_filt[0].squeeze(0)
+        kp = np.array(keypoints_filt.cpu())
+        num_kpts = len(keypoint_text_prompt)
+        Z = kp[:num_kpts * 2] * np.array([W, H] * num_kpts)
+        Z = Z.reshape(num_kpts * 2)
+        x = Z[0::2]
+        y = Z[1::2]
+        return np.stack((x, y), axis=1)
+
+    def warmup(self):
+        self.timer.tic()
+
+        img_rgb = Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))
+        self.run(img_rgb, 'face', 'face', box_threshold=0.0, IoU_threshold=0.0)
+
+        elapse = self.timer.toc()
+        log(f'XPoseRunner warmup time: {elapse:.3f}s')
diff --git a/src/utils/camera.py b/src/utils/camera.py
new file mode 100644
index 0000000..485a1f7
--- /dev/null
+++ b/src/utils/camera.py
@@ -0,0 +1,73 @@
+# coding: utf-8
+
+"""
+functions for processing and transforming 3D facial keypoints
+"""
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+PI = np.pi
+
+
+def headpose_pred_to_degree(pred):
+    """
+    pred: (bs, 66) or (bs, 1) or others
+    """
+    if pred.ndim > 1 and pred.shape[1] == 66:
+        # NOTE: note that the average is modified to 97.5
+        device = pred.device
+        idx_tensor = [idx for idx in range(0, 66)]
+        idx_tensor = torch.FloatTensor(idx_tensor).to(device)
+        pred = F.softmax(pred, dim=1)
+        degree = torch.sum(pred*idx_tensor, axis=1) * 3 - 97.5
+
+        return degree
+
+    return pred
+
+
+def get_rotation_matrix(pitch_, yaw_, roll_):
+    """ the input is in degree
+    """
+    # transform to radian
+    pitch = pitch_ / 180 * PI
+    yaw = yaw_ / 180 * PI
+    roll = roll_ / 180 * PI
+
+    device = pitch.device
+
+    if pitch.ndim == 1:
+        pitch = pitch.unsqueeze(1)
+    if yaw.ndim == 1:
+        yaw = yaw.unsqueeze(1)
+    if roll.ndim == 1:
+        roll = roll.unsqueeze(1)
+
+    # calculate the euler matrix
+    bs = pitch.shape[0]
+    ones = torch.ones([bs, 1]).to(device)
+    zeros = torch.zeros([bs, 1]).to(device)
+    x, y, z = pitch, yaw, roll
+
+    rot_x = torch.cat([
+        ones, zeros, zeros,
+        zeros, torch.cos(x), -torch.sin(x),
+        zeros, torch.sin(x), torch.cos(x)
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot_y = torch.cat([
+        torch.cos(y), zeros, torch.sin(y),
+        zeros, ones, zeros,
+        -torch.sin(y), zeros, torch.cos(y)
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot_z = torch.cat([
+        torch.cos(z), -torch.sin(z), zeros,
+        torch.sin(z), torch.cos(z), zeros,
+        zeros, zeros, ones
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot = rot_z @ rot_y @ rot_x
+    return rot.permute(0, 2, 1)  # transpose
diff --git a/src/utils/check_windows_port.py b/src/utils/check_windows_port.py
new file mode 100644
index 0000000..2a5f928
--- /dev/null
+++ b/src/utils/check_windows_port.py
@@ -0,0 +1,18 @@
+import socket
+import sys
+
+if len(sys.argv) != 2:
+    print("Usage: python check_port.py <port>")
+    sys.exit(1)
+
+port = int(sys.argv[1])
+
+sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+sock.settimeout(1)
+result = sock.connect_ex(('127.0.0.1', port))
+
+if result == 0:
+    print("LISTENING")
+else:
+    print("NOT LISTENING")
+sock.close
diff --git a/src/utils/crop.py b/src/utils/crop.py
new file mode 100644
index 0000000..5671e95
--- /dev/null
+++ b/src/utils/crop.py
@@ -0,0 +1,423 @@
+# coding: utf-8
+
+"""
+cropping function and the related preprocess functions for cropping
+"""
+
+import numpy as np
+import os.path as osp
+from math import sin, cos, acos, degrees
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False) # NOTE: enforce single thread
+from .rprint import rprint as print
+
+DTYPE = np.float32
+CV2_INTERP = cv2.INTER_LINEAR
+
+def make_abs_path(fn):
+    return osp.join(osp.dirname(osp.realpath(__file__)), fn)
+
+def _transform_img(img, M, dsize, flags=CV2_INTERP, borderMode=None):
+    """ conduct similarity or affine transformation to the image, do not do border operation!
+    img:
+    M: 2x3 matrix or 3x3 matrix
+    dsize: target shape (width, height)
+    """
+    if isinstance(dsize, tuple) or isinstance(dsize, list):
+        _dsize = tuple(dsize)
+    else:
+        _dsize = (dsize, dsize)
+
+    if borderMode is not None:
+        return cv2.warpAffine(img, M[:2, :], dsize=_dsize, flags=flags, borderMode=borderMode, borderValue=(0, 0, 0))
+    else:
+        return cv2.warpAffine(img, M[:2, :], dsize=_dsize, flags=flags)
+
+
+def _transform_pts(pts, M):
+    """ conduct similarity or affine transformation to the pts
+    pts: Nx2 ndarray
+    M: 2x3 matrix or 3x3 matrix
+    return: Nx2
+    """
+    return pts @ M[:2, :2].T + M[:2, 2]
+
+
+def parse_pt2_from_pt101(pt101, use_lip=True):
+    """
+    parsing the 2 points according to the 101 points, which cancels the roll
+    """
+    # the former version use the eye center, but it is not robust, now use interpolation
+    pt_left_eye = np.mean(pt101[[39, 42, 45, 48]], axis=0)  # left eye center
+    pt_right_eye = np.mean(pt101[[51, 54, 57, 60]], axis=0)  # right eye center
+
+    if use_lip:
+        # use lip
+        pt_center_eye = (pt_left_eye + pt_right_eye) / 2
+        pt_center_lip = (pt101[75] + pt101[81]) / 2
+        pt2 = np.stack([pt_center_eye, pt_center_lip], axis=0)
+    else:
+        pt2 = np.stack([pt_left_eye, pt_right_eye], axis=0)
+    return pt2
+
+
+def parse_pt2_from_pt106(pt106, use_lip=True):
+    """
+    parsing the 2 points according to the 106 points, which cancels the roll
+    """
+    pt_left_eye = np.mean(pt106[[33, 35, 40, 39]], axis=0)  # left eye center
+    pt_right_eye = np.mean(pt106[[87, 89, 94, 93]], axis=0)  # right eye center
+
+    if use_lip:
+        # use lip
+        pt_center_eye = (pt_left_eye + pt_right_eye) / 2
+        pt_center_lip = (pt106[52] + pt106[61]) / 2
+        pt2 = np.stack([pt_center_eye, pt_center_lip], axis=0)
+    else:
+        pt2 = np.stack([pt_left_eye, pt_right_eye], axis=0)
+    return pt2
+
+
+def parse_pt2_from_pt203(pt203, use_lip=True):
+    """
+    parsing the 2 points according to the 203 points, which cancels the roll
+    """
+    pt_left_eye = np.mean(pt203[[0, 6, 12, 18]], axis=0)  # left eye center
+    pt_right_eye = np.mean(pt203[[24, 30, 36, 42]], axis=0)  # right eye center
+    if use_lip:
+        # use lip
+        pt_center_eye = (pt_left_eye + pt_right_eye) / 2
+        pt_center_lip = (pt203[48] + pt203[66]) / 2
+        pt2 = np.stack([pt_center_eye, pt_center_lip], axis=0)
+    else:
+        pt2 = np.stack([pt_left_eye, pt_right_eye], axis=0)
+    return pt2
+
+
+def parse_pt2_from_pt68(pt68, use_lip=True):
+    """
+    parsing the 2 points according to the 68 points, which cancels the roll
+    """
+    lm_idx = np.array([31, 37, 40, 43, 46, 49, 55], dtype=np.int32) - 1
+    if use_lip:
+        pt5 = np.stack([
+            np.mean(pt68[lm_idx[[1, 2]], :], 0),  # left eye
+            np.mean(pt68[lm_idx[[3, 4]], :], 0),  # right eye
+            pt68[lm_idx[0], :],  # nose
+            pt68[lm_idx[5], :],  # lip
+            pt68[lm_idx[6], :]   # lip
+        ], axis=0)
+
+        pt2 = np.stack([
+            (pt5[0] + pt5[1]) / 2,
+            (pt5[3] + pt5[4]) / 2
+        ], axis=0)
+    else:
+        pt2 = np.stack([
+            np.mean(pt68[lm_idx[[1, 2]], :], 0),  # left eye
+            np.mean(pt68[lm_idx[[3, 4]], :], 0),  # right eye
+        ], axis=0)
+
+    return pt2
+
+
+def parse_pt2_from_pt5(pt5, use_lip=True):
+    """
+    parsing the 2 points according to the 5 points, which cancels the roll
+    """
+    if use_lip:
+        pt2 = np.stack([
+            (pt5[0] + pt5[1]) / 2,
+            (pt5[3] + pt5[4]) / 2
+        ], axis=0)
+    else:
+        pt2 = np.stack([
+            pt5[0],
+            pt5[1]
+        ], axis=0)
+    return pt2
+
+def parse_pt2_from_pt9(pt9, use_lip=True):
+    '''
+    parsing the 2 points according to the 9 points, which cancels the roll
+    ['right eye right', 'right eye left', 'left eye right', 'left eye left', 'nose tip', 'lip right', 'lip left', 'upper lip', 'lower lip']
+    '''
+    if use_lip:
+        pt9 = np.stack([
+            (pt9[2] + pt9[3]) / 2, # left eye
+            (pt9[0] + pt9[1]) / 2, # right eye
+            pt9[4],
+            (pt9[5] + pt9[6] ) / 2 # lip
+        ], axis=0)
+        pt2 = np.stack([
+            (pt9[0] + pt9[1]) / 2, # eye
+            pt9[3] # lip
+        ], axis=0)
+    else:
+        pt2 = np.stack([
+            (pt9[2] + pt9[3]) / 2,
+            (pt9[0] + pt9[1]) / 2,
+        ], axis=0)
+
+    return pt2
+
+def parse_pt2_from_pt_x(pts, use_lip=True):
+    if pts.shape[0] == 101:
+        pt2 = parse_pt2_from_pt101(pts, use_lip=use_lip)
+    elif pts.shape[0] == 106:
+        pt2 = parse_pt2_from_pt106(pts, use_lip=use_lip)
+    elif pts.shape[0] == 68:
+        pt2 = parse_pt2_from_pt68(pts, use_lip=use_lip)
+    elif pts.shape[0] == 5:
+        pt2 = parse_pt2_from_pt5(pts, use_lip=use_lip)
+    elif pts.shape[0] == 203:
+        pt2 = parse_pt2_from_pt203(pts, use_lip=use_lip)
+    elif pts.shape[0] > 101:
+        # take the first 101 points
+        pt2 = parse_pt2_from_pt101(pts[:101], use_lip=use_lip)
+    elif pts.shape[0] == 9:
+        pt2 = parse_pt2_from_pt9(pts, use_lip=use_lip)
+    else:
+        raise Exception(f'Unknow shape: {pts.shape}')
+
+    if not use_lip:
+        # NOTE: to compile with the latter code, need to rotate the pt2 90 degrees clockwise manually
+        v = pt2[1] - pt2[0]
+        pt2[1, 0] = pt2[0, 0] - v[1]
+        pt2[1, 1] = pt2[0, 1] + v[0]
+
+    return pt2
+
+
+def parse_rect_from_landmark(
+    pts,
+    scale=1.5,
+    need_square=True,
+    vx_ratio=0,
+    vy_ratio=0,
+    use_deg_flag=False,
+    **kwargs
+):
+    """parsing center, size, angle from 101/68/5/x landmarks
+    vx_ratio: the offset ratio along the pupil axis x-axis, multiplied by size
+    vy_ratio: the offset ratio along the pupil axis y-axis, multiplied by size, which is used to contain more forehead area
+
+    judge with pts.shape
+    """
+    pt2 = parse_pt2_from_pt_x(pts, use_lip=kwargs.get('use_lip', True))
+
+    uy = pt2[1] - pt2[0]
+    l = np.linalg.norm(uy)
+    if l <= 1e-3:
+        uy = np.array([0, 1], dtype=DTYPE)
+    else:
+        uy /= l
+    ux = np.array((uy[1], -uy[0]), dtype=DTYPE)
+
+    # the rotation degree of the x-axis, the clockwise is positive, the counterclockwise is negative (image coordinate system)
+    # print(uy)
+    # print(ux)
+    angle = acos(ux[0])
+    if ux[1] < 0:
+        angle = -angle
+
+    # rotation matrix
+    M = np.array([ux, uy])
+
+    # calculate the size which contains the angle degree of the bbox, and the center
+    center0 = np.mean(pts, axis=0)
+    rpts = (pts - center0) @ M.T  # (M @ P.T).T = P @ M.T
+    lt_pt = np.min(rpts, axis=0)
+    rb_pt = np.max(rpts, axis=0)
+    center1 = (lt_pt + rb_pt) / 2
+
+    size = rb_pt - lt_pt
+    if need_square:
+        m = max(size[0], size[1])
+        size[0] = m
+        size[1] = m
+
+    size *= scale  # scale size
+    center = center0 + ux * center1[0] + uy * center1[1]  # counterclockwise rotation, equivalent to M.T @ center1.T
+    center = center + ux * (vx_ratio * size) + uy * \
+        (vy_ratio * size)  # considering the offset in vx and vy direction
+
+    if use_deg_flag:
+        angle = degrees(angle)
+
+    return center, size, angle
+
+
+def parse_bbox_from_landmark(pts, **kwargs):
+    center, size, angle = parse_rect_from_landmark(pts, **kwargs)
+    cx, cy = center
+    w, h = size
+
+    # calculate the vertex positions before rotation
+    bbox = np.array([
+        [cx-w/2, cy-h/2],  # left, top
+        [cx+w/2, cy-h/2],
+        [cx+w/2, cy+h/2],  # right, bottom
+        [cx-w/2, cy+h/2]
+    ], dtype=DTYPE)
+
+    # construct rotation matrix
+    bbox_rot = bbox.copy()
+    R = np.array([
+        [np.cos(angle), -np.sin(angle)],
+        [np.sin(angle),  np.cos(angle)]
+    ], dtype=DTYPE)
+
+    # calculate the relative position of each vertex from the rotation center, then rotate these positions, and finally add the coordinates of the rotation center
+    bbox_rot = (bbox_rot - center) @ R.T + center
+
+    return {
+        'center': center,  # 2x1
+        'size': size,  # scalar
+        'angle': angle,  # rad, counterclockwise
+        'bbox': bbox,  # 4x2
+        'bbox_rot': bbox_rot,  # 4x2
+    }
+
+
+def crop_image_by_bbox(img, bbox, lmk=None, dsize=512, angle=None, flag_rot=False, **kwargs):
+    left, top, right, bot = bbox
+    if int(right - left) != int(bot - top):
+        print(f'right-left {right-left} != bot-top {bot-top}')
+    size = right - left
+
+    src_center = np.array([(left + right) / 2, (top + bot) / 2], dtype=DTYPE)
+    tgt_center = np.array([dsize / 2, dsize / 2], dtype=DTYPE)
+
+    s = dsize / size  # scale
+    if flag_rot and angle is not None:
+        costheta, sintheta = cos(angle), sin(angle)
+        cx, cy = src_center[0], src_center[1]  # ori center
+        tcx, tcy = tgt_center[0], tgt_center[1]  # target center
+        # need to infer
+        M_o2c = np.array(
+            [[s * costheta, s * sintheta, tcx - s * (costheta * cx + sintheta * cy)],
+             [-s * sintheta, s * costheta, tcy - s * (-sintheta * cx + costheta * cy)]],
+            dtype=DTYPE
+        )
+    else:
+        M_o2c = np.array(
+            [[s, 0, tgt_center[0] - s * src_center[0]],
+             [0, s, tgt_center[1] - s * src_center[1]]],
+            dtype=DTYPE
+        )
+
+    # if flag_rot and angle is None:
+        # print('angle is None, but flag_rotate is True', style="bold yellow")
+
+    img_crop = _transform_img(img, M_o2c, dsize=dsize, borderMode=kwargs.get('borderMode', None))
+    lmk_crop = _transform_pts(lmk, M_o2c) if lmk is not None else None
+
+    M_o2c = np.vstack([M_o2c, np.array([0, 0, 1], dtype=DTYPE)])
+    M_c2o = np.linalg.inv(M_o2c)
+
+    # cv2.imwrite('crop.jpg', img_crop)
+
+    return {
+        'img_crop': img_crop,
+        'lmk_crop': lmk_crop,
+        'M_o2c': M_o2c,
+        'M_c2o': M_c2o,
+    }
+
+
+def _estimate_similar_transform_from_pts(
+    pts,
+    dsize,
+    scale=1.5,
+    vx_ratio=0,
+    vy_ratio=-0.1,
+    flag_do_rot=True,
+    **kwargs
+):
+    """ calculate the affine matrix of the cropped image from sparse points, the original image to the cropped image, the inverse is the cropped image to the original image
+    pts: landmark, 101 or 68 points or other points, Nx2
+    scale: the larger scale factor, the smaller face ratio
+    vx_ratio: x shift
+    vy_ratio: y shift, the smaller the y shift, the lower the face region
+    rot_flag: if it is true, conduct correction
+    """
+    center, size, angle = parse_rect_from_landmark(
+        pts, scale=scale, vx_ratio=vx_ratio, vy_ratio=vy_ratio,
+        use_lip=kwargs.get('use_lip', True)
+    )
+
+    s = dsize / size[0]  # scale
+    tgt_center = np.array([dsize / 2, dsize / 2], dtype=DTYPE)  # center of dsize
+
+    if flag_do_rot:
+        costheta, sintheta = cos(angle), sin(angle)
+        cx, cy = center[0], center[1]  # ori center
+        tcx, tcy = tgt_center[0], tgt_center[1]  # target center
+        # need to infer
+        M_INV = np.array(
+            [[s * costheta, s * sintheta, tcx - s * (costheta * cx + sintheta * cy)],
+             [-s * sintheta, s * costheta, tcy - s * (-sintheta * cx + costheta * cy)]],
+            dtype=DTYPE
+        )
+    else:
+        M_INV = np.array(
+            [[s, 0, tgt_center[0] - s * center[0]],
+             [0, s, tgt_center[1] - s * center[1]]],
+            dtype=DTYPE
+        )
+
+    M_INV_H = np.vstack([M_INV, np.array([0, 0, 1])])
+    M = np.linalg.inv(M_INV_H)
+
+    # M_INV is from the original image to the cropped image, M is from the cropped image to the original image
+    return M_INV, M[:2, ...]
+
+
+def crop_image(img, pts: np.ndarray, **kwargs):
+    dsize = kwargs.get('dsize', 224)
+    scale = kwargs.get('scale', 1.5)  # 1.5 | 1.6
+    vy_ratio = kwargs.get('vy_ratio', -0.1)  # -0.0625 | -0.1
+
+    M_INV, _ = _estimate_similar_transform_from_pts(
+        pts,
+        dsize=dsize,
+        scale=scale,
+        vy_ratio=vy_ratio,
+        flag_do_rot=kwargs.get('flag_do_rot', True),
+    )
+
+    img_crop = _transform_img(img, M_INV, dsize)  # origin to crop
+    pt_crop = _transform_pts(pts, M_INV)
+
+    M_o2c = np.vstack([M_INV, np.array([0, 0, 1], dtype=DTYPE)])
+    M_c2o = np.linalg.inv(M_o2c)
+
+    ret_dct = {
+        'M_o2c': M_o2c,  # from the original image to the cropped image 3x3
+        'M_c2o': M_c2o,  # from the cropped image to the original image 3x3
+        'img_crop': img_crop,  # the cropped image
+        'pt_crop': pt_crop,  # the landmarks of the cropped image
+    }
+
+    return ret_dct
+
+def average_bbox_lst(bbox_lst):
+    if len(bbox_lst) == 0:
+        return None
+    bbox_arr = np.array(bbox_lst)
+    return np.mean(bbox_arr, axis=0).tolist()
+
+def prepare_paste_back(mask_crop, crop_M_c2o, dsize):
+    """prepare mask for later image paste back
+    """
+    mask_ori = _transform_img(mask_crop, crop_M_c2o, dsize)
+    mask_ori = mask_ori.astype(np.float32) / 255.
+    return mask_ori
+
+def paste_back(img_crop, M_c2o, img_ori, mask_ori):
+    """paste back the image
+    """
+    dsize = (img_ori.shape[1], img_ori.shape[0])
+    result = _transform_img(img_crop, M_c2o, dsize=dsize)
+    result = np.clip(mask_ori * result + (1 - mask_ori) * img_ori, 0, 255).astype(np.uint8)
+    return result
diff --git a/src/utils/cropper.py b/src/utils/cropper.py
new file mode 100644
index 0000000..f07bc8c
--- /dev/null
+++ b/src/utils/cropper.py
@@ -0,0 +1,313 @@
+# coding: utf-8
+
+import os.path as osp
+import torch
+import numpy as np
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False)
+
+from PIL import Image
+from typing import List, Tuple, Union
+from dataclasses import dataclass, field
+
+from ..config.crop_config import CropConfig
+from .crop import (
+    average_bbox_lst,
+    crop_image,
+    crop_image_by_bbox,
+    parse_bbox_from_landmark,
+)
+from .io import contiguous
+from .rprint import rlog as log
+from .face_analysis_diy import FaceAnalysisDIY
+from .human_landmark_runner import LandmarkRunner as HumanLandmark
+
+def make_abs_path(fn):
+    return osp.join(osp.dirname(osp.realpath(__file__)), fn)
+
+
+@dataclass
+class Trajectory:
+    start: int = -1  # start frame
+    end: int = -1  # end frame
+    lmk_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # lmk list
+    bbox_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # bbox list
+    M_c2o_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # M_c2o list
+
+    frame_rgb_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # frame list
+    lmk_crop_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # lmk list
+    frame_rgb_crop_lst: Union[Tuple, List, np.ndarray] = field(default_factory=list)  # frame crop list
+
+
+class Cropper(object):
+    def __init__(self, **kwargs) -> None:
+        self.crop_cfg: CropConfig = kwargs.get("crop_cfg", None)
+        self.image_type = kwargs.get("image_type", 'human_face')
+        device_id = kwargs.get("device_id", 0)
+        flag_force_cpu = kwargs.get("flag_force_cpu", False)
+        if flag_force_cpu:
+            device = "cpu"
+            face_analysis_wrapper_provider = ["CPUExecutionProvider"]
+        else:
+            try:
+                if torch.backends.mps.is_available():
+                    # Shape inference currently fails with CoreMLExecutionProvider
+                    # for the retinaface model
+                    device = "mps"
+                    face_analysis_wrapper_provider = ["CPUExecutionProvider"]
+                else:
+                    device = "cuda"
+                    face_analysis_wrapper_provider = ["CUDAExecutionProvider"]
+            except:
+                    device = "cuda"
+                    face_analysis_wrapper_provider = ["CUDAExecutionProvider"]
+        self.face_analysis_wrapper = FaceAnalysisDIY(
+                    name="buffalo_l",
+                    root=self.crop_cfg.insightface_root,
+                    providers=face_analysis_wrapper_provider,
+                )
+        self.face_analysis_wrapper.prepare(ctx_id=device_id, det_size=(512, 512), det_thresh=self.crop_cfg.det_thresh)
+        self.face_analysis_wrapper.warmup()
+
+        self.human_landmark_runner = HumanLandmark(
+            ckpt_path=self.crop_cfg.landmark_ckpt_path,
+            onnx_provider=device,
+            device_id=device_id,
+        )
+        self.human_landmark_runner.warmup()
+
+        if self.image_type == "animal_face":
+            from .animal_landmark_runner import XPoseRunner as AnimalLandmarkRunner
+            self.animal_landmark_runner = AnimalLandmarkRunner(
+                    model_config_path=self.crop_cfg.xpose_config_file_path,
+                    model_checkpoint_path=self.crop_cfg.xpose_ckpt_path,
+                    embeddings_cache_path=self.crop_cfg.xpose_embedding_cache_path,
+                    flag_use_half_precision=kwargs.get("flag_use_half_precision", True),
+                )
+            self.animal_landmark_runner.warmup()
+
+    def update_config(self, user_args):
+        for k, v in user_args.items():
+            if hasattr(self.crop_cfg, k):
+                setattr(self.crop_cfg, k, v)
+
+    def crop_source_image(self, img_rgb_: np.ndarray, crop_cfg: CropConfig):
+        # crop a source image and get neccessary information
+        img_rgb = img_rgb_.copy()  # copy it
+        img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
+
+        if self.image_type == "human_face":
+            src_face = self.face_analysis_wrapper.get(
+                img_bgr,
+                flag_do_landmark_2d_106=True,
+                direction=crop_cfg.direction,
+                max_face_num=crop_cfg.max_face_num,
+            )
+
+            if len(src_face) == 0:
+                log("No face detected in the source image.")
+                return None
+            elif len(src_face) > 1:
+                log(f"More than one face detected in the image, only pick one face by rule {crop_cfg.direction}.")
+
+            # NOTE: temporarily only pick the first face, to support multiple face in the future
+            src_face = src_face[0]
+            lmk = src_face.landmark_2d_106  # this is the 106 landmarks from insightface
+        else:
+            tmp_dct = {
+                'animal_face_9': 'animal_face',
+                'animal_face_68': 'face'
+            }
+
+            img_rgb_pil = Image.fromarray(img_rgb)
+            lmk = self.animal_landmark_runner.run(
+                img_rgb_pil,
+                'face',
+                tmp_dct[crop_cfg.animal_face_type],
+                0,
+                0
+            )
+
+        # crop the face
+        ret_dct = crop_image(
+            img_rgb,  # ndarray
+            lmk,  # 106x2 or Nx2
+            dsize=crop_cfg.dsize,
+            scale=crop_cfg.scale,
+            vx_ratio=crop_cfg.vx_ratio,
+            vy_ratio=crop_cfg.vy_ratio,
+            flag_do_rot=crop_cfg.flag_do_rot,
+        )
+
+        # update a 256x256 version for network input
+        ret_dct["img_crop_256x256"] = cv2.resize(ret_dct["img_crop"], (256, 256), interpolation=cv2.INTER_AREA)
+        if self.image_type == "human_face":
+            lmk = self.human_landmark_runner.run(img_rgb, lmk)
+            ret_dct["lmk_crop"] = lmk
+            ret_dct["lmk_crop_256x256"] = ret_dct["lmk_crop"] * 256 / crop_cfg.dsize
+        else:
+            # 68x2 or 9x2
+            ret_dct["lmk_crop"] = lmk
+
+        return ret_dct
+
+    def calc_lmk_from_cropped_image(self, img_rgb_, **kwargs):
+        direction = kwargs.get("direction", "large-small")
+        src_face = self.face_analysis_wrapper.get(
+            contiguous(img_rgb_[..., ::-1]),  # convert to BGR
+            flag_do_landmark_2d_106=True,
+            direction=direction,
+        )
+        if len(src_face) == 0:
+            log("No face detected in the source image.")
+            return None
+        elif len(src_face) > 1:
+            log(f"More than one face detected in the image, only pick one face by rule {direction}.")
+        src_face = src_face[0]
+        lmk = src_face.landmark_2d_106
+        lmk = self.human_landmark_runner.run(img_rgb_, lmk)
+
+        return lmk
+
+    # TODO: support skipping frame with NO FACE
+    def crop_source_video(self, source_rgb_lst, crop_cfg: CropConfig, **kwargs):
+        """Tracking based landmarks/alignment and cropping"""
+        trajectory = Trajectory()
+        direction = kwargs.get("direction", "large-small")
+        for idx, frame_rgb in enumerate(source_rgb_lst):
+            if idx == 0 or trajectory.start == -1:
+                src_face = self.face_analysis_wrapper.get(
+                    contiguous(frame_rgb[..., ::-1]),
+                    flag_do_landmark_2d_106=True,
+                    direction=crop_cfg.direction,
+                    max_face_num=crop_cfg.max_face_num,
+                )
+                if len(src_face) == 0:
+                    log(f"No face detected in the frame #{idx}")
+                    continue
+                elif len(src_face) > 1:
+                    log(f"More than one face detected in the source frame_{idx}, only pick one face by rule {direction}.")
+                src_face = src_face[0]
+                lmk = src_face.landmark_2d_106
+                lmk = self.human_landmark_runner.run(frame_rgb, lmk)
+                trajectory.start, trajectory.end = idx, idx
+            else:
+                # TODO: add IOU check for tracking
+                lmk = self.human_landmark_runner.run(frame_rgb, trajectory.lmk_lst[-1])
+                trajectory.end = idx
+
+            trajectory.lmk_lst.append(lmk)
+
+            # crop the face
+            ret_dct = crop_image(
+                frame_rgb,  # ndarray
+                lmk,  # 106x2 or Nx2
+                dsize=crop_cfg.dsize,
+                scale=crop_cfg.scale,
+                vx_ratio=crop_cfg.vx_ratio,
+                vy_ratio=crop_cfg.vy_ratio,
+                flag_do_rot=crop_cfg.flag_do_rot,
+            )
+            lmk = self.human_landmark_runner.run(frame_rgb, lmk)
+            ret_dct["lmk_crop"] = lmk
+
+            # update a 256x256 version for network input
+            ret_dct["img_crop_256x256"] = cv2.resize(ret_dct["img_crop"], (256, 256), interpolation=cv2.INTER_AREA)
+            ret_dct["lmk_crop_256x256"] = ret_dct["lmk_crop"] * 256 / crop_cfg.dsize
+
+            trajectory.frame_rgb_crop_lst.append(ret_dct["img_crop_256x256"])
+            trajectory.lmk_crop_lst.append(ret_dct["lmk_crop_256x256"])
+            trajectory.M_c2o_lst.append(ret_dct['M_c2o'])
+
+        return {
+            "frame_crop_lst": trajectory.frame_rgb_crop_lst,
+            "lmk_crop_lst": trajectory.lmk_crop_lst,
+            "M_c2o_lst": trajectory.M_c2o_lst,
+        }
+
+    def crop_driving_video(self, driving_rgb_lst, **kwargs):
+        """Tracking based landmarks/alignment and cropping"""
+        trajectory = Trajectory()
+        direction = kwargs.get("direction", "large-small")
+        for idx, frame_rgb in enumerate(driving_rgb_lst):
+            if idx == 0 or trajectory.start == -1:
+                src_face = self.face_analysis_wrapper.get(
+                    contiguous(frame_rgb[..., ::-1]),
+                    flag_do_landmark_2d_106=True,
+                    direction=direction,
+                )
+                if len(src_face) == 0:
+                    log(f"No face detected in the frame #{idx}")
+                    continue
+                elif len(src_face) > 1:
+                    log(f"More than one face detected in the driving frame_{idx}, only pick one face by rule {direction}.")
+                src_face = src_face[0]
+                lmk = src_face.landmark_2d_106
+                lmk = self.human_landmark_runner.run(frame_rgb, lmk)
+                trajectory.start, trajectory.end = idx, idx
+            else:
+                lmk = self.human_landmark_runner.run(frame_rgb, trajectory.lmk_lst[-1])
+                trajectory.end = idx
+
+            trajectory.lmk_lst.append(lmk)
+            ret_bbox = parse_bbox_from_landmark(
+                lmk,
+                scale=self.crop_cfg.scale_crop_driving_video,
+                vx_ratio_crop_driving_video=self.crop_cfg.vx_ratio_crop_driving_video,
+                vy_ratio=self.crop_cfg.vy_ratio_crop_driving_video,
+            )["bbox"]
+            bbox = [
+                ret_bbox[0, 0],
+                ret_bbox[0, 1],
+                ret_bbox[2, 0],
+                ret_bbox[2, 1],
+            ]  # 4,
+            trajectory.bbox_lst.append(bbox)  # bbox
+            trajectory.frame_rgb_lst.append(frame_rgb)
+
+        global_bbox = average_bbox_lst(trajectory.bbox_lst)
+
+        for idx, (frame_rgb, lmk) in enumerate(zip(trajectory.frame_rgb_lst, trajectory.lmk_lst)):
+            ret_dct = crop_image_by_bbox(
+                frame_rgb,
+                global_bbox,
+                lmk=lmk,
+                dsize=kwargs.get("dsize", 512),
+                flag_rot=False,
+                borderValue=(0, 0, 0),
+            )
+            trajectory.frame_rgb_crop_lst.append(ret_dct["img_crop"])
+            trajectory.lmk_crop_lst.append(ret_dct["lmk_crop"])
+
+        return {
+            "frame_crop_lst": trajectory.frame_rgb_crop_lst,
+            "lmk_crop_lst": trajectory.lmk_crop_lst,
+        }
+
+
+    def calc_lmks_from_cropped_video(self, driving_rgb_crop_lst, **kwargs):
+        """Tracking based landmarks/alignment"""
+        trajectory = Trajectory()
+        direction = kwargs.get("direction", "large-small")
+
+        for idx, frame_rgb_crop in enumerate(driving_rgb_crop_lst):
+            if idx == 0 or trajectory.start == -1:
+                src_face = self.face_analysis_wrapper.get(
+                    contiguous(frame_rgb_crop[..., ::-1]),  # convert to BGR
+                    flag_do_landmark_2d_106=True,
+                    direction=direction,
+                )
+                if len(src_face) == 0:
+                    log(f"No face detected in the frame #{idx}")
+                    raise Exception(f"No face detected in the frame #{idx}")
+                elif len(src_face) > 1:
+                    log(f"More than one face detected in the driving frame_{idx}, only pick one face by rule {direction}.")
+                src_face = src_face[0]
+                lmk = src_face.landmark_2d_106
+                lmk = self.human_landmark_runner.run(frame_rgb_crop, lmk)
+                trajectory.start, trajectory.end = idx, idx
+            else:
+                lmk = self.human_landmark_runner.run(frame_rgb_crop, trajectory.lmk_lst[-1])
+                trajectory.end = idx
+
+            trajectory.lmk_lst.append(lmk)
+        return trajectory.lmk_lst
diff --git a/src/utils/face_analysis_diy.py b/src/utils/face_analysis_diy.py
new file mode 100644
index 0000000..359ac17
--- /dev/null
+++ b/src/utils/face_analysis_diy.py
@@ -0,0 +1,79 @@
+# coding: utf-8
+
+"""
+face detectoin and alignment using InsightFace
+"""
+
+import numpy as np
+from .rprint import rlog as log
+from .dependencies.insightface.app import FaceAnalysis
+from .dependencies.insightface.app.common import Face
+from .timer import Timer
+
+
+def sort_by_direction(faces, direction: str = 'large-small', face_center=None):
+    if len(faces) <= 0:
+        return faces
+
+    if direction == 'left-right':
+        return sorted(faces, key=lambda face: face['bbox'][0])
+    if direction == 'right-left':
+        return sorted(faces, key=lambda face: face['bbox'][0], reverse=True)
+    if direction == 'top-bottom':
+        return sorted(faces, key=lambda face: face['bbox'][1])
+    if direction == 'bottom-top':
+        return sorted(faces, key=lambda face: face['bbox'][1], reverse=True)
+    if direction == 'small-large':
+        return sorted(faces, key=lambda face: (face['bbox'][2] - face['bbox'][0]) * (face['bbox'][3] - face['bbox'][1]))
+    if direction == 'large-small':
+        return sorted(faces, key=lambda face: (face['bbox'][2] - face['bbox'][0]) * (face['bbox'][3] - face['bbox'][1]), reverse=True)
+    if direction == 'distance-from-retarget-face':
+        return sorted(faces, key=lambda face: (((face['bbox'][2]+face['bbox'][0])/2-face_center[0])**2+((face['bbox'][3]+face['bbox'][1])/2-face_center[1])**2)**0.5)
+    return faces
+
+
+class FaceAnalysisDIY(FaceAnalysis):
+    def __init__(self, name='buffalo_l', root='~/.insightface', allowed_modules=None, **kwargs):
+        super().__init__(name=name, root=root, allowed_modules=allowed_modules, **kwargs)
+
+        self.timer = Timer()
+
+    def get(self, img_bgr, **kwargs):
+        max_num = kwargs.get('max_face_num', 0)  # the number of the detected faces, 0 means no limit
+        flag_do_landmark_2d_106 = kwargs.get('flag_do_landmark_2d_106', True)  # whether to do 106-point detection
+        direction = kwargs.get('direction', 'large-small')  # sorting direction
+        face_center = None
+
+        bboxes, kpss = self.det_model.detect(img_bgr, max_num=max_num, metric='default')
+        if bboxes.shape[0] == 0:
+            return []
+        ret = []
+        for i in range(bboxes.shape[0]):
+            bbox = bboxes[i, 0:4]
+            det_score = bboxes[i, 4]
+            kps = None
+            if kpss is not None:
+                kps = kpss[i]
+            face = Face(bbox=bbox, kps=kps, det_score=det_score)
+            for taskname, model in self.models.items():
+                if taskname == 'detection':
+                    continue
+
+                if (not flag_do_landmark_2d_106) and taskname == 'landmark_2d_106':
+                    continue
+
+                # print(f'taskname: {taskname}')
+                model.get(img_bgr, face)
+            ret.append(face)
+
+        ret = sort_by_direction(ret, direction, face_center)
+        return ret
+
+    def warmup(self):
+        self.timer.tic()
+
+        img_bgr = np.zeros((512, 512, 3), dtype=np.uint8)
+        self.get(img_bgr)
+
+        elapse = self.timer.toc()
+        log(f'FaceAnalysisDIY warmup time: {elapse:.3f}s')
diff --git a/src/utils/filter.py b/src/utils/filter.py
new file mode 100644
index 0000000..a7b1692
--- /dev/null
+++ b/src/utils/filter.py
@@ -0,0 +1,19 @@
+# coding: utf-8
+
+import torch
+import numpy as np
+from pykalman import KalmanFilter
+
+
+def smooth(x_d_lst, shape, device, observation_variance=3e-7, process_variance=1e-5):
+    x_d_lst_reshape = [x.reshape(-1) for x in x_d_lst]
+    x_d_stacked = np.vstack(x_d_lst_reshape)
+    kf = KalmanFilter(
+        initial_state_mean=x_d_stacked[0],
+        n_dim_obs=x_d_stacked.shape[1],
+        transition_covariance=process_variance * np.eye(x_d_stacked.shape[1]),
+        observation_covariance=observation_variance * np.eye(x_d_stacked.shape[1])
+    )
+    smoothed_state_means, _ = kf.smooth(x_d_stacked)
+    x_d_lst_smooth = [torch.tensor(state_mean.reshape(shape[-2:]), dtype=torch.float32, device=device) for state_mean in smoothed_state_means]
+    return x_d_lst_smooth
diff --git a/src/utils/helper.py b/src/utils/helper.py
new file mode 100644
index 0000000..cf97033
--- /dev/null
+++ b/src/utils/helper.py
@@ -0,0 +1,199 @@
+# coding: utf-8
+
+"""
+utility functions and classes to handle feature extraction and model loading
+"""
+
+import os
+import os.path as osp
+import torch
+from collections import OrderedDict
+import numpy as np
+from scipy.spatial import ConvexHull # pylint: disable=E0401,E0611
+from typing import Union
+import cv2
+
+from ..modules.spade_generator import SPADEDecoder
+from ..modules.warping_network import WarpingNetwork
+from ..modules.motion_extractor import MotionExtractor
+from ..modules.appearance_feature_extractor import AppearanceFeatureExtractor
+from ..modules.stitching_retargeting_network import StitchingRetargetingNetwork
+
+
+def tensor_to_numpy(data: Union[np.ndarray, torch.Tensor]) -> np.ndarray:
+    """transform torch.Tensor into numpy.ndarray"""
+    if isinstance(data, torch.Tensor):
+        return data.data.cpu().numpy()
+    return data
+
+def calc_motion_multiplier(
+    kp_source: Union[np.ndarray, torch.Tensor],
+    kp_driving_initial: Union[np.ndarray, torch.Tensor]
+) -> float:
+    """calculate motion_multiplier based on the source image and the first driving frame"""
+    kp_source_np = tensor_to_numpy(kp_source)
+    kp_driving_initial_np = tensor_to_numpy(kp_driving_initial)
+
+    source_area = ConvexHull(kp_source_np.squeeze(0)).volume
+    driving_area = ConvexHull(kp_driving_initial_np.squeeze(0)).volume
+    motion_multiplier = np.sqrt(source_area) / np.sqrt(driving_area)
+    # motion_multiplier = np.cbrt(source_area) / np.cbrt(driving_area)
+
+    return motion_multiplier
+
+def suffix(filename):
+    """a.jpg -> jpg"""
+    pos = filename.rfind(".")
+    if pos == -1:
+        return ""
+    return filename[pos + 1:]
+
+
+def prefix(filename):
+    """a.jpg -> a"""
+    pos = filename.rfind(".")
+    if pos == -1:
+        return filename
+    return filename[:pos]
+
+
+def basename(filename):
+    """a/b/c.jpg -> c"""
+    return prefix(osp.basename(filename))
+
+
+def remove_suffix(filepath):
+    """a/b/c.jpg -> a/b/c"""
+    return osp.join(osp.dirname(filepath), basename(filepath))
+
+
+def is_image(file_path):
+    image_extensions = ('.jpg', '.jpeg', '.png', '.gif', '.bmp', '.tiff', '.webp')
+    return file_path.lower().endswith(image_extensions)
+
+
+def is_video(file_path):
+    if file_path.lower().endswith((".mp4", ".mov", ".avi", ".webm")) or osp.isdir(file_path):
+        return True
+    return False
+
+
+def is_template(file_path):
+    if file_path.endswith(".pkl"):
+        return True
+    return False
+
+
+def mkdir(d, log=False):
+    # return self-assined `d`, for one line code
+    if not osp.exists(d):
+        os.makedirs(d, exist_ok=True)
+        if log:
+            print(f"Make dir: {d}")
+    return d
+
+
+def squeeze_tensor_to_numpy(tensor):
+    out = tensor.data.squeeze(0).cpu().numpy()
+    return out
+
+
+def dct2device(dct: dict, device):
+    for key in dct:
+        if isinstance(dct[key], torch.Tensor):
+            dct[key] = dct[key].to(device)
+        else:
+            dct[key] = torch.tensor(dct[key]).to(device)
+    return dct
+
+
+def concat_feat(kp_source: torch.Tensor, kp_driving: torch.Tensor) -> torch.Tensor:
+    """
+    kp_source: (bs, k, 3)
+    kp_driving: (bs, k, 3)
+    Return: (bs, 2k*3)
+    """
+    bs_src = kp_source.shape[0]
+    bs_dri = kp_driving.shape[0]
+    assert bs_src == bs_dri, 'batch size must be equal'
+
+    feat = torch.cat([kp_source.view(bs_src, -1), kp_driving.view(bs_dri, -1)], dim=1)
+    return feat
+
+
+def remove_ddp_dumplicate_key(state_dict):
+    state_dict_new = OrderedDict()
+    for key in state_dict.keys():
+        state_dict_new[key.replace('module.', '')] = state_dict[key]
+    return state_dict_new
+
+
+def load_model(ckpt_path, model_config, device, model_type):
+    model_params = model_config['model_params'][f'{model_type}_params']
+
+    if model_type == 'appearance_feature_extractor':
+        model = AppearanceFeatureExtractor(**model_params).to(device)
+    elif model_type == 'motion_extractor':
+        model = MotionExtractor(**model_params).to(device)
+    elif model_type == 'warping_module':
+        model = WarpingNetwork(**model_params).to(device)
+    elif model_type == 'spade_generator':
+        model = SPADEDecoder(**model_params).to(device)
+    elif model_type == 'stitching_retargeting_module':
+        # Special handling for stitching and retargeting module
+        config = model_config['model_params']['stitching_retargeting_module_params']
+        checkpoint = torch.load(ckpt_path, map_location=lambda storage, loc: storage)
+
+        stitcher = StitchingRetargetingNetwork(**config.get('stitching'))
+        stitcher.load_state_dict(remove_ddp_dumplicate_key(checkpoint['retarget_shoulder']))
+        stitcher = stitcher.to(device)
+        stitcher.eval()
+
+        retargetor_lip = StitchingRetargetingNetwork(**config.get('lip'))
+        retargetor_lip.load_state_dict(remove_ddp_dumplicate_key(checkpoint['retarget_mouth']))
+        retargetor_lip = retargetor_lip.to(device)
+        retargetor_lip.eval()
+
+        retargetor_eye = StitchingRetargetingNetwork(**config.get('eye'))
+        retargetor_eye.load_state_dict(remove_ddp_dumplicate_key(checkpoint['retarget_eye']))
+        retargetor_eye = retargetor_eye.to(device)
+        retargetor_eye.eval()
+
+        return {
+            'stitching': stitcher,
+            'lip': retargetor_lip,
+            'eye': retargetor_eye
+        }
+    else:
+        raise ValueError(f"Unknown model type: {model_type}")
+
+    model.load_state_dict(torch.load(ckpt_path, map_location=lambda storage, loc: storage))
+    model.eval()
+    return model
+
+
+def load_description(fp):
+    with open(fp, 'r', encoding='utf-8') as f:
+        content = f.read()
+    return content
+
+
+def is_square_video(video_path):
+    video = cv2.VideoCapture(video_path)
+
+    width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+    video.release()
+    # if width != height:
+        # gr.Info(f"Uploaded video is not square, force do crop (driving) to be True")
+
+    return width == height
+
+def clean_state_dict(state_dict):
+    new_state_dict = OrderedDict()
+    for k, v in state_dict.items():
+        if k[:7] == 'module.':
+            k = k[7:]  # remove `module.`
+        new_state_dict[k] = v
+    return new_state_dict
diff --git a/src/utils/human_landmark_runner.py b/src/utils/human_landmark_runner.py
new file mode 100644
index 0000000..e25e5a4
--- /dev/null
+++ b/src/utils/human_landmark_runner.py
@@ -0,0 +1,95 @@
+# coding: utf-8
+
+import os.path as osp
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False)
+import torch
+import numpy as np
+import onnxruntime
+from .timer import Timer
+from .rprint import rlog
+from .crop import crop_image, _transform_pts
+
+
+def make_abs_path(fn):
+    return osp.join(osp.dirname(osp.realpath(__file__)), fn)
+
+
+def to_ndarray(obj):
+    if isinstance(obj, torch.Tensor):
+        return obj.cpu().numpy()
+    elif isinstance(obj, np.ndarray):
+        return obj
+    else:
+        return np.array(obj)
+
+
+class LandmarkRunner(object):
+    """landmark runner"""
+
+    def __init__(self, **kwargs):
+        ckpt_path = kwargs.get('ckpt_path')
+        onnx_provider = kwargs.get('onnx_provider', 'cuda')  # 默认用cuda
+        device_id = kwargs.get('device_id', 0)
+        self.dsize = kwargs.get('dsize', 224)
+        self.timer = Timer()
+
+        if onnx_provider.lower() == 'cuda':
+            self.session = onnxruntime.InferenceSession(
+                ckpt_path, providers=[
+                    ('CUDAExecutionProvider', {'device_id': device_id})
+                ]
+            )
+        elif onnx_provider.lower() == 'mps':
+            self.session = onnxruntime.InferenceSession(
+                ckpt_path, providers=[
+                    'CoreMLExecutionProvider'
+                ]
+            )
+        else:
+            opts = onnxruntime.SessionOptions()
+            opts.intra_op_num_threads = 4  # 默认线程数为 4
+            self.session = onnxruntime.InferenceSession(
+                ckpt_path, providers=['CPUExecutionProvider'],
+                sess_options=opts
+            )
+
+    def _run(self, inp):
+        out = self.session.run(None, {'input': inp})
+        return out
+
+    def run(self, img_rgb: np.ndarray, lmk=None):
+        if lmk is not None:
+            crop_dct = crop_image(img_rgb, lmk, dsize=self.dsize, scale=1.5, vy_ratio=-0.1)
+            img_crop_rgb = crop_dct['img_crop']
+        else:
+            # NOTE: force resize to 224x224, NOT RECOMMEND!
+            img_crop_rgb = cv2.resize(img_rgb, (self.dsize, self.dsize))
+            scale = max(img_rgb.shape[:2]) / self.dsize
+            crop_dct = {
+                'M_c2o': np.array([
+                    [scale, 0., 0.],
+                    [0., scale, 0.],
+                    [0., 0., 1.],
+                ], dtype=np.float32),
+            }
+
+        inp = (img_crop_rgb.astype(np.float32) / 255.).transpose(2, 0, 1)[None, ...]  # HxWx3 (BGR) -> 1x3xHxW (RGB!)
+
+        out_lst = self._run(inp)
+        out_pts = out_lst[2]
+
+        # 2d landmarks 203 points
+        lmk = to_ndarray(out_pts[0]).reshape(-1, 2) * self.dsize  # scale to 0-224
+        lmk = _transform_pts(lmk, M=crop_dct['M_c2o'])
+
+        return lmk
+
+    def warmup(self):
+        self.timer.tic()
+
+        dummy_image = np.zeros((1, 3, self.dsize, self.dsize), dtype=np.float32)
+
+        _ = self._run(dummy_image)
+
+        elapse = self.timer.toc()
+        rlog(f'LandmarkRunner warmup time: {elapse:.3f}s')
diff --git a/src/utils/io.py b/src/utils/io.py
new file mode 100644
index 0000000..28ca742
--- /dev/null
+++ b/src/utils/io.py
@@ -0,0 +1,117 @@
+# coding: utf-8
+
+import os.path as osp
+import imageio
+import numpy as np
+import pickle
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False)
+
+from .helper import mkdir, suffix
+
+
+def load_image_rgb(image_path: str):
+    if not osp.exists(image_path):
+        raise FileNotFoundError(f"Image not found: {image_path}")
+    img = cv2.imread(image_path, cv2.IMREAD_COLOR)
+    return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+
+
+def load_video(video_info, n_frames=-1):
+    reader = imageio.get_reader(video_info, "ffmpeg")
+
+    ret = []
+    for idx, frame_rgb in enumerate(reader):
+        if n_frames > 0 and idx >= n_frames:
+            break
+        ret.append(frame_rgb)
+
+    reader.close()
+    return ret
+
+
+def contiguous(obj):
+    if not obj.flags.c_contiguous:
+        obj = obj.copy(order="C")
+    return obj
+
+
+def resize_to_limit(img: np.ndarray, max_dim=1920, division=2):
+    """
+    ajust the size of the image so that the maximum dimension does not exceed max_dim, and the width and the height of the image are multiples of n.
+    :param img: the image to be processed.
+    :param max_dim: the maximum dimension constraint.
+    :param n: the number that needs to be multiples of.
+    :return: the adjusted image.
+    """
+    h, w = img.shape[:2]
+
+    # ajust the size of the image according to the maximum dimension
+    if max_dim > 0 and max(h, w) > max_dim:
+        if h > w:
+            new_h = max_dim
+            new_w = int(w * (max_dim / h))
+        else:
+            new_w = max_dim
+            new_h = int(h * (max_dim / w))
+        img = cv2.resize(img, (new_w, new_h))
+
+    # ensure that the image dimensions are multiples of n
+    division = max(division, 1)
+    new_h = img.shape[0] - (img.shape[0] % division)
+    new_w = img.shape[1] - (img.shape[1] % division)
+
+    if new_h == 0 or new_w == 0:
+        # when the width or height is less than n, no need to process
+        return img
+
+    if new_h != img.shape[0] or new_w != img.shape[1]:
+        img = img[:new_h, :new_w]
+
+    return img
+
+
+def load_img_online(obj, mode="bgr", **kwargs):
+    max_dim = kwargs.get("max_dim", 1920)
+    n = kwargs.get("n", 2)
+    if isinstance(obj, str):
+        if mode.lower() == "gray":
+            img = cv2.imread(obj, cv2.IMREAD_GRAYSCALE)
+        else:
+            img = cv2.imread(obj, cv2.IMREAD_COLOR)
+    else:
+        img = obj
+
+    # Resize image to satisfy constraints
+    img = resize_to_limit(img, max_dim=max_dim, division=n)
+
+    if mode.lower() == "bgr":
+        return contiguous(img)
+    elif mode.lower() == "rgb":
+        return contiguous(img[..., ::-1])
+    else:
+        raise Exception(f"Unknown mode {mode}")
+
+
+def load(fp):
+    suffix_ = suffix(fp)
+
+    if suffix_ == "npy":
+        return np.load(fp)
+    elif suffix_ == "pkl":
+        return pickle.load(open(fp, "rb"))
+    else:
+        raise Exception(f"Unknown type: {suffix}")
+
+
+def dump(wfp, obj):
+    wd = osp.split(wfp)[0]
+    if wd != "" and not osp.exists(wd):
+        mkdir(wd)
+
+    _suffix = suffix(wfp)
+    if _suffix == "npy":
+        np.save(wfp, obj)
+    elif _suffix == "pkl":
+        pickle.dump(obj, open(wfp, "wb"))
+    else:
+        raise Exception("Unknown type: {}".format(_suffix))
diff --git a/src/utils/resources/clip_embedding_68.pkl b/src/utils/resources/clip_embedding_68.pkl
new file mode 100644
index 0000000..d69ca7f
Binary files /dev/null and b/src/utils/resources/clip_embedding_68.pkl differ
diff --git a/src/utils/resources/clip_embedding_9.pkl b/src/utils/resources/clip_embedding_9.pkl
new file mode 100644
index 0000000..2b1e6ae
Binary files /dev/null and b/src/utils/resources/clip_embedding_9.pkl differ
diff --git a/src/utils/resources/lip_array.pkl b/src/utils/resources/lip_array.pkl
new file mode 100644
index 0000000..61fe803
Binary files /dev/null and b/src/utils/resources/lip_array.pkl differ
diff --git a/src/utils/resources/mask_template.png b/src/utils/resources/mask_template.png
new file mode 100644
index 0000000..bca6ca5
Binary files /dev/null and b/src/utils/resources/mask_template.png differ
diff --git a/src/utils/retargeting_utils.py b/src/utils/retargeting_utils.py
new file mode 100644
index 0000000..cbe9d1c
--- /dev/null
+++ b/src/utils/retargeting_utils.py
@@ -0,0 +1,24 @@
+
+"""
+Functions to compute distance ratios between specific pairs of facial landmarks
+"""
+
+import numpy as np
+
+
+def calculate_distance_ratio(lmk: np.ndarray, idx1: int, idx2: int, idx3: int, idx4: int, eps: float = 1e-6) -> np.ndarray:
+    return (np.linalg.norm(lmk[:, idx1] - lmk[:, idx2], axis=1, keepdims=True) /
+            (np.linalg.norm(lmk[:, idx3] - lmk[:, idx4], axis=1, keepdims=True) + eps))
+
+
+def calc_eye_close_ratio(lmk: np.ndarray, target_eye_ratio: np.ndarray = None) -> np.ndarray:
+    lefteye_close_ratio = calculate_distance_ratio(lmk, 6, 18, 0, 12)
+    righteye_close_ratio = calculate_distance_ratio(lmk, 30, 42, 24, 36)
+    if target_eye_ratio is not None:
+        return np.concatenate([lefteye_close_ratio, righteye_close_ratio, target_eye_ratio], axis=1)
+    else:
+        return np.concatenate([lefteye_close_ratio, righteye_close_ratio], axis=1)
+
+
+def calc_lip_close_ratio(lmk: np.ndarray) -> np.ndarray:
+    return calculate_distance_ratio(lmk, 90, 102, 48, 66)
diff --git a/src/utils/rprint.py b/src/utils/rprint.py
new file mode 100644
index 0000000..18614e8
--- /dev/null
+++ b/src/utils/rprint.py
@@ -0,0 +1,16 @@
+# coding: utf-8
+
+"""
+custom print and log functions 
+"""
+
+__all__ = ['rprint', 'rlog']
+
+try:
+    from rich.console import Console
+    console = Console()
+    rprint = console.print
+    rlog = console.log
+except:
+    rprint = print
+    rlog = print
diff --git a/src/utils/timer.py b/src/utils/timer.py
new file mode 100644
index 0000000..792af24
--- /dev/null
+++ b/src/utils/timer.py
@@ -0,0 +1,29 @@
+# coding: utf-8
+
+"""
+tools to measure elapsed time
+"""
+
+import time
+
+class Timer(object):
+    """A simple timer."""
+
+    def __init__(self):
+        self.total_time = 0.
+        self.calls = 0
+        self.start_time = 0.
+        self.diff = 0.
+
+    def tic(self):
+        # using time.time instead of time.clock because time time.clock
+        # does not normalize for multithreading
+        self.start_time = time.time()
+
+    def toc(self, average=True):
+        self.diff = time.time() - self.start_time
+        return self.diff
+
+    def clear(self):
+        self.start_time = 0.
+        self.diff = 0.
diff --git a/src/utils/video.py b/src/utils/video.py
new file mode 100644
index 0000000..9331c90
--- /dev/null
+++ b/src/utils/video.py
@@ -0,0 +1,218 @@
+# coding: utf-8
+
+"""
+Functions for processing video
+
+ATTENTION: you need to install ffmpeg and ffprobe in your env!
+"""
+
+import os.path as osp
+import numpy as np
+import subprocess
+import imageio
+import cv2
+from rich.progress import track
+
+from .rprint import rlog as log
+from .rprint import rprint as print
+from .helper import prefix
+
+
+def exec_cmd(cmd):
+    return subprocess.run(cmd, shell=True, check=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+
+
+def images2video(images, wfp, **kwargs):
+    fps = kwargs.get('fps', 30)
+    video_format = kwargs.get('format', 'mp4')  # default is mp4 format
+    codec = kwargs.get('codec', 'libx264')  # default is libx264 encoding
+    quality = kwargs.get('quality')  # video quality
+    pixelformat = kwargs.get('pixelformat', 'yuv420p')  # video pixel format
+    image_mode = kwargs.get('image_mode', 'rgb')
+    macro_block_size = kwargs.get('macro_block_size', 2)
+    ffmpeg_params = ['-crf', str(kwargs.get('crf', 18))]
+
+    writer = imageio.get_writer(
+        wfp, fps=fps, format=video_format,
+        codec=codec, quality=quality, ffmpeg_params=ffmpeg_params, pixelformat=pixelformat, macro_block_size=macro_block_size
+    )
+
+    n = len(images)
+    for i in track(range(n), description='Writing', transient=True):
+        if image_mode.lower() == 'bgr':
+            writer.append_data(images[i][..., ::-1])
+        else:
+            writer.append_data(images[i])
+
+    writer.close()
+
+
+def video2gif(video_fp, fps=30, size=256):
+    if osp.exists(video_fp):
+        d = osp.split(video_fp)[0]
+        fn = prefix(osp.basename(video_fp))
+        palette_wfp = osp.join(d, 'palette.png')
+        gif_wfp = osp.join(d, f'{fn}.gif')
+        # generate the palette
+        cmd = f'ffmpeg -i "{video_fp}" -vf "fps={fps},scale={size}:-1:flags=lanczos,palettegen" "{palette_wfp}" -y'
+        exec_cmd(cmd)
+        # use the palette to generate the gif
+        cmd = f'ffmpeg -i "{video_fp}" -i "{palette_wfp}" -filter_complex "fps={fps},scale={size}:-1:flags=lanczos[x];[x][1:v]paletteuse" "{gif_wfp}" -y'
+        exec_cmd(cmd)
+        return gif_wfp
+    else:
+        raise FileNotFoundError(f"video_fp: {video_fp} not exists!")
+
+
+def merge_audio_video(video_fp, audio_fp, wfp):
+    if osp.exists(video_fp) and osp.exists(audio_fp):
+        cmd = f'ffmpeg -i "{video_fp}" -i "{audio_fp}" -c:v copy -c:a aac "{wfp}" -y'
+        exec_cmd(cmd)
+        print(f'merge {video_fp} and {audio_fp} to {wfp}')
+    else:
+        print(f'video_fp: {video_fp} or audio_fp: {audio_fp} not exists!')
+
+
+def blend(img: np.ndarray, mask: np.ndarray, background_color=(255, 255, 255)):
+    mask_float = mask.astype(np.float32) / 255.
+    background_color = np.array(background_color).reshape([1, 1, 3])
+    bg = np.ones_like(img) * background_color
+    img = np.clip(mask_float * img + (1 - mask_float) * bg, 0, 255).astype(np.uint8)
+    return img
+
+
+def concat_frames(driving_image_lst, source_image_lst, I_p_lst):
+    # TODO: add more concat style, e.g., left-down corner driving
+    out_lst = []
+    h, w, _ = I_p_lst[0].shape
+    source_image_resized_lst = [cv2.resize(img, (w, h)) for img in source_image_lst]
+
+    for idx, _ in track(enumerate(I_p_lst), total=len(I_p_lst), description='Concatenating result...'):
+        I_p = I_p_lst[idx]
+        source_image_resized = source_image_resized_lst[idx] if len(source_image_lst) > 1 else source_image_resized_lst[0]
+
+        if driving_image_lst is None:
+            out = np.hstack((source_image_resized, I_p))
+        else:
+            driving_image = driving_image_lst[idx]
+            driving_image_resized = cv2.resize(driving_image, (w, h))
+            out = np.hstack((driving_image_resized, source_image_resized, I_p))
+
+        out_lst.append(out)
+    return out_lst
+
+
+class VideoWriter:
+    def __init__(self, **kwargs):
+        self.fps = kwargs.get('fps', 30)
+        self.wfp = kwargs.get('wfp', 'video.mp4')
+        self.video_format = kwargs.get('format', 'mp4')
+        self.codec = kwargs.get('codec', 'libx264')
+        self.quality = kwargs.get('quality')
+        self.pixelformat = kwargs.get('pixelformat', 'yuv420p')
+        self.image_mode = kwargs.get('image_mode', 'rgb')
+        self.ffmpeg_params = kwargs.get('ffmpeg_params')
+
+        self.writer = imageio.get_writer(
+            self.wfp, fps=self.fps, format=self.video_format,
+            codec=self.codec, quality=self.quality,
+            ffmpeg_params=self.ffmpeg_params, pixelformat=self.pixelformat
+        )
+
+    def write(self, image):
+        if self.image_mode.lower() == 'bgr':
+            self.writer.append_data(image[..., ::-1])
+        else:
+            self.writer.append_data(image)
+
+    def close(self):
+        if self.writer is not None:
+            self.writer.close()
+
+
+def change_video_fps(input_file, output_file, fps=20, codec='libx264', crf=12):
+    cmd = f'ffmpeg -i "{input_file}" -c:v {codec} -crf {crf} -r {fps} "{output_file}" -y'
+    exec_cmd(cmd)
+
+
+def get_fps(filepath, default_fps=25):
+    try:
+        fps = cv2.VideoCapture(filepath).get(cv2.CAP_PROP_FPS)
+
+        if fps in (0, None):
+            fps = default_fps
+    except Exception as e:
+        log(e)
+        fps = default_fps
+
+    return fps
+
+
+def has_audio_stream(video_path: str) -> bool:
+    """
+    Check if the video file contains an audio stream.
+
+    :param video_path: Path to the video file
+    :return: True if the video contains an audio stream, False otherwise
+    """
+    if osp.isdir(video_path):
+        return False
+
+    cmd = [
+        'ffprobe',
+        '-v', 'error',
+        '-select_streams', 'a',
+        '-show_entries', 'stream=codec_type',
+        '-of', 'default=noprint_wrappers=1:nokey=1',
+        f'"{video_path}"'
+    ]
+
+    try:
+        # result = subprocess.run(cmd, capture_output=True, text=True)
+        result = exec_cmd(' '.join(cmd))
+        if result.returncode != 0:
+            log(f"Error occurred while probing video: {result.stderr}")
+            return False
+
+        # Check if there is any output from ffprobe command
+        return bool(result.stdout.strip())
+    except Exception as e:
+        log(
+            f"Error occurred while probing video: {video_path}, "
+            "you may need to install ffprobe! (https://ffmpeg.org/download.html) "
+            "Now set audio to false!",
+            style="bold red"
+        )
+    return False
+
+
+def add_audio_to_video(silent_video_path: str, audio_video_path: str, output_video_path: str):
+    cmd = [
+        'ffmpeg',
+        '-y',
+        '-i', f'"{silent_video_path}"',
+        '-i', f'"{audio_video_path}"',
+        '-map', '0:v',
+        '-map', '1:a',
+        '-c:v', 'copy',
+        '-shortest',
+        f'"{output_video_path}"'
+    ]
+
+    try:
+        exec_cmd(' '.join(cmd))
+        log(f"Video with audio generated successfully: {output_video_path}")
+    except subprocess.CalledProcessError as e:
+        log(f"Error occurred: {e}")
+
+
+def bb_intersection_over_union(boxA, boxB):
+    xA = max(boxA[0], boxB[0])
+    yA = max(boxA[1], boxB[1])
+    xB = min(boxA[2], boxB[2])
+    yB = min(boxA[3], boxB[3])
+    interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
+    boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)
+    boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)
+    iou = interArea / float(boxAArea + boxBArea - interArea)
+    return iou
diff --git a/src/utils/viz.py b/src/utils/viz.py
new file mode 100644
index 0000000..59b35d2
--- /dev/null
+++ b/src/utils/viz.py
@@ -0,0 +1,19 @@
+# coding: utf-8
+
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False)
+
+
+def viz_lmk(img_, vps, **kwargs):
+    """可视化点"""
+    lineType = kwargs.get("lineType", cv2.LINE_8)  # cv2.LINE_AA
+    img_for_viz = img_.copy()
+    for pt in vps:
+        cv2.circle(
+            img_for_viz,
+            (int(pt[0]), int(pt[1])),
+            radius=kwargs.get("radius", 1),
+            color=(0, 255, 0),
+            thickness=kwargs.get("thickness", 1),
+            lineType=lineType,
+        )
+    return img_for_viz