diff --git a/flux_style_shaper_api/flux_precisecam_demo.py b/flux_style_shaper_api/flux_precisecam_demo.py
new file mode 100644
index 00000000..44055804
--- /dev/null
+++ b/flux_style_shaper_api/flux_precisecam_demo.py
@@ -0,0 +1,253 @@
+import gradio as gr
+import numpy as np
+import torch
+import os
+from PIL import Image
+from perspective_fields import pano_utils as pu
+from diffusers import (
+    AutoencoderKL,
+    ControlNetModel,
+    FluxControlNetPipeline,
+    FlowMatchEulerDiscreteScheduler,
+    FluxControlNetModel,
+)
+from transformers import CLIPTextModel, CLIPTokenizer, CLIPImageProcessor, CLIPVisionModelWithProjection
+from transformers import T5EncoderModel, T5TokenizerFast
+from safetensors.torch import load_file
+# 加载本地模型文件路径
+current_dir = os.path.dirname(os.path.abspath(__file__))
+project_dir = os.path.dirname(current_dir)
+
+# 修正路径格式，使用os.path.join确保路径分隔符正确
+flux_model_dir = os.path.join(project_dir, "models", "diffusion_models")
+
+print(f"Loading models from:")
+print(f"Flux dir: {flux_model_dir}")
+
+# 创建缓存目录（如果不存在）
+os.makedirs(os.path.join(project_dir, "cache"), exist_ok=True)
+
+# 加载所有需要的组件
+print("Loading components...")
+
+# 1. 加载调度器
+scheduler = FlowMatchEulerDiscreteScheduler.from_config(
+    {
+        "num_train_timesteps": 1000
+    }
+)
+
+# 5. 加载ControlNet - 使用与FLUX.1-dev兼容的模型
+print("Loading ControlNet compatible with FLUX.1-dev...")
+controlnet_model = "InstantX/FLUX.1-dev-controlnet-canny"
+controlnet = FluxControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16,cache_dir=os.path.join(project_dir, "cache"))
+
+
+# 2. 加载VAE
+try:
+    vae = AutoencoderKL.from_pretrained(
+        "diffusers/FLUX.1-vae", 
+        torch_dtype=torch.bfloat16,  # 使用bfloat16匹配controlnet
+        cache_dir=os.path.join(project_dir, "cache")
+    )
+except Exception as e:
+    print(f"Error loading VAE: {e}")
+    print("Loading default VAE...")
+    vae = AutoencoderKL.from_pretrained(
+        "diffusers/FLUX.1-vae",
+        torch_dtype=torch.bfloat16,  # 使用bfloat16匹配controlnet
+        cache_dir=os.path.join(project_dir, "cache")
+    )
+
+# 3. 加载文本编码器和标记器
+text_encoder = CLIPTextModel.from_pretrained(
+    "openai/clip-vit-large-patch14",
+    torch_dtype=torch.bfloat16,  # 使用bfloat16匹配controlnet
+    cache_dir=os.path.join(project_dir, "cache")
+)
+tokenizer = CLIPTokenizer.from_pretrained(
+    "openai/clip-vit-large-patch14",
+    cache_dir=os.path.join(project_dir, "cache")
+)
+
+# 4. 加载第二个文本编码器和标记器
+print("从本地加载预先下载的T5模型...")
+# 指定本地模型路径 - 假设已经手动下载了魔搭的模型
+t5_model_path = os.path.join(project_dir, "models", "t5_xxl")  # 根据实际路径调整
+
+# 检查模型路径是否存在
+if not os.path.exists(t5_model_path):
+    print(f"警告: 本地模型路径 {t5_model_path} 不存在。请先手动下载魔搭的T5-XXL模型到此路径。")
+    print("继续尝试从HuggingFace下载，但这可能会很慢...")
+    t5_model_path = "google/t5-v1_1-xxl"
+
+text_encoder_2 = T5EncoderModel.from_pretrained(
+    t5_model_path,  # 使用本地路径
+    torch_dtype=torch.bfloat16,
+    cache_dir=os.path.join(project_dir, "cache")
+)
+tokenizer_2 = T5TokenizerFast.from_pretrained(
+    t5_model_path,  # 使用本地路径
+    cache_dir=os.path.join(project_dir, "cache")
+)
+
+
+
+# 6. 加载特征提取器和图像编码器
+feature_extractor = CLIPImageProcessor.from_pretrained(
+    "openai/clip-vit-large-patch14",
+    cache_dir=os.path.join(project_dir, "cache")
+)
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+    "openai/clip-vit-large-patch14",
+    torch_dtype=torch.bfloat16,  # 使用bfloat16匹配controlnet
+    cache_dir=os.path.join(project_dir, "cache")
+)
+
+# 7. 加载PipelineImageProcessor (transformer)
+base_model = "black-forest-labs/FLUX.1-dev"
+
+# 不加载transformer组件，因为这会导致类型错误
+# pipeline会自动初始化正确类型的transformer
+
+# 8. 加载Pipeline - 直接从HuggingFace加载
+try:
+    print("Loading FLUX model from HuggingFace...")
+    # 设置缓存相关环境变量
+    os.environ["HF_HOME"] = os.path.join(project_dir, "cache")
+    os.environ["TRANSFORMERS_CACHE"] = os.path.join(project_dir, "cache", "transformers")
+    os.environ["DIFFUSERS_CACHE"] = os.path.join(project_dir, "cache", "diffusers")
+    # 设置缓存路径参数
+    cache_path = os.path.join(project_dir, "cache")
+    
+    pipe = FluxControlNetPipeline.from_pretrained(
+        base_model,
+        scheduler=scheduler,
+        vae=vae,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        text_encoder_2=text_encoder_2,
+        tokenizer_2=tokenizer_2,
+        feature_extractor=feature_extractor,
+        image_encoder=image_encoder,
+        controlnet=controlnet,
+        torch_dtype=torch.bfloat16,  # 使用bfloat16匹配controlnet
+        low_cpu_mem_usage=False,
+        cache_dir=cache_path,
+        resume_download=True,  # 如果下载中断，尝试继续下载
+        force_download=False   # 如果已缓存，不重新下载
+    )
+except Exception as e:
+    print(f"Error loading from HuggingFace: {e}")
+    print("Trying alternative configuration...")
+
+# 启用CPU卸载以减少GPU内存使用
+pipe.enable_model_cpu_offload()
+
+def inference(prompt, pf_image, n_steps=2, seed=13):
+    """Generates an image based on the given prompt and perspective field image."""
+    pf = Image.fromarray(
+        np.concatenate(
+            [np.expand_dims(pf_image[:, :, i], axis=-1) for i in [2, 1, 0]], axis=2
+        )
+    )
+    pf_condition = pf.resize((1024, 1024))
+    generator = torch.manual_seed(seed)
+    print("Control image size:", pf_condition.size)  # 应该是 (1024, 1024)
+    return pipe(
+        prompt=prompt,
+        num_inference_steps=n_steps,
+        generator=generator,
+        control_image=pf_condition,  # 只保留ControlNet的输入参数
+        # 移除不支持的参数
+        # image=pf_condition,  # 需要为img2img提供一个基础图像
+        # strength=0.8,  # 控制处理强度
+    ).images[0]
+
+
+def obtain_pf(xi, roll, pitch, vfov):
+    """Computes perspective fields given camera parameters."""
+    w, h = (1024, 1024)
+    equi_img = np.zeros((h, w, 3), dtype=np.uint8)
+    x = -np.sin(np.radians(vfov / 2))
+    z = np.sqrt(1 - x**2)
+    f_px_effective = -0.5 * (w / 2) * (xi + z) / x
+
+    crop, _, _, _, up, lat, _ = pu.crop_distortion(
+        equi_img, f=f_px_effective, xi=xi, H=h, W=w, az=10, el=-pitch, roll=roll
+    )
+
+    gravity = (up + 1) * 127.5
+    latitude = np.expand_dims((np.degrees(lat) + 90) * (255 / 180), axis=-1)
+    pf_image = np.concatenate([gravity, latitude], axis=2).astype(np.uint8)
+    
+    # 尝试生成blend图像，如果失败则创建一个替代的blend图像
+    try:
+        blend = pu.draw_perspective_fields(crop, up, np.radians(np.degrees(lat)))
+    except Exception as e:
+        print(f"Error generating blend image: {e}")
+        # 创建一个替代的blend图像
+        blend = np.zeros_like(pf_image)
+        # 添加一些基本的可视化
+        blend[:,:,0] = (np.degrees(lat) + 90) * (255 / 180)  # 使用latitude作为红色通道
+        blend[:,:,1] = (up + 1) * 127.5  # 使用up作为绿色通道
+        blend[:,:,2] = 128  # 蓝色通道设置为中性值
+
+    return pf_image, blend
+
+
+# Gradio UI
+demo = gr.Blocks(theme=gr.themes.Soft())
+
+with demo:
+    gr.Markdown("""---""")
+    gr.Markdown("""# PreciseCam: Precise Camera Control for Text-to-Image Generation""")
+    gr.Markdown("""1. Set the camera parameters (Roll, Pitch, Vertical FOV, ξ)""")
+    gr.Markdown("""2. Click "Compute PF-US" to generate the perspective field image""")
+    gr.Markdown("""3. Enter a prompt for the image generation""")
+    gr.Markdown("""4. Click "Generate Image" to create the final image""")
+    gr.Markdown("""---""")
+
+    with gr.Row():
+        with gr.Column():
+            roll = gr.Slider(-90, 90, 0, label="Roll")
+            pitch = gr.Slider(-90, 90, 1, label="Pitch")
+            vfov = gr.Slider(15, 140, 50, label="Vertical FOV")
+            xi = gr.Slider(0.0, 1, 0.2, label="ξ")
+            prompt = gr.Textbox(
+                lines=4,
+                label="Prompt",
+                show_copy_button=True,
+                value="A colorful autumn park with leaves of orange, red, and yellow scattered across a winding path.",
+            )
+            pf_btn = gr.Button("Compute PF-US", variant="primary")
+        with gr.Row():
+            pf_img = gr.Image(height=1024 // 4, width=1024 // 4, label="PF-US")
+            condition_img = gr.Image(
+                height=1024 // 4, width=1024 // 4, label="Internal PF-US (RGB)"
+            )
+
+        with gr.Column(scale=2):
+            result_img = gr.Image(label="Generated Image", height=1024 // 2)
+            inf_btn = gr.Button("Generate Image", variant="primary")
+    gr.Markdown("""---""")
+
+    pf_btn.click(
+        obtain_pf, inputs=[xi, roll, pitch, vfov], outputs=[condition_img, pf_img]
+    )
+    inf_btn.click(inference, inputs=[prompt, condition_img], outputs=[result_img])
+
+# 设置环境变量，禁用代理
+os.environ["HTTP_PROXY"] = ""
+os.environ["HTTPS_PROXY"] = ""
+os.environ["NO_PROXY"] = "localhost,127.0.0.1"
+
+# 启动Gradio应用，指定内联选项和服务器名称
+demo.launch(
+    server_name="0.0.0.0",
+    server_port=9946,
+    share=False,
+    favicon_path=None,
+    show_api=False,
+    prevent_thread_lock=True
+)
diff --git a/flux_style_shaper_api/perspective_fields/README.md b/flux_style_shaper_api/perspective_fields/README.md
new file mode 100644
index 00000000..b82f2316
--- /dev/null
+++ b/flux_style_shaper_api/perspective_fields/README.md
@@ -0,0 +1,11 @@
+Code based on the original implementation from [jinlinyi/PerspectiveFields](https://github.com/jinlinyi/PerspectiveFields).
+
+**Citation:**
+```bibtex
+@inproceedings{jin2023perspective,
+      title={Perspective Fields for Single Image Camera Calibration},
+      author={Linyi Jin and Jianming Zhang and Yannick Hold-Geoffroy and Oliver Wang and Kevin Matzen and Matthew Sticha and David F. Fouhey},
+      booktitle = {CVPR},
+      year={2023}
+}
+```
\ No newline at end of file
diff --git a/flux_style_shaper_api/perspective_fields/pano_utils.py b/flux_style_shaper_api/perspective_fields/pano_utils.py
new file mode 100644
index 00000000..b528dd10
--- /dev/null
+++ b/flux_style_shaper_api/perspective_fields/pano_utils.py
@@ -0,0 +1,261 @@
+#######################################################################
+# Adapted code from https://github.com/jinlinyi/PerspectiveFields
+#######################################################################
+import torch 
+import numpy as np
+from equilib import grid_sample
+from sklearn.preprocessing import normalize
+from perspective_fields.visualizer import VisualizerPerspective
+from numpy.lib.scimath import sqrt as csqrt
+
+def deg2rad(deg):
+    """convert degrees to radians"""
+    return deg * np.pi / 180
+
+def diskradius(xi, f):  # compute the disk radius when the image is catadioptric
+    return np.sqrt(-(f * f) / (1 - xi * xi))
+
+def crop_distortion(image360_path, f, xi, H, W, az, el, roll):
+        """
+        Reference: https://github.com/dompm/spherical-distortion-dataset/blob/main/spherical_distortion/spherical_distortion.py
+        Crop distorted image with specified camera parameters
+
+        Args:
+            image360_path (str): path to image which to crop from
+            f (float): focal_length of cropped image
+            xi:
+            H (int): height of cropped image
+            W (int): width of cropped image
+            az: camera rotation about camera frame y-axis of cropped image (degrees)
+            el: camera rotation about camera frame x-axis of cropped image (degrees)
+            roll: camera rotation about camera frame z-axis of cropped image (degrees)
+        Returns:
+            im (np.ndarray): cropped, distorted image
+        """
+
+        u0 = W / 2.0
+        v0 = H / 2.0
+
+        grid_x, grid_y = np.meshgrid(list(range(W)), list(range(H)))
+
+        image360 = image360_path.copy()
+
+        ImPano_W = np.shape(image360)[1]
+        ImPano_H = np.shape(image360)[0]
+        x_ref = 1
+        y_ref = 1
+
+        fmin = minfocal(
+            u0, v0, xi, x_ref, y_ref
+        )  # compute minimal focal length for the image to ve catadioptric with given xi
+
+        # 1. Projection on the camera plane
+
+        X_Cam = np.divide(grid_x - u0, f)
+        Y_Cam = -np.divide(grid_y - v0, f)
+
+        # 2. Projection on the sphere
+
+        AuxVal = np.multiply(X_Cam, X_Cam) + np.multiply(Y_Cam, Y_Cam)
+
+        alpha_cam = np.real(xi + csqrt(1 + np.multiply((1 - xi * xi), AuxVal)))
+
+        alpha_div = AuxVal + 1
+
+        alpha_cam_div = np.divide(alpha_cam, alpha_div)
+
+        X_Sph = np.multiply(X_Cam, alpha_cam_div)
+        Y_Sph = np.multiply(Y_Cam, alpha_cam_div)
+        Z_Sph = alpha_cam_div - xi
+
+        # 3. Rotation of the sphere
+        coords = np.vstack((X_Sph.ravel(), Y_Sph.ravel(), Z_Sph.ravel()))
+        rot_el = np.array(
+            [
+                1.0,
+                0.0,
+                0.0,
+                0.0,
+                np.cos(deg2rad(el)),
+                -np.sin(deg2rad(el)),
+                0.0,
+                np.sin(deg2rad(el)),
+                np.cos(deg2rad(el)),
+            ]
+        ).reshape((3, 3))
+        rot_az = np.array(
+            [
+                np.cos(deg2rad(az)),
+                0.0,
+                -np.sin(deg2rad(az)),
+                0.0,
+                1.0,
+                0.0,
+                np.sin(deg2rad(az)),
+                0.0,
+                np.cos(deg2rad(az)),
+            ]
+        ).reshape((3, 3))
+        rot_roll = np.array(
+            [
+                np.cos(deg2rad(roll)),
+                np.sin(deg2rad(roll)),
+                0.0,
+                -np.sin(deg2rad(roll)),
+                np.cos(deg2rad(roll)),
+                0.0,
+                0.0,
+                0.0,
+                1.0,
+            ]
+        ).reshape((3, 3))
+        sph = rot_el.dot(rot_roll.T.dot(coords))
+        sph = rot_az.dot(sph)
+
+        sph = sph.reshape((3, H, W)).transpose((1, 2, 0))
+        X_Sph, Y_Sph, Z_Sph = sph[:, :, 0], sph[:, :, 1], sph[:, :, 2]
+
+        # 4. cart 2 sph
+        ntheta = np.arctan2(X_Sph, Z_Sph)
+        nphi = np.arctan2(Y_Sph, np.sqrt(Z_Sph**2 + X_Sph**2))
+
+        pi = np.pi
+
+        # 5. Sphere to pano
+        min_theta = -pi
+        max_theta = pi
+        min_phi = -pi / 2.0
+        max_phi = pi / 2.0
+
+        min_x = 0
+        max_x = ImPano_W - 1.0
+        min_y = 0
+        max_y = ImPano_H - 1.0
+
+        ## for x
+        a = (max_theta - min_theta) / (max_x - min_x)
+        b = max_theta - a * max_x  # from y=ax+b %% -a;
+        nx = (1.0 / a) * (ntheta - b)
+
+        ## for y
+        a = (min_phi - max_phi) / (max_y - min_y)
+        b = max_phi - a * min_y  # from y=ax+b %% -a;
+        ny = (1.0 / a) * (nphi - b)
+        lat = nphi.copy()
+        xy_map = np.stack((nx, ny)).transpose(1, 2, 0)
+
+        # 6. Final step interpolation and mapping
+        # im = np.array(my_interpol.interp2linear(image360, nx, ny), dtype=np.uint8)
+        im = grid_sample.numpy_grid_sample.default(
+            image360.transpose(2, 0, 1), np.stack((ny, nx))
+        ).transpose(1, 2, 0)
+        if (
+            f < fmin
+        ):  # if it is a catadioptric image, apply mask and a disk in the middle
+            r = diskradius(xi, f)
+            DIM = im.shape
+            ci = (np.round(DIM[0] / 2), np.round(DIM[1] / 2))
+            xx, yy = np.meshgrid(
+                list(range(DIM[0])) - ci[0], list(range(DIM[1])) - ci[1]
+            )
+            mask = np.double((np.multiply(xx, xx) + np.multiply(yy, yy)) < r * r)
+            mask_3channel = np.stack([mask, mask, mask], axis=-1).transpose((1, 0, 2))
+            im = np.array(np.multiply(im, mask_3channel), dtype=np.uint8)
+
+        col = nphi[:, W // 2]
+        zero_crossings_rows = np.where(np.diff(np.sign(col)))[0]
+        if len(zero_crossings_rows) >= 2:
+            print("WARNING | Number of zero crossings:", len(zero_crossings_rows))
+            zero_crossings_rows = [zero_crossings_rows[0]]
+
+        if len(zero_crossings_rows) == 0:
+            offset = np.nan
+        else:
+            assert col[zero_crossings_rows[0]] >= 0
+            assert col[zero_crossings_rows[0] + 1] <= 0
+            dy = col[zero_crossings_rows[0] + 1] - col[zero_crossings_rows[0]]
+            offset = zero_crossings_rows[0] - col[zero_crossings_rows[0]] / dy
+            assert col[zero_crossings_rows[0]] / dy <= 1.0
+        # Reproject [nx, ny+epsilon] back
+        epsilon = 1e-5
+        end_vector_x = nx.copy()
+        end_vector_y = ny.copy() - epsilon
+        # -5. pano to Sphere
+        a = (max_theta - min_theta) / (max_x - min_x)
+        b = max_theta - a * max_x  # from y=ax+b %% -a;
+        ntheta_end = end_vector_x * a + b
+        ## for y
+        a = (min_phi - max_phi) / (max_y - min_y)
+        b = max_phi - a * min_y
+        nphi_end = end_vector_y * a + b
+        # -4. sph 2 cart
+        Y_Sph = np.sin(nphi)
+        X_Sph = np.cos(nphi_end) * np.sin(ntheta_end)
+        Z_Sph = np.cos(nphi_end) * np.cos(ntheta_end)
+        # -3. Reverse Rotation of the sphere
+        coords = np.vstack((X_Sph.ravel(), Y_Sph.ravel(), Z_Sph.ravel()))
+        sph = rot_roll.dot(rot_el.T.dot(rot_az.T.dot(coords)))
+        sph = sph.reshape((3, H, W)).transpose((1, 2, 0))
+        X_Sph, Y_Sph, Z_Sph = sph[:, :, 0], sph[:, :, 1], sph[:, :, 2]
+
+        # -1. Projection on the image plane
+
+        X_Cam = X_Sph * f / (xi * csqrt(X_Sph**2 + Y_Sph**2 + Z_Sph**2) + Z_Sph) + u0
+        Y_Cam = -Y_Sph * f / (xi * csqrt(X_Sph**2 + Y_Sph**2 + Z_Sph**2) + Z_Sph) + v0
+        up = np.stack((X_Cam - grid_x, Y_Cam - grid_y)).transpose(1, 2, 0)
+        up = normalize(up.reshape(-1, 2)).reshape(up.shape)
+
+        return im, ntheta, nphi, offset, up, lat, xy_map
+
+def minfocal(u0, v0, xi, xref=1, yref=1):
+    """compute the minimum focal for the image to be catadioptric given xi"""
+    value = -(1 - xi * xi) * ((xref - u0) * (xref - u0) + (yref - v0) * (yref - v0))
+
+    if value < 0:
+        return 0
+    else:
+        return np.sqrt(value) * 1.0001
+
+#------------------------------------------------
+
+def draw_perspective_fields(
+    img_rgb, up, latimap, color=None, density=10, arrow_inv_len=20, return_img=True
+):
+    """draw perspective field on top of input image
+
+    Args:
+        img_rgb (np.ndarray): input image
+        up (np.ndarray): gravity field (h, w, 2)
+        latimap (np.ndarray): latitude map (h, w) (radians)
+        color ((float, float, float), optional): RGB color for up vectors. [0, 1]
+                                                 Defaults to None.
+        density (int, optional): Value to control density of up vectors.
+                                 Each row has (width // density) vectors.
+                                 Each column has (height // density) vectors.
+                                 Defaults to 10.
+        arrow_inv_len (int, optional): Value to control vector length
+                                       Vector length set to (image plane diagonal // arrow_inv_len).
+                                       Defaults to 20.
+        return_img (bool, optional): bool to control if to return np array or VisImage
+
+    Returns:
+        image blended with perspective fields.
+    """
+    visualizer = VisualizerPerspective(img_rgb.copy())
+    vis_output = visualizer.draw_lati(latimap)
+    if torch.is_tensor(up):
+        up = up.numpy().transpose(1, 2, 0)
+    im_h, im_w, _ = img_rgb.shape
+    x, y = np.meshgrid(
+        np.arange(0, im_w, im_w // density), np.arange(0, im_h, im_h // density)
+    )
+    x, y = x.ravel(), y.ravel()
+    arrow_len = np.sqrt(im_w**2 + im_h**2) // arrow_inv_len
+    end = up[y, x, :] * arrow_len
+    if color is None:
+        color = (0, 1, 0)
+    vis_output = visualizer.draw_arrow(x, y, end[:, 0], -end[:, 1], color=color)
+    if return_img:
+        return vis_output.get_image()
+    else:
+        return vis_output
diff --git a/flux_style_shaper_api/perspective_fields/visualizer.py b/flux_style_shaper_api/perspective_fields/visualizer.py
new file mode 100644
index 00000000..f6f9d4bf
--- /dev/null
+++ b/flux_style_shaper_api/perspective_fields/visualizer.py
@@ -0,0 +1,283 @@
+#######################################################################
+# Adapted code from https://github.com/jinlinyi/PerspectiveFields
+#######################################################################
+# Modified from https://github.com/facebookresearch/detectron2/blob/main/detectron2/utils/visualizer.py
+import matplotlib.colors as mplc
+import matplotlib.figure as mplfigure
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from matplotlib.backends.backend_agg import FigureCanvasAgg
+
+
+class VisImage:
+    def __init__(self, img, scale=1.0):
+        """
+        Args:
+            img (ndarray): an RGB image of shape (H, W, 3) in range [0, 255].
+            scale (float): scale the input image
+        """
+        self.img = img
+        self.scale = scale
+        self.width, self.height = img.shape[1], img.shape[0]
+        self._setup_figure(img)
+
+    def _setup_figure(self, img):
+        """
+        Args:
+            Same as in :meth:`__init__()`.
+
+        Returns:
+            fig (matplotlib.pyplot.figure): top level container for all the image plot elements.
+            ax (matplotlib.pyplot.Axes): contains figure elements and sets the coordinate system.
+        """
+        fig = mplfigure.Figure(frameon=False)
+        self.dpi = fig.get_dpi()
+        # add a small 1e-2 to avoid precision lost due to matplotlib's truncation
+        # (https://github.com/matplotlib/matplotlib/issues/15363)
+        fig.set_size_inches(
+            (self.width * self.scale + 1e-2) / self.dpi,
+            (self.height * self.scale + 1e-2) / self.dpi,
+        )
+        self.canvas = FigureCanvasAgg(fig)
+        # self.canvas = mpl.backends.backend_cairo.FigureCanvasCairo(fig)
+        ax = fig.add_axes([0.0, 0.0, 1.0, 1.0])
+        ax.axis("off")
+        self.fig = fig
+        self.ax = ax
+        self.reset_image(img)
+
+    def reset_image(self, img):
+        """
+        Args:
+            img: same as in __init__
+        """
+        img = img.astype("uint8")
+        self.ax.imshow(
+            img, extent=(0, self.width, self.height, 0), interpolation="nearest"
+        )
+
+    def save(self, filepath):
+        """
+        Args:
+            filepath (str): a string that contains the absolute path, including the file name, where
+                the visualized image will be saved.
+        """
+        self.fig.savefig(filepath)
+
+    def get_image(self):
+        """
+        Returns:
+            ndarray:
+                the visualized image of shape (H, W, 3) (RGB) in uint8 type.
+                The shape is scaled w.r.t the input image using the given `scale` argument.
+        """
+        canvas = self.canvas
+        s, (width, height) = canvas.print_to_buffer()
+        # buf = io.BytesIO()  # works for cairo backend
+        # canvas.print_rgba(buf)
+        # width, height = self.width, self.height
+        # s = buf.getvalue()
+
+        buffer = np.frombuffer(s, dtype="uint8")
+
+        img_rgba = buffer.reshape(height, width, 4)
+        rgb, alpha = np.split(img_rgba, [3], axis=2)
+        return rgb.astype("uint8")
+
+
+class Visualizer:
+    """
+    Visualizer that draws data about detection/segmentation on images.
+
+    It contains methods like `draw_{text,box,circle,line,binary_mask,polygon}`
+    that draw primitive objects to images, as well as high-level wrappers like
+    `draw_{instance_predictions,sem_seg,panoptic_seg_predictions,dataset_dict}`
+    that draw composite data in some pre-defined style.
+
+    Note that the exact visualization style for the high-level wrappers are subject to change.
+    Style such as color, opacity, label contents, visibility of labels, or even the visibility
+    of objects themselves (e.g. when the object is too small) may change according
+    to different heuristics, as long as the results still look visually reasonable.
+
+    To obtain a consistent style, you can implement custom drawing functions with the
+    abovementioned primitive methods instead. If you need more customized visualization
+    styles, you can process the data yourself following their format documented in
+    tutorials (:doc:`/tutorials/models`, :doc:`/tutorials/datasets`). This class does not
+    intend to satisfy everyone's preference on drawing styles.
+
+    This visualizer focuses on high rendering quality rather than performance. It is not
+    designed to be used for real-time applications.
+    """
+
+    def __init__(self, img_rgb, scale=1.0, font_size_scale=1.0):
+        """
+        Args:
+            img_rgb: a numpy array of shape (H, W, C), where H and W correspond to
+                the height and width of the image respectively. C is the number of
+                color channels. The image is required to be in RGB format since that
+                is a requirement of the Matplotlib library. The image is also expected
+                to be in the range [0, 255].
+            font_size_scale: extra scaling of font size on top of default font size
+        """
+        self.img = np.asarray(img_rgb).clip(0, 255).astype(np.uint8)
+        self.output = VisImage(self.img, scale=scale)
+        self.cpu_device = torch.device("cpu")
+
+        # too small texts are useless, therefore clamp to 9
+        self._default_font_size = (
+            max(np.sqrt(self.output.height * self.output.width) // 90, 10 // scale)
+            * font_size_scale
+        )
+
+    """
+    Primitive drawing functions:
+    """
+
+    def draw_text(
+        self,
+        text,
+        position,
+        *,
+        font_size=None,
+        color="g",
+        horizontal_alignment="center",
+        rotation=0,
+    ):
+        """
+        Args:
+            text (str): class label
+            position (tuple): a tuple of the x and y coordinates to place text on image.
+            font_size (int, optional): font of the text. If not provided, a font size
+                proportional to the image width is calculated and used.
+            color: color of the text. Refer to `matplotlib.colors` for full list
+                of formats that are accepted.
+            horizontal_alignment (str): see `matplotlib.text.Text`
+            rotation: rotation angle in degrees CCW
+
+        Returns:
+            output (VisImage): image object with text drawn.
+        """
+        if not font_size:
+            font_size = self._default_font_size
+
+        # since the text background is dark, we don't want the text to be dark
+        color = np.maximum(list(mplc.to_rgb(color)), 0.2)
+        color[np.argmax(color)] = max(0.8, np.max(color))
+
+        x, y = position
+        self.output.ax.text(
+            x,
+            y,
+            text,
+            size=font_size * self.output.scale,
+            family="sans-serif",
+            bbox={"facecolor": "black", "alpha": 0.8, "pad": 0.7, "edgecolor": "none"},
+            verticalalignment="top",
+            horizontalalignment=horizontal_alignment,
+            color=color,
+            zorder=10,
+            rotation=rotation,
+        )
+        return self.output
+
+    def get_output(self):
+        """
+        Returns:
+            output (VisImage): the image output containing the visualizations added
+            to the image.
+        """
+        return self.output
+
+
+class VisualizerPerspective(Visualizer):
+    def draw_arrow(
+        self,
+        x_pos,
+        y_pos,
+        x_direct,
+        y_direct,
+        color=None,
+        linestyle="-",
+        linewidth=None,
+    ):
+        """
+        Args:
+            x_data (list[int]): a list containing x values of all the points being drawn.
+                Length of list should match the length of y_data.
+            y_data (list[int]): a list containing y values of all the points being drawn.
+                Length of list should match the length of x_data.
+            color: color of the line. Refer to `matplotlib.colors` for a full list of
+                formats that are accepted.
+            linestyle: style of the line. Refer to `matplotlib.lines.Line2D`
+                for a full list of formats that are accepted.
+            linewidth (float or None): width of the line. When it's None,
+                a default value will be computed and used.
+
+        Returns:
+            output (VisImage): image object with line drawn.
+        """
+        if linewidth is None:
+            linewidth = self._default_font_size / 3
+        linewidth = max(linewidth, 1)
+        self.output.ax.quiver(
+            x_pos,
+            y_pos,
+            x_direct,
+            y_direct,
+            color=color,
+            scale_units="xy",
+            scale=1,
+            antialiased=True,
+            headaxislength=3.5,
+            linewidths=0.1,  # , width=0.01
+        )
+        return self.output
+
+    def draw_lati(
+        self, latimap, alpha_contourf=0.4, alpha_contour=0.9, contour_only=False
+    ):
+        """Blend latitude map"""
+        height, width = latimap.shape
+        y, x = np.mgrid[0:height, 0:width]
+        cmap = plt.get_cmap("seismic")
+        bands = 20
+        levels = np.linspace(-np.pi / 2, np.pi / 2, bands - 1)
+        if not contour_only:
+            pp = self.output.ax.contourf(
+                x,
+                y,
+                latimap,
+                levels=levels,
+                cmap=cmap,
+                alpha=alpha_contourf,
+                antialiased=True,
+            )
+            pp2 = self.output.ax.contour(
+                x,
+                y,
+                latimap,
+                pp.levels,
+                cmap=cmap,
+                alpha=alpha_contour,
+                antialiased=True,
+                linewidths=5,
+            )
+            try:
+                for c in pp2.collections:
+                    c.set_linestyle("solid")
+            except AttributeError:
+                pass
+        else:
+            # only plot central contour
+            pp = self.output.ax.contour(
+                x,
+                y,
+                latimap,
+                levels=[0],
+                cmap=cmap,
+                alpha=alpha_contour,
+                antialiased=True,
+                linewidths=15,
+            )
+        return self.output