local_viewer.py

# 
# Toyota Motor Europe NV/SA and its affiliated companies retain all intellectual 
# property and proprietary rights in and to this software and related documentation. 
# Any commercial use, reproduction, disclosure or distribution of this software and 
# related documentation without an express license agreement from Toyota Motor Europe NV/SA 
# is strictly prohibited.
#

import json
import math
import tyro
from dataclasses import dataclass, field
from typing import Literal, Optional
from pathlib import Path
import time
import dearpygui.dearpygui as dpg
import numpy as np
import torch
from PIL import Image
from scipy.spatial.transform import Rotation as R
from scipy.interpolate import interp1d
import matplotlib

from utils.viewer_utils import Mini3DViewer, Mini3DViewerConfig
from gaussian_renderer import GaussianModel, FlameGaussianModel
from gaussian_renderer import render
from mesh_renderer import NVDiffRenderer


@dataclass
class PipelineConfig:
    debug: bool = False
    compute_cov3D_python: bool = False
    convert_SHs_python: bool = False


@dataclass
class Config(Mini3DViewerConfig):
    pipeline: PipelineConfig = field(default_factory=PipelineConfig)
    """Pipeline settings for gaussian splatting rendering"""
    cam_convention: Literal["opengl", "opencv"] = "opencv"
    """Camera convention"""
    point_path: Optional[Path] = None
    """Path to the gaussian splatting file"""
    motion_path: Optional[Path] = None
    """Path to the motion file (npz)"""
    sh_degree: int = 3
    """Spherical Harmonics degree"""
    background_color: tuple[float] = (1., 1., 1.)
    """default GUI background color"""
    save_folder: Path = Path("./viewer_output")
    """default saving folder"""
    fps: int = 25
    """default fps for recording"""
    keyframe_interval: int = 1
    """default keyframe interval"""
    ref_json: Optional[Path] = None
    """ Path to a reference json file. We copy file paths from a reference json into 
    the exported trajectory json file as placeholders so that `render.py` can directly
    load it like a normal sequence. """
    demo_mode: bool = False
    """The UI will be simplified in demo mode."""

class LocalViewer(Mini3DViewer):
    def __init__(self, cfg: Config):
        self.cfg = cfg
        
        # recording settings
        self.keyframes = []  # list of state dicts of keyframes
        self.all_frames = {}  # state dicts of all frames {key: [num_frames, ...]}
        self.num_record_timeline = 0
        self.playing = False

        print("Initializing 3D Gaussians...")
        self.init_gaussians()

        if self.gaussians.binding is not None:
            # rendering settings
            self.mesh_color = torch.tensor([1, 1, 1, 0.5])
            self.face_colors = None
            print("Initializing mesh renderer...")
            self.mesh_renderer = NVDiffRenderer(use_opengl=False)
        
        # FLAME parameters
        if self.gaussians.binding is not None:
            print("Initializing FLAME parameters...")
            self.reset_flame_param()
        
        super().__init__(cfg, 'GaussianAvatars - Local Viewer')

        if self.gaussians.binding is not None:
            self.num_timesteps = self.gaussians.num_timesteps
            dpg.configure_item("_slider_timestep", max_value=self.num_timesteps - 1)

            self.gaussians.select_mesh_by_timestep(self.timestep)

    def init_gaussians(self):
        # load gaussians
        if (Path(self.cfg.point_path).parent / "flame_param.npz").exists():
            self.gaussians = FlameGaussianModel(self.cfg.sh_degree)
        else:
            self.gaussians = GaussianModel(self.cfg.sh_degree)

        # selected_fid = self.gaussians.flame_model.mask.get_fid_by_region(['left_half'])
        # selected_fid = self.gaussians.flame_model.mask.get_fid_by_region(['right_half'])
        # unselected_fid = self.gaussians.flame_model.mask.get_fid_except_fids(selected_fid)
        unselected_fid = []
        
        if self.cfg.point_path is not None:
            if self.cfg.point_path.exists():
                self.gaussians.load_ply(self.cfg.point_path, has_target=False, motion_path=self.cfg.motion_path, disable_fid=unselected_fid)
            else:
                raise FileNotFoundError(f'{self.cfg.point_path} does not exist.')

    def refresh_stat(self):
        if self.last_time_fresh is not None:
            elapsed = time.time() - self.last_time_fresh
            fps = 1 / elapsed
            dpg.set_value("_log_fps", f'{int(fps):<4d}')
        self.last_time_fresh = time.time()
    
    def update_record_timeline(self):
        cycles = dpg.get_value("_input_cycles")
        if cycles == 0:
            self.num_record_timeline = sum([keyframe['interval'] for keyframe in self.keyframes[:-1]])
        else:
            self.num_record_timeline = sum([keyframe['interval'] for keyframe in self.keyframes]) * cycles

        dpg.configure_item("_slider_record_timestep", min_value=0, max_value=self.num_record_timeline-1)

        if len(self.keyframes) <= 0:
            self.all_frames = {}
            return
        else:
            k_x = []

            keyframes = self.keyframes.copy()
            if cycles > 0:
                # pad a cycle at the beginning and the end to ensure smooth transition
                keyframes = self.keyframes * (cycles + 2)
                t_couter = -sum([keyframe['interval'] for keyframe in self.keyframes])
            else:
                t_couter = 0

            for keyframe in keyframes:
                k_x.append(t_couter)
                t_couter += keyframe['interval']
            
            x = np.arange(self.num_record_timeline)
            self.all_frames = {}

            if len(keyframes) <= 1:
                for k in keyframes[0]:
                    k_y = np.concatenate([np.array(keyframe[k])[None] for keyframe in keyframes], axis=0)
                    self.all_frames[k] = np.tile(k_y, (self.num_record_timeline, 1))
            else:
                kind = 'linear' if len(keyframes) <= 3 else 'cubic'
            
                for k in keyframes[0]:
                    if k == 'interval':
                        continue
                    k_y = np.concatenate([np.array(keyframe[k])[None] for keyframe in keyframes], axis=0)
                  
                    interp_funcs = [interp1d(k_x, k_y[:, i], kind=kind, fill_value='extrapolate') for i in range(k_y.shape[1])]

                    y = np.array([interp_func(x) for interp_func in interp_funcs]).transpose(1, 0)
                    self.all_frames[k] = y

    def get_state_dict(self):
        return {
            'rot': self.cam.rot.as_quat(),
            'look_at': np.array(self.cam.look_at),
            'radius': np.array([self.cam.radius]).astype(np.float32),
            'fovy': np.array([self.cam.fovy]).astype(np.float32),
            'interval': self.cfg.fps*self.cfg.keyframe_interval,
        }

    def get_state_dict_record(self):
        record_timestep = dpg.get_value("_slider_record_timestep")
        state_dict = {k: self.all_frames[k][record_timestep] for k in self.all_frames}
        return state_dict

    def apply_state_dict(self, state_dict):
        if 'rot' in state_dict:
            self.cam.rot = R.from_quat(state_dict['rot'])
        if 'look_at' in state_dict:
            self.cam.look_at = state_dict['look_at']
        if 'radius' in state_dict:
            self.cam.radius = state_dict['radius'].item()
        if 'fovy' in state_dict:
            self.cam.fovy = state_dict['fovy'].item()
    
    def parse_ref_json(self):
        if self.cfg.ref_json is None:
            return {}
        else:
            with open(self.cfg.ref_json, 'r') as f:
                ref_dict = json.load(f)

        tid2paths = {}
        for frame in ref_dict['frames']:
            tid = frame['timestep_index']
            if tid not in tid2paths:
                tid2paths[tid] = frame
        return tid2paths
    
    def export_trajectory(self):
        tid2paths = self.parse_ref_json()

        if self.num_record_timeline <= 0:
            return
        
        timestamp = f"{time.strftime('%Y-%m-%d_%H-%M-%S')}"
        traj_dict = {'frames': []}
        timestep_indices = []
        camera_indices = []
        for i in range(self.num_record_timeline):
            # update
            dpg.set_value("_slider_record_timestep", i)
            state_dict = self.get_state_dict_record()
            self.apply_state_dict(state_dict)

            self.need_update = True
            while self.need_update:
                time.sleep(0.001)

            # save image
            save_folder = self.cfg.save_folder / timestamp
            if not save_folder.exists():
                save_folder.mkdir(parents=True)
            path = save_folder / f"{i:05d}.png"
            print(f"Saving image to {path}")
            Image.fromarray((np.clip(self.render_buffer, 0, 1) * 255).astype(np.uint8)).save(path)

            # cache camera parameters
            cx = self.cam.intrinsics[2]
            cy = self.cam.intrinsics[3]
            fl_x = self.cam.intrinsics[0].item() if isinstance(self.cam.intrinsics[0], np.ndarray) else self.cam.intrinsics[0]
            fl_y = self.cam.intrinsics[1].item() if isinstance(self.cam.intrinsics[1], np.ndarray) else self.cam.intrinsics[1]
            h = self.cam.image_height
            w = self.cam.image_width
            angle_x = math.atan(w / (fl_x * 2)) * 2
            angle_y = math.atan(h / (fl_y * 2)) * 2

            c2w = self.cam.pose.copy()  # opencv convention
            c2w[:, [1, 2]] *= -1  # opencv to opengl
            # transform_matrix = np.linalg.inv(c2w).tolist()  # world2cam
            
            timestep_index = self.timestep
            camera_indx = i
            timestep_indices.append(timestep_index)
            camera_indices.append(camera_indx)
            
            tid2paths[timestep_index]['file_path']

            frame = {
                "cx": cx,
                "cy": cy,
                "fl_x": fl_x,
                "fl_y": fl_y,
                "h": h,
                "w": w,
                "camera_angle_x": angle_x,
                "camera_angle_y": angle_y,
                "transform_matrix": c2w.tolist(),
                'timestep_index': timestep_index,
                'camera_indx': camera_indx,
            }
            if timestep_index in tid2paths:
                frame['file_path'] = tid2paths[timestep_index]['file_path']
                frame['fg_mask_path'] = tid2paths[timestep_index]['fg_mask_path']
                frame['flame_param_path'] = tid2paths[timestep_index]['flame_param_path']
            traj_dict['frames'].append(frame)

            # update timestep
            if dpg.get_value("_checkbox_dynamic_record"):
                self.timestep = min(self.timestep + 1, self.num_timesteps - 1)
                dpg.set_value("_slider_timestep", self.timestep)
                self.gaussians.select_mesh_by_timestep(self.timestep)
        
        traj_dict['timestep_indices'] = sorted(list(set(timestep_indices)))
        traj_dict['camera_indices'] = sorted(list(set(camera_indices)))
        
        # save camera parameters
        path = save_folder / f"trajectory.json"
        print(f"Saving trajectory to {path}")
        with open(path, 'w') as f:
            json.dump(traj_dict, f, indent=4)

    def reset_flame_param(self):
        self.flame_param = {
            'expr': torch.zeros(1, self.gaussians.n_expr),
            'rotation': torch.zeros(1, 3),
            'neck': torch.zeros(1, 3),
            'jaw': torch.zeros(1, 3),
            'eyes': torch.zeros(1, 6),
            'translation': torch.zeros(1, 3),
        }

    def define_gui(self):
        super().define_gui()

        # window: rendering options ==================================================================================================
        with dpg.window(label="Render", tag="_render_window", autosize=True):

            with dpg.group(horizontal=True):
                dpg.add_text("FPS:", show=not self.cfg.demo_mode)
                dpg.add_text("0   ", tag="_log_fps", show=not self.cfg.demo_mode)

            dpg.add_text(f"number of points: {self.gaussians._xyz.shape[0]}")
            
            with dpg.group(horizontal=True):
                # show splatting
                def callback_show_splatting(sender, app_data):
                    self.need_update = True
                dpg.add_checkbox(label="show splatting", default_value=True, callback=callback_show_splatting, tag="_checkbox_show_splatting")

                dpg.add_spacer(width=10)

                if self.gaussians.binding is not None:
                    # show mesh
                    def callback_show_mesh(sender, app_data):
                        self.need_update = True
                    dpg.add_checkbox(label="show mesh", default_value=False, callback=callback_show_mesh, tag="_checkbox_show_mesh")

                    # # show original mesh
                    # def callback_original_mesh(sender, app_data):
                    #     self.original_mesh = app_data
                    #     self.need_update = True
                    # dpg.add_checkbox(label="original mesh", default_value=self.original_mesh, callback=callback_original_mesh)
            
            # timestep slider and buttons
            if self.num_timesteps != None:
                def callback_set_current_frame(sender, app_data):
                    if sender == "_slider_timestep":
                        self.timestep = app_data
                    elif sender in ["_button_timestep_plus", "_mvKey_Right"]:
                        self.timestep = min(self.timestep + 1, self.num_timesteps - 1)
                    elif sender in ["_button_timestep_minus", "_mvKey_Left"]:
                        self.timestep = max(self.timestep - 1, 0)
                    elif sender == "_mvKey_Home":
                        self.timestep = 0
                    elif sender == "_mvKey_End":
                        self.timestep = self.num_timesteps - 1

                    dpg.set_value("_slider_timestep", self.timestep)
                    self.gaussians.select_mesh_by_timestep(self.timestep)

                    self.need_update = True
                with dpg.group(horizontal=True):
                    dpg.add_button(label='-', tag="_button_timestep_minus", callback=callback_set_current_frame)
                    dpg.add_button(label='+', tag="_button_timestep_plus", callback=callback_set_current_frame)
                    dpg.add_slider_int(label="timestep", tag='_slider_timestep', width=153, min_value=0, max_value=self.num_timesteps - 1, format="%d", default_value=0, callback=callback_set_current_frame)

            # # render_mode combo
            # def callback_change_mode(sender, app_data):
            #     self.render_mode = app_data
            #     self.need_update = True
            # dpg.add_combo(('rgb', 'depth', 'opacity'), label='render mode', default_value=self.render_mode, callback=callback_change_mode)

            # scaling_modifier slider
            def callback_set_scaling_modifier(sender, app_data):
                self.need_update = True
            dpg.add_slider_float(label="Scale modifier", min_value=0, max_value=1, format="%.2f", width=200, default_value=1, callback=callback_set_scaling_modifier, tag="_slider_scaling_modifier")

            # fov slider
            def callback_set_fovy(sender, app_data):
                self.cam.fovy = app_data
                self.need_update = True
            dpg.add_slider_int(label="FoV (vertical)", min_value=1, max_value=120, width=200, format="%d deg", default_value=self.cam.fovy, callback=callback_set_fovy, tag="_slider_fovy", show=not self.cfg.demo_mode)

            if self.gaussians.binding is not None:
                # visualization options
                def callback_visual_options(sender, app_data):
                    if app_data == 'number of points per face':
                        value, ct = self.gaussians.binding.unique(return_counts=True)
                        ct = torch.log10(ct + 1)
                        ct = ct.float() / ct.max()
                        cmap = matplotlib.colormaps["plasma"]
                        self.face_colors = torch.from_numpy(cmap(ct.cpu())[None, :, :3]).to(self.gaussians.verts)
                    else:
                        self.face_colors = self.mesh_color[:3].to(self.gaussians.verts)[None, None, :].repeat(1, self.gaussians.face_center.shape[0], 1)  # (1, F, 3)
                    
                    dpg.set_value('_checkbox_show_mesh', True)
                    self.need_update = True
                dpg.add_combo(["none", "number of points per face"], default_value="none", label='visualization', width=200, callback=callback_visual_options, tag="_visual_options")

                # mesh_color picker
                def callback_change_mesh_color(sender, app_data):
                    self.mesh_color = torch.tensor(app_data, dtype=torch.float32)  # only need RGB in [0, 1]
                    if dpg.get_value("_visual_options") == 'none':
                        self.face_colors = self.mesh_color[:3].to(self.gaussians.verts)[None, None, :].repeat(1, self.gaussians.face_center.shape[0], 1)
                    self.need_update = True
                dpg.add_color_edit((self.mesh_color*255).tolist(), label="Mesh Color", width=200, callback=callback_change_mesh_color, show=not self.cfg.demo_mode)

            # # bg_color picker
            # def callback_change_bg(sender, app_data):
            #     self.bg_color = torch.tensor(app_data[:3], dtype=torch.float32)  # only need RGB in [0, 1]
            #     self.need_update = True
            # dpg.add_color_edit((self.bg_color*255).tolist(), label="Background Color", width=200, no_alpha=True, callback=callback_change_bg)

            # # near slider
            # def callback_set_near(sender, app_data):
            #     self.cam.znear = app_data
            #     self.need_update = True
            # dpg.add_slider_int(label="near", min_value=1e-8, max_value=2, format="%.2f", default_value=self.cam.znear, callback=callback_set_near, tag="_slider_near")

            # # far slider
            # def callback_set_far(sender, app_data):
            #     self.cam.zfar = app_data
            #     self.need_update = True
            # dpg.add_slider_int(label="far", min_value=1e-3, max_value=10, format="%.2f", default_value=self.cam.zfar, callback=callback_set_far, tag="_slider_far")
            
            # camera
            with dpg.group(horizontal=True):
                def callback_reset_camera(sender, app_data):
                    self.cam.reset()
                    self.need_update = True
                    dpg.set_value("_slider_fovy", self.cam.fovy)
                dpg.add_button(label="reset camera", tag="_button_reset_pose", callback=callback_reset_camera, show=not self.cfg.demo_mode)
                
                def callback_cache_camera(sender, app_data):
                    self.cam.save()
                dpg.add_button(label="cache camera", tag="_button_cache_pose", callback=callback_cache_camera, show=not self.cfg.demo_mode)

                def callback_clear_cache(sender, app_data):
                    self.cam.clear()
                dpg.add_button(label="clear cache", tag="_button_clear_cache", callback=callback_clear_cache, show=not self.cfg.demo_mode)
                
        # window: recording ==================================================================================================
        with dpg.window(label="Record", tag="_record_window", autosize=True, pos=(0, self.H//2)):
            dpg.add_text("Keyframes")
            with dpg.group(horizontal=True):
                # list keyframes
                def callback_set_current_keyframe(sender, app_data):
                    idx = int(dpg.get_value("_listbox_keyframes"))
                    self.apply_state_dict(self.keyframes[idx])

                    record_timestep = sum([keyframe['interval'] for keyframe in self.keyframes[:idx]])
                    dpg.set_value("_slider_record_timestep", record_timestep)

                    self.need_update = True
                dpg.add_listbox(self.keyframes, width=200, tag="_listbox_keyframes", callback=callback_set_current_keyframe)

                # edit keyframes
                with dpg.group():
                    # add
                    def callback_add_keyframe(sender, app_data):
                        if len(self.keyframes) == 0:
                            new_idx = 0
                        else:
                            new_idx = int(dpg.get_value("_listbox_keyframes")) + 1

                        states = self.get_state_dict()
                        
                        self.keyframes.insert(new_idx, states)
                        dpg.configure_item("_listbox_keyframes", items=list(range(len(self.keyframes))))
                        dpg.set_value("_listbox_keyframes", new_idx)

                        self.update_record_timeline()
                    dpg.add_button(label="add", tag="_button_add_keyframe", callback=callback_add_keyframe)

                    # delete
                    def callback_delete_keyframe(sender, app_data):
                        idx = int(dpg.get_value("_listbox_keyframes"))
                        self.keyframes.pop(idx)
                        dpg.configure_item("_listbox_keyframes", items=list(range(len(self.keyframes))))
                        dpg.set_value("_listbox_keyframes", idx-1)

                        self.update_record_timeline()
                    dpg.add_button(label="delete", tag="_button_delete_keyframe", callback=callback_delete_keyframe)

                    # update
                    def callback_update_keyframe(sender, app_data):
                        if len(self.keyframes) == 0:
                            return
                        else:
                            idx = int(dpg.get_value("_listbox_keyframes"))

                        states = self.get_state_dict()
                        states['interval'] = self.cfg.fps*self.cfg.keyframe_interval

                        self.keyframes[idx] = states
                    dpg.add_button(label="update", tag="_button_update_keyframe", callback=callback_update_keyframe)

            with dpg.group(horizontal=True):
                def callback_set_record_cycles(sender, app_data):
                    self.update_record_timeline()
                dpg.add_input_int(label="cycles", tag="_input_cycles", default_value=0, width=70, callback=callback_set_record_cycles)

                def callback_set_keyframe_interval(sender, app_data):
                    self.cfg.keyframe_interval = app_data
                    for keyframe in self.keyframes:
                        keyframe['interval'] = self.cfg.fps*self.cfg.keyframe_interval
                    self.update_record_timeline()
                dpg.add_input_int(label="interval", tag="_input_interval", default_value=self.cfg.keyframe_interval, width=70, callback=callback_set_keyframe_interval)
            
            def callback_set_record_timestep(sender, app_data):
                state_dict = self.get_state_dict_record()
                
                self.apply_state_dict(state_dict)
                self.need_update = True
            dpg.add_slider_int(label="timeline", tag='_slider_record_timestep', width=200, min_value=0, max_value=0, format="%d", default_value=0, callback=callback_set_record_timestep)
            
            with dpg.group(horizontal=True):
                dpg.add_checkbox(label="dynamic", default_value=False, tag="_checkbox_dynamic_record")
                dpg.add_checkbox(label="loop", default_value=True, tag="_checkbox_loop_record")
            
            with dpg.group(horizontal=True):
                def callback_play(sender, app_data):
                    self.playing = not self.playing
                    self.need_update = True
                dpg.add_button(label="play", tag="_button_play", callback=callback_play)

                def callback_export_trajectory(sender, app_data):
                    self.export_trajectory()
                dpg.add_button(label="export traj", tag="_button_export_traj", callback=callback_export_trajectory)
            
            def callback_save_image(sender, app_data):
                if not self.cfg.save_folder.exists():
                    self.cfg.save_folder.mkdir(parents=True)
                path = self.cfg.save_folder / f"{time.strftime('%Y-%m-%d_%H-%M-%S')}_{self.timestep}.png"
                print(f"Saving image to {path}")
                Image.fromarray((np.clip(self.render_buffer, 0, 1) * 255).astype(np.uint8)).save(path)
            with dpg.group(horizontal=True):
                dpg.add_button(label="save image", tag="_button_save_image", callback=callback_save_image)

        # window: FLAME ==================================================================================================
        if self.gaussians.binding is not None:
            with dpg.window(label="FLAME parameters", tag="_flame_window", autosize=True, pos=(self.W-300, 0)):
                def callback_enable_control(sender, app_data):
                    if app_data:
                        self.gaussians.update_mesh_by_param_dict(self.flame_param)
                    else:
                        self.gaussians.select_mesh_by_timestep(self.timestep)
                    self.need_update = True
                dpg.add_checkbox(label="enable control", default_value=False, tag="_checkbox_enable_control", callback=callback_enable_control)

                dpg.add_separator()

                def callback_set_pose(sender, app_data):
                    joint, axis = sender.split('-')[1:3]
                    axis_idx = {'x': 0, 'y': 1, 'z': 2}[axis]
                    self.flame_param[joint][0, axis_idx] = app_data
                    if joint == 'eyes':
                        self.flame_param[joint][0, 3+axis_idx] = app_data
                    if not dpg.get_value("_checkbox_enable_control"):
                        dpg.set_value("_checkbox_enable_control", True)
                    self.gaussians.update_mesh_by_param_dict(self.flame_param)
                    self.need_update = True
                dpg.add_text(f'Joints')
                self.pose_sliders = []
                max_rot = 0.5
                for joint in ['neck', 'jaw', 'eyes']:
                    if joint in self.flame_param:
                        with dpg.group(horizontal=True):
                            dpg.add_slider_float(min_value=-max_rot, max_value=max_rot, format="%.2f", default_value=self.flame_param[joint][0, 0], callback=callback_set_pose, tag=f"_slider-{joint}-x", width=70)
                            dpg.add_slider_float(min_value=-max_rot, max_value=max_rot, format="%.2f", default_value=self.flame_param[joint][0, 1], callback=callback_set_pose, tag=f"_slider-{joint}-y", width=70)
                            dpg.add_slider_float(min_value=-max_rot, max_value=max_rot, format="%.2f", default_value=self.flame_param[joint][0, 2], callback=callback_set_pose, tag=f"_slider-{joint}-z", width=70)
                            self.pose_sliders.append(f"_slider-{joint}-x")
                            self.pose_sliders.append(f"_slider-{joint}-y")
                            self.pose_sliders.append(f"_slider-{joint}-z")
                            dpg.add_text(f'{joint:4s}')
                dpg.add_text('   roll       pitch      yaw')
                
                dpg.add_separator()
                
                def callback_set_expr(sender, app_data):
                    expr_i = int(sender.split('-')[2])
                    self.flame_param['expr'][0, expr_i] = app_data
                    if not dpg.get_value("_checkbox_enable_control"):
                        dpg.set_value("_checkbox_enable_control", True)
                    self.gaussians.update_mesh_by_param_dict(self.flame_param)
                    self.need_update = True
                self.expr_sliders = []
                dpg.add_text(f'Expressions')
                for i in range(5):
                    dpg.add_slider_float(label=f"{i}", min_value=-3, max_value=3, format="%.2f", default_value=0, callback=callback_set_expr, tag=f"_slider-expr-{i}", width=250)
                    self.expr_sliders.append(f"_slider-expr-{i}")

                def callback_reset_flame(sender, app_data):
                    self.reset_flame_param()
                    if not dpg.get_value("_checkbox_enable_control"):
                        dpg.set_value("_checkbox_enable_control", True)
                    self.gaussians.update_mesh_by_param_dict(self.flame_param)
                    self.need_update = True
                    for slider in self.pose_sliders + self.expr_sliders:
                        dpg.set_value(slider, 0)
                dpg.add_button(label="reset FLAME", tag="_button_reset_flame", callback=callback_reset_flame)

        # widget-dependent handlers ========================================================================================
        with dpg.handler_registry():
            dpg.add_key_press_handler(dpg.mvKey_Left, callback=callback_set_current_frame, tag='_mvKey_Left')
            dpg.add_key_press_handler(dpg.mvKey_Right, callback=callback_set_current_frame, tag='_mvKey_Right')
            dpg.add_key_press_handler(dpg.mvKey_Home, callback=callback_set_current_frame, tag='_mvKey_Home')
            dpg.add_key_press_handler(dpg.mvKey_End, callback=callback_set_current_frame, tag='_mvKey_End')

            def callbackmouse_wheel_slider(sender, app_data):
                delta = app_data
                if dpg.is_item_hovered("_slider_timestep"):
                    self.timestep = min(max(self.timestep - delta, 0), self.num_timesteps - 1)
                    dpg.set_value("_slider_timestep", self.timestep)
                    self.gaussians.select_mesh_by_timestep(self.timestep)
                    self.need_update = True
            dpg.add_mouse_wheel_handler(callback=callbackmouse_wheel_slider)

    def prepare_camera(self):
        @dataclass
        class Cam:
            FoVx = float(np.radians(self.cam.fovx))
            FoVy = float(np.radians(self.cam.fovy))
            image_height = self.cam.image_height
            image_width = self.cam.image_width
            world_view_transform = torch.tensor(self.cam.world_view_transform).float().cuda().T  # the transpose is required by gaussian splatting rasterizer
            full_proj_transform = torch.tensor(self.cam.full_proj_transform).float().cuda().T  # the transpose is required by gaussian splatting rasterizer
            camera_center = torch.tensor(self.cam.pose[:3, 3]).cuda()
        return Cam

    @torch.no_grad()
    def run(self):
        print("Running LocalViewer...")

        while dpg.is_dearpygui_running():

            if self.need_update or self.playing:
                cam = self.prepare_camera()

                if dpg.get_value("_checkbox_show_splatting"):
                    # rgb
                    rgb_splatting = render(cam, self.gaussians, self.cfg.pipeline, torch.tensor(self.cfg.background_color).cuda(), scaling_modifier=dpg.get_value("_slider_scaling_modifier"))["render"].permute(1, 2, 0).contiguous()

                    # opacity
                    # override_color = torch.ones_like(self.gaussians._xyz).cuda()
                    # background_color = torch.tensor(self.cfg.background_color).cuda() * 0
                    # rgb_splatting = render(cam, self.gaussians, self.cfg.pipeline, background_color, scaling_modifier=dpg.get_value("_slider_scaling_modifier"), override_color=override_color)["render"].permute(1, 2, 0).contiguous()

                if self.gaussians.binding is not None and dpg.get_value("_checkbox_show_mesh"):
                    out_dict = self.mesh_renderer.render_from_camera(self.gaussians.verts, self.gaussians.faces, cam, face_colors=self.face_colors)

                    rgba_mesh = out_dict['rgba'].squeeze(0)  # (H, W, C)
                    rgb_mesh = rgba_mesh[:, :, :3]
                    alpha_mesh = rgba_mesh[:, :, 3:]
                    mesh_opacity = self.mesh_color[3:].cuda()

                if dpg.get_value("_checkbox_show_splatting") and dpg.get_value("_checkbox_show_mesh"):
                    rgb = rgb_mesh * alpha_mesh * mesh_opacity  + rgb_splatting * (alpha_mesh * (1 - mesh_opacity) + (1 - alpha_mesh))
                elif dpg.get_value("_checkbox_show_splatting") and not dpg.get_value("_checkbox_show_mesh"):
                    rgb = rgb_splatting
                elif not dpg.get_value("_checkbox_show_splatting") and dpg.get_value("_checkbox_show_mesh"):
                    rgb = rgb_mesh
                else:
                    rgb = torch.ones([self.H, self.W, 3])

                self.render_buffer = rgb.cpu().numpy()
                if self.render_buffer.shape[0] != self.H or self.render_buffer.shape[1] != self.W:
                    continue
                dpg.set_value("_texture", self.render_buffer)

                self.refresh_stat()
                self.need_update = False

                if self.playing:
                    record_timestep = dpg.get_value("_slider_record_timestep")
                    if record_timestep >= self.num_record_timeline - 1:
                        if not dpg.get_value("_checkbox_loop_record"):
                            self.playing = False
                        dpg.set_value("_slider_record_timestep", 0)
                    else:
                        dpg.set_value("_slider_record_timestep", record_timestep + 1)
                        if dpg.get_value("_checkbox_dynamic_record"):
                            self.timestep = min(self.timestep + 1, self.num_timesteps - 1)
                            dpg.set_value("_slider_timestep", self.timestep)
                            self.gaussians.select_mesh_by_timestep(self.timestep)

                        state_dict = self.get_state_dict_record()
                        self.apply_state_dict(state_dict)

            dpg.render_dearpygui_frame()


if __name__ == "__main__":
    cfg = tyro.cli(Config)
    gui = LocalViewer(cfg)
    gui.run()