-
Notifications
You must be signed in to change notification settings - Fork 15
/
app.py
384 lines (302 loc) · 16.8 KB
/
app.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
import gradio as gr
from addict import Dict
import os
import torch
from system_gm_ldm import GMLDMSystem
from system_traj_dit import TrajDiTSystem
import numpy as np
from utils import export_ply_for_gaussians, import_str, export_video
import torch.nn.functional as F
import argparse
from omegaconf import OmegaConf
import functools
import gradio as gr
import base64
import io
import matplotlib
import matplotlib.pyplot as plt
import plotly
from pytorch3d.renderer import PerspectiveCameras
from typing import NamedTuple
from pytorch3d.vis.plotly_vis import plot_scene
import tempfile
import shutil
from contextlib import contextmanager
from pytorch3d.renderer.cameras import FoVPerspectiveCameras, look_at_view_transform
import tqdm
def get_args_parser():
parser = argparse.ArgumentParser()
parser.add_argument("--config", default="configs/main.yaml")
parser.add_argument("--ckpt", default="model.ckpt")
parser.add_argument("--gpu", type=int, default=0)
return parser
def get_w2c(camera):
w2c = np.linalg.inv(camera)
return w2c
HTML_TEMPLATE = """<html><head><meta charset="utf-8"/></head>
<body><img src="data:image/png;charset=utf-8;base64, {image_encoded}"/>
{plotly_html}</body></html>"""
class AxisArgs(NamedTuple): # pragma: no cover
showgrid: bool = False
zeroline: bool = False
showline: bool = False
ticks: str = ""
showticklabels: bool = False
backgroundcolor: str = "#fff"
showaxeslabels: bool = False
@contextmanager
def temporary_directory():
tmp_dir = tempfile.mkdtemp()
try:
yield tmp_dir
finally:
shutil.rmtree(tmp_dir)
class App():
def __init__(self):
self.init_data()
self.parser = get_args_parser()
self.args, _ = self.parser.parse_known_args()
self.opt = OmegaConf.load(self.args.config)
self.device = f'cuda:{self.args.gpu}'
self.ckpt = self.args.ckpt
self.app = gr.Blocks(css=""".gradio-container {margin: 0 !important; min-width: 100%};""", title="Director3d Gradio Demo")
with self.app:
self.init_app()
def init_data(self) :
self.data = Dict(
)
def init_models(self):
params = torch.load(self.ckpt, map_location=self.device)
self.system_gm_ldm = GMLDMSystem(self.opt).to(self.device).eval()
self.system_traj_dit = TrajDiTSystem(self.opt).to(self.device).eval()
self.system_gm_ldm.model.load_state_dict(params['gm_ldm'], strict=False)
self.system_traj_dit.model.load_state_dict(params['traj_dit'], strict=False)
self.refiner = import_str(self.opt['inference']['refiner']['module'])(**self.opt['inference']['refiner']['args'], total_iterations=1000).to(self.device)
def init_app(self):
with temporary_directory() as tmp_dir:
data = gr.State(
Dict({
"cameras": [],
"results": [],
"text": "",
"tmp_dir":tmp_dir
})
)
with gr.Row():
helps = gr.Markdown(value=
"""
# Gradio Demo of "🎥 Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text"
<div>
<a style="display:inline-block" href="http://arxiv.org/abs/2404.11613"><img src='https://img.shields.io/badge/arXiv-Director3D-red?logo=arxiv' alt='Paper PDF'></a>
<a style="display:inline-block; margin-left: .5em" href='https://github.com/imlixinyang/director3d'><img src='https://img.shields.io/badge/Github-Director3D-blue?logo=github' alt='Project Page'></a>
<a style="display:inline-block; margin-left: .5em" href='https://imlixinyang.github.io/director3d-page'><img src='https://img.shields.io/badge/Project_Page-Director3D-green' alt='Project Page'></a>
</div>
Abs: We introduce Director3D, a robust open-world text-to-3D generation framework, designed to generate both real-world 3D scenes and adaptive camera trajectories. To achieve this, (1) we first utilize a Trajectory Diffusion Transformer, acting as the <a style="background-color:#EFF185">Cinematographer</a>, to model the distribution of camera trajectories based on textual descriptions. Next, a Gaussian-driven Multi-view Latent Diffusion Model serves as the <a style="background-color:#93D881">Decorator</a>, modeling the image sequence distribution given the camera trajectories and texts. This model, fine-tuned from a 2D diffusion model, directly generates pixel-aligned 3D Gaussians as an immediate 3D scene representation for consistent denoising. Lastly, the 3D Gaussians are further refined by a novel SDS++ loss as the <a style="background-color:#6FE4E4">Detailer</a>, which incorporates the prior of the 2D diffusion model.
"""
)
with gr.Column():
with gr.Row():
input_text = gr.Textbox(value="A delicious hamburger on a wooden table", label="Input Text",interactive=True)
with gr.Row():
with gr.Column(scale=1):
with gr.Tab("Step 1: Cinematographer", interactive=True):
run_btn0 = gr.Button("Generate Camera Trajectories", variant='primary')
# outcameras = gr.Gallery(label='Camera Trajectories', interactive=False, show_download_button=False, show_share_button=False)
with gr.Column(scale=1):
with gr.Row():
outcamera1 = gr.Image(label='1', interactive=False, show_download_button=False, show_share_button=False)
outcamera2 = gr.Image(label='2', interactive=False, show_download_button=False, show_share_button=False)
with gr.Column(scale=1):
with gr.Row():
outcamera3 = gr.Image(label='3', interactive=False, show_download_button=False, show_share_button=False)
outcamera4 = gr.Image(label='4', interactive=False, show_download_button=False, show_share_button=False)
radio0 = gr.Radio(
[1, 2, 3, 4], label="Selected Trajectory", value=1, interactive=True, info="Which trajectory do you prefer for the next step?"
)
with gr.Column(scale=1):
with gr.Tab("Step 2: Decorator", interactive=True):
run_btn1 = gr.Button("Generate Coarse 3D Scenes", variant='primary')
with gr.Column(scale=1):
with gr.Row():
outvideo1 = gr.Video(label='1', interactive=False, show_download_button=False, show_share_button=False, autoplay=True)
outvideo2 = gr.Video(label='2', interactive=False, show_download_button=False, show_share_button=False, autoplay=True)
with gr.Column(scale=1):
with gr.Row():
outvideo3 = gr.Video(label='3', interactive=False, show_download_button=False, show_share_button=False, autoplay=True)
outvideo4 = gr.Video(label='4', interactive=False, show_download_button=False, show_share_button=False, autoplay=True)
radio1 = gr.Radio(
[1, 2, 3, 4], label="Selected Scene", value=1, interactive=True, info="Which scene do you prefer for the next step?"
)
with gr.Column(scale=1):
with gr.Tab("Step 3: Detailer", interactive=True):
run_btn2 = gr.Button("Refine 3D Scene", variant='primary')
outvideo_refined = gr.Video(label='Refined Scene', interactive=False, show_download_button=False, show_share_button=False, autoplay=True)
run_btn0.click(fn= self.generate_camera_trajectory,
inputs=[input_text, data],
outputs=[outcamera1, outcamera2, outcamera3, outcamera4, data],
show_progress='full'
)
run_btn1.click(fn= self.generate_coarse_3D_scenes,
inputs=[input_text, radio0, data],
outputs=[outvideo1, outvideo2, outvideo3, outvideo4, data],
show_progress='full'
)
run_btn2.click(fn=self.refine_3D_scene,
inputs=[input_text, radio1, data],
outputs=[outvideo_refined, data],
show_progress='full'
)
def plotly_scene_visualization(self, npy_file):
camera_scene=[]
cameras = np.load(npy_file)
cameras = cameras[0]
num_frames = cameras.shape[0]
cameras = torch.from_numpy(cameras)
camera = {}
for i in range(0,num_frames,4):
c2w = torch.eye(4)
c2w[:3,:] =cameras[i][:12].reshape(3,4)
fx,fy,cx,cy,H,W = cameras[i][12:].chunk(6,-1)
K = [
[fx, 0, cx, 0],
[0, fy, cy, 0],
[0, 0, 0, 1],
[0, 0, 1, 0],
]
K = torch.tensor(K).unsqueeze(0)
#fov = 2 * np.arctan2(H / 2, K[0, 0, 0])
c2w[:3,1] *= -1
c2w[:3,2] *= -1
w2c = get_w2c(c2w)
camera[i] = PerspectiveCameras(R=w2c[None,:3,:3], T=w2c[None,:3,3],K=K)
camera_scene.append(camera)
# for visual convinience
array_axs =[1, 0.5, 0,-0.5, -1]
array_axs = [x * 1.5 for x in array_axs]
range_axs = [-1, 1]
range_axs = [x * 1.5 for x in range_axs]
dist = -2.7
elev = -52
azim = 180
# demo default view transform
R, T = look_at_view_transform(dist, elev, azim)
# craeate view camera
cameras_view = FoVPerspectiveCameras(R=R, T=T)
fig = plot_scene(
{" ":camera_scene[0]},
yaxis={ "title": "",
"backgroundcolor":"rgb(200, 200, 230)",
'tickmode': 'array',
'tickvals': array_axs,
'range':range_axs,
},
zaxis={ "title": "",
'tickmode': 'array',
'tickvals': array_axs,
'range':range_axs,
},
xaxis={ "title": "",
'tickmode': 'array',
'tickvals': array_axs,
'range':range_axs,
},
camera_scale=0.08,
axis_args=AxisArgs(showline=False,showgrid=True,zeroline=False,showticklabels=False,showaxeslabels=False),
viewpoint_cameras=cameras_view,
)
cmap = plt.get_cmap("hsv")
for i in range(len(fig.data)):
fig.data[i].line.color = matplotlib.colors.to_hex(cmap((i*4%32) / (num_frames)))
fig.update_layout(showlegend=False)
return fig
def generate_camera_trajectory(self, text, data, progress=gr.Progress(track_tqdm=True)):
data['cameras'] = []
data['text'] = text
# create cameras traj
with torch.no_grad():
video_paths= []
for i in tqdm.tqdm(range(4)):
print(i)
filename = text
extra_filename = f'_{i}'
cameras = self.system_traj_dit.inference(data["text"])
data['cameras'].append(cameras)
os.makedirs(os.path.join(data['tmp_dir'], 'camera'), exist_ok=True)
np.save(os.path.join(data['tmp_dir'], 'camera', f'{filename}{extra_filename}.npy'), cameras.detach().cpu().numpy())
# camearas traj visualization
tmp_dir = data['tmp_dir']
image_out = f'{tmp_dir}/camera_visualization/images'
html_out = f'{tmp_dir}/camera_visualization/htmls'
os.makedirs(image_out, exist_ok=True)
os.makedirs(html_out, exist_ok=True)
directory = os.path.join(data['tmp_dir'], 'camera')
files = os.listdir(directory)
files = sorted(files)
image_paths = []
for file in files:
if not file.endswith('.npy'):
continue
if not file.startswith(data["text"]):
continue
file_path = os.path.join(directory, file)
npy_files = file_path
fig = self.plotly_scene_visualization(npy_files)
image_path = f"{image_out}/{file[:-4]}.png"
image_paths.append(image_path)
plotly.io.write_image(fig, image_path, width=800, height=800)
html_plot = plotly.io.to_html(fig, full_html=True, include_plotlyjs="cdn")
s = io.BytesIO()
plt.savefig(s, format="png", bbox_inches="tight")
plt.close()
image_encoded = base64.b64encode(s.getvalue()).decode("utf-8").replace("\n", "")
html_path_out = f"{html_out}/{file[:-4]}.html"
with open(html_path_out, "w") as f:
s = HTML_TEMPLATE.format(
image_encoded=image_encoded,
plotly_html=html_plot,
)
f.write(s)
return image_paths[0], image_paths[1], image_paths[2], image_paths[3], data
def generate_coarse_3D_scenes(self, text, camera_id, data, progress=gr.Progress(track_tqdm=True)):
data['text'] = text
data['results'] = []
data['camera_id'] = camera_id
cameras = data['cameras'][camera_id - 1]
with torch.no_grad():
video_paths = []
for i in tqdm.tqdm(range(4)):
filename = data["text"]
extra_filename = f'_{i}'
sparse_cameras = cameras[:, ::int((cameras.shape[1]-1)/(self.system_gm_ldm.num_input_views-1))]
result = self.system_gm_ldm.inference(sparse_cameras, data["text"], use_3d_mode_every_m_steps=10)
data['results'].append(result)
os.makedirs(os.path.join(data['tmp_dir'], 'video'), exist_ok=True)
render_fn = lambda cameras, h=512, w=512: self.system_gm_ldm.model.render(cameras, result['gaussians'], h=h, w=w, bg_color=None)[:2]
video_path = os.path.join(os.path.join(data['tmp_dir'], 'video'), f'{filename}{extra_filename}.mp4')
export_video(render_fn, os.path.join(data['tmp_dir'], 'video'), f'{filename}{extra_filename}', cameras, device=self.device, fps=20, num_frames=240)
video_paths.append(video_path)
return video_paths[0], video_paths[1], video_paths[2], video_paths[3], data
def refine_3D_scene(self, text, scene_id, data, progress=gr.Progress(track_tqdm=True)):
data['text'] = text
cameras = data['cameras'][data['camera_id'] - 1]
gaussians = data['results'][scene_id - 1]['gaussians']
filename = text
extra_filename = f'_{scene_id - 1}'
gaussians = self.refiner.refine_gaussians(gaussians, text, dense_cameras=cameras)
render_fn = lambda cameras, h=512, w=512: self.system_gm_ldm.model.render(cameras, gaussians, h=h, w=w, bg_color=None)[:2]
video_path = os.path.join(os.path.join(data['tmp_dir'], 'video'), f'{filename}{extra_filename}_refined.mp4')
export_video(render_fn, os.path.join(data['tmp_dir'], 'video'), f'{filename}{extra_filename}_refined', cameras, device=self.device)
return video_path, data
def launch(self, only_layout=False, **kwargs):
if not only_layout:
self.init_models()
self.app.launch(**kwargs)
def queue(self, **kwargs):
self.app.queue(**kwargs)
if __name__ == "__main__":
max_threads = 2
api_open = True
max_size = 10
app = App()
app.queue(api_open=api_open, max_size=max_size)
app.launch(only_layout=False, share=False)