File size: 13,553 Bytes
a6e635b |
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 |
import sys
import os
os.system("git clone https://github.com/3DTopia/3DTopia")
sys.path.append("3DTopia")
import os
import cv2
import time
import json
import torch
import mcubes
import trimesh
import datetime
import argparse
import subprocess
import numpy as np
import gradio as gr
from tqdm import tqdm
import imageio.v2 as imageio
import pytorch_lightning as pl
from omegaconf import OmegaConf
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
from ldm.models.diffusion.dpm_solver import DPMSolverSampler
from utility.initialize import instantiate_from_config, get_obj_from_str
from utility.triplane_renderer.eg3d_renderer import sample_from_planes, generate_planes
from utility.triplane_renderer.renderer import get_rays, to8b
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
import warnings
warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)
def add_text(rgb, caption):
font = cv2.FONT_HERSHEY_SIMPLEX
# org
gap = 10
org = (gap, gap)
# fontScale
fontScale = 0.3
# Blue color in BGR
color = (255, 0, 0)
# Line thickness of 2 px
thickness = 1
break_caption = []
for i in range(len(caption) // 30 + 1):
break_caption_i = caption[i*30:(i+1)*30]
break_caption.append(break_caption_i)
for i, bci in enumerate(break_caption):
cv2.putText(rgb, bci, (gap, gap*(i+1)), font, fontScale, color, thickness, cv2.LINE_AA)
return rgb
config = "configs/default.yaml"
local_ckpt = "checkpoints/3dtopia_diffusion_state_dict.ckpt"
if os.path.exists(local_ckpt):
ckpt = local_ckpt
else:
ckpt = hf_hub_download(repo_id="hongfz16/3DTopia", filename="model.safetensors")
configs = OmegaConf.load(config)
os.makedirs("tmp", exist_ok=True)
if ckpt.endswith(".ckpt"):
model = get_obj_from_str(configs.model["target"]).load_from_checkpoint(ckpt, map_location='cpu', strict=False, **configs.model.params)
elif ckpt.endswith(".safetensors"):
model = get_obj_from_str(configs.model["target"])(**configs.model.params)
model_ckpt = load_file(ckpt)
model.load_state_dict(model_ckpt)
else:
raise NotImplementedError
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model = model.to(device)
sampler = DDIMSampler(model)
img_size = configs.model.params.unet_config.params.image_size
channels = configs.model.params.unet_config.params.in_channels
shape = [channels, img_size, img_size * 3]
pose_folder = 'assets/sample_data/pose'
poses_fname = sorted([os.path.join(pose_folder, f) for f in os.listdir(pose_folder)])
batch_rays_list = []
H = 128
ratio = 512 // H
for p in poses_fname:
c2w = np.loadtxt(p).reshape(4, 4)
c2w[:3, 3] *= 2.2
c2w = np.array([
[1, 0, 0, 0],
[0, 0, -1, 0],
[0, 1, 0, 0],
[0, 0, 0, 1]
]) @ c2w
k = np.array([
[560 / ratio, 0, H * 0.5],
[0, 560 / ratio, H * 0.5],
[0, 0, 1]
])
rays_o, rays_d = get_rays(H, H, torch.Tensor(k), torch.Tensor(c2w[:3, :4]))
coords = torch.stack(torch.meshgrid(torch.linspace(0, H-1, H), torch.linspace(0, H-1, H), indexing='ij'), -1)
coords = torch.reshape(coords, [-1,2]).long()
rays_o = rays_o[coords[:, 0], coords[:, 1]]
rays_d = rays_d[coords[:, 0], coords[:, 1]]
batch_rays = torch.stack([rays_o, rays_d], 0)
batch_rays_list.append(batch_rays)
batch_rays_list = torch.stack(batch_rays_list, 0)
def marching_cube(b, text, global_info):
# prepare volumn for marching cube
res = 128
assert 'decode_res' in global_info
decode_res = global_info['decode_res']
c_list = torch.linspace(-1.2, 1.2, steps=res)
grid_x, grid_y, grid_z = torch.meshgrid(
c_list, c_list, c_list, indexing='ij'
)
coords = torch.stack([grid_x, grid_y, grid_z], -1).to(device)
plane_axes = generate_planes()
feats = sample_from_planes(
plane_axes, decode_res[b:b+1].reshape(1, 3, -1, 256, 256), coords.reshape(1, -1, 3), padding_mode='zeros', box_warp=2.4
)
fake_dirs = torch.zeros_like(coords)
fake_dirs[..., 0] = 1
out = model.first_stage_model.triplane_decoder.decoder(feats, fake_dirs)
u = out['sigma'].reshape(res, res, res).detach().cpu().numpy()
del out
# marching cube
vertices, triangles = mcubes.marching_cubes(u, 10)
min_bound = np.array([-1.2, -1.2, -1.2])
max_bound = np.array([1.2, 1.2, 1.2])
vertices = vertices / (res - 1) * (max_bound - min_bound)[None, :] + min_bound[None, :]
pt_vertices = torch.from_numpy(vertices).to(device)
# extract vertices color
res_triplane = 256
render_kwargs = {
'depth_resolution': 128,
'disparity_space_sampling': False,
'box_warp': 2.4,
'depth_resolution_importance': 128,
'clamp_mode': 'softplus',
'white_back': True,
'det': True
}
rays_o_list = [
np.array([0, 0, 2]),
np.array([0, 0, -2]),
np.array([0, 2, 0]),
np.array([0, -2, 0]),
np.array([2, 0, 0]),
np.array([-2, 0, 0]),
]
rgb_final = None
diff_final = None
for rays_o in tqdm(rays_o_list):
rays_o = torch.from_numpy(rays_o.reshape(1, 3)).repeat(vertices.shape[0], 1).float().to(device)
rays_d = pt_vertices.reshape(-1, 3) - rays_o
rays_d = rays_d / torch.norm(rays_d, dim=-1).reshape(-1, 1)
dist = torch.norm(pt_vertices.reshape(-1, 3) - rays_o, dim=-1).cpu().numpy().reshape(-1)
render_out = model.first_stage_model.triplane_decoder(
decode_res[b:b+1].reshape(1, 3, -1, res_triplane, res_triplane),
rays_o.unsqueeze(0), rays_d.unsqueeze(0), render_kwargs,
whole_img=False, tvloss=False
)
rgb = render_out['rgb_marched'].reshape(-1, 3).detach().cpu().numpy()
depth = render_out['depth_final'].reshape(-1).detach().cpu().numpy()
depth_diff = np.abs(dist - depth)
if rgb_final is None:
rgb_final = rgb.copy()
diff_final = depth_diff.copy()
else:
ind = diff_final > depth_diff
rgb_final[ind] = rgb[ind]
diff_final[ind] = depth_diff[ind]
# bgr to rgb
rgb_final = np.stack([
rgb_final[:, 2], rgb_final[:, 1], rgb_final[:, 0]
], -1)
# export to ply
mesh = trimesh.Trimesh(vertices, triangles, vertex_colors=(rgb_final * 255).astype(np.uint8))
path = os.path.join('tmp', f"{text.replace(' ', '_')}_{str(datetime.datetime.now()).replace(' ', '_')}.ply")
trimesh.exchange.export.export_mesh(mesh, path, file_type='ply')
del vertices, triangles, rgb_final
torch.cuda.empty_cache()
return path
def infer(prompt, samples, steps, scale, seed, global_info):
prompt = prompt.replace('/', '')
pl.seed_everything(seed)
batch_size = samples
with torch.no_grad():
noise = None
c = model.get_learned_conditioning([prompt])
unconditional_c = torch.zeros_like(c)
sample, _ = sampler.sample(
S=steps,
batch_size=batch_size,
shape=shape,
verbose=False,
x_T = noise,
conditioning = c.repeat(batch_size, 1, 1),
unconditional_guidance_scale=scale,
unconditional_conditioning=unconditional_c.repeat(batch_size, 1, 1)
)
decode_res = model.decode_first_stage(sample)
big_video_list = []
global_info['decode_res'] = decode_res
for b in range(batch_size):
def render_img(v):
rgb_sample, _ = model.first_stage_model.render_triplane_eg3d_decoder(
decode_res[b:b+1], batch_rays_list[v:v+1].to(device), torch.zeros(1, H, H, 3).to(device),
)
rgb_sample = to8b(rgb_sample.detach().cpu().numpy())[0]
rgb_sample = np.stack(
[rgb_sample[..., 2], rgb_sample[..., 1], rgb_sample[..., 0]], -1
)
rgb_sample = add_text(rgb_sample, str(b))
return rgb_sample
view_num = len(batch_rays_list)
video_list = []
for v in tqdm(range(view_num//8*3, view_num//8*5, 2)):
rgb_sample = render_img(v)
video_list.append(rgb_sample)
big_video_list.append(video_list)
# if batch_size == 2:
# cat_video_list = [
# np.concatenate([big_video_list[j][i] for j in range(len(big_video_list))], 1) \
# for i in range(len(big_video_list[0]))
# ]
# elif batch_size > 2:
# if batch_size == 3:
# big_video_list.append(
# [np.zeros_like(f) for f in big_video_list[0]]
# )
# cat_video_list = [
# np.concatenate([
# np.concatenate([big_video_list[0][i], big_video_list[1][i]], 1),
# np.concatenate([big_video_list[2][i], big_video_list[3][i]], 1),
# ], 0) \
# for i in range(len(big_video_list[0]))
# ]
# else:
# cat_video_list = big_video_list[0]
for _ in range(4 - batch_size):
big_video_list.append(
[np.zeros_like(f) + 255 for f in big_video_list[0]]
)
cat_video_list = [
np.concatenate([
np.concatenate([big_video_list[0][i], big_video_list[1][i]], 1),
np.concatenate([big_video_list[2][i], big_video_list[3][i]], 1),
], 0) \
for i in range(len(big_video_list[0]))
]
path = f"tmp/{prompt.replace(' ', '_')}_{str(datetime.datetime.now()).replace(' ', '_')}.mp4"
imageio.mimwrite(path, np.stack(cat_video_list, 0))
return global_info, path
def infer_stage2(prompt, selection, seed, global_info):
prompt = prompt.replace('/', '')
mesh_path = marching_cube(int(selection), prompt, global_info)
mesh_name = mesh_path.split('/')[-1][:-4]
if2_cmd = f"threefiner if2 --mesh {mesh_path} --prompt \"{prompt}\" --outdir tmp --save {mesh_name}_if2.glb --text_dir --front_dir=-y"
print(if2_cmd)
# os.system(if2_cmd)
subprocess.Popen(if2_cmd, shell=True).wait()
torch.cuda.empty_cache()
video_path = f"tmp/{prompt.replace(' ', '_')}_{str(datetime.datetime.now()).replace(' ', '_')}.mp4"
render_cmd = f"kire {os.path.join('tmp', mesh_name + '_if2.glb')} --save_video {video_path} --wogui --force_cuda_rast --H 256 --W 256"
print(render_cmd)
# os.system(render_cmd)
subprocess.Popen(render_cmd, shell=True).wait()
torch.cuda.empty_cache()
return video_path, os.path.join('tmp', mesh_name + '_if2.glb')
markdown=f'''
# 3DTopia
A two-stage text-to-3D generation model. The first stage uses diffusion model to quickly generate candidates. The second stage refines the assets chosen from the first stage.
### Usage:
First enter prompt for a 3D object, hit "Generate 3D". Then choose one candidate from the dropdown options for the second stage refinement and hit "Start Refinement". The final mesh can be downloaded from the bottom right box.
### Runtime:
The first stage takes 30s if generating 4 samples. The second stage takes roughly 3 min.
### Useful links:
[Github Repo](https://github.com/3DTopia/3DTopia)
'''
block = gr.Blocks()
with block:
global_info = gr.State(dict())
gr.Markdown(markdown)
with gr.Row():
with gr.Column():
with gr.Row():
text = gr.Textbox(
label = "Enter your prompt",
max_lines = 1,
placeholder = "Enter your prompt",
container = False,
)
btn = gr.Button("Generate 3D")
gallery = gr.Video(height=512)
# advanced_button = gr.Button("Advanced Options", elem_id="advanced-btn")
with gr.Row(elem_id="advanced-options"):
with gr.Tab("Advanced options"):
samples = gr.Slider(label="Number of Samples", minimum=1, maximum=4, value=4, step=1)
steps = gr.Slider(label="Steps", minimum=1, maximum=500, value=50, step=1)
scale = gr.Slider(
label="Guidance Scale", minimum=0, maximum=50, value=7.5, step=0.1
)
seed = gr.Slider(
label="Seed",
minimum=0,
maximum=2147483647,
step=1,
randomize=True,
)
gr.on([text.submit, btn.click], infer, inputs=[text, samples, steps, scale, seed, global_info], outputs=[global_info, gallery])
# advanced_button.click(
# None,
# [],
# text,
# )
with gr.Column():
with gr.Row():
dropdown = gr.Dropdown(
['0', '1', '2', '3'], label="Choose a candidate for stage2", value='0'
)
btn_stage2 = gr.Button("Start Refinement")
gallery = gr.Video(height=512)
download = gr.File(label="Download mesh", file_count="single", height=100)
gr.on([btn_stage2.click], infer_stage2, inputs=[text, dropdown, seed, global_info], outputs=[gallery, download])
block.launch(share=True)
|