Spaces:
Runtime error
Runtime error
File size: 15,062 Bytes
c24da45 |
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 |
import numpy as np
import torch
from imagedream.camera_utils import get_camera_for_index
from imagedream.ldm.util import set_seed, add_random_background
from libs.base_utils import do_resize_content
from imagedream.ldm.models.diffusion.ddim import DDIMSampler
from torchvision import transforms as T
class ImageDreamDiffusion:
def __init__(
self,
model,
device,
dtype,
mode,
num_frames,
camera_views,
ref_position,
random_background=False,
offset_noise=False,
resize_rate=1,
image_size=256,
seed=1234,
) -> None:
assert mode in ["pixel", "local"]
size = image_size
self.seed = seed
batch_size = max(4, num_frames)
neg_texts = "uniform low no texture ugly, boring, bad anatomy, blurry, pixelated, obscure, unnatural colors, poor lighting, dull, and unclear."
uc = model.get_learned_conditioning([neg_texts]).to(device)
sampler = DDIMSampler(model)
# pre-compute camera matrices
camera = [get_camera_for_index(i).squeeze() for i in camera_views]
camera[ref_position] = torch.zeros_like(camera[ref_position]) # set ref camera to zero
camera = torch.stack(camera)
camera = camera.repeat(batch_size // num_frames, 1).to(device)
self.image_transform = T.Compose(
[
T.Resize((size, size)),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
self.dtype = dtype
self.ref_position = ref_position
self.mode = mode
self.random_background = random_background
self.resize_rate = resize_rate
self.num_frames = num_frames
self.size = size
self.device = device
self.batch_size = batch_size
self.model = model
self.sampler = sampler
self.uc = uc
self.camera = camera
self.offset_noise = offset_noise
@staticmethod
def i2i(
model,
image_size,
prompt,
uc,
sampler,
ip=None,
step=20,
scale=5.0,
batch_size=8,
ddim_eta=0.0,
dtype=torch.float32,
device="cuda",
camera=None,
num_frames=4,
pixel_control=False,
transform=None,
offset_noise=False,
):
""" The function supports additional image prompt.
Args:
model (_type_): the image dream model
image_size (_type_): size of diffusion output (standard 256)
prompt (_type_): text prompt for the image (prompt in type str)
uc (_type_): unconditional vector (tensor in shape [1, 77, 1024])
sampler (_type_): imagedream.ldm.models.diffusion.ddim.DDIMSampler
ip (Image, optional): the image prompt. Defaults to None.
step (int, optional): _description_. Defaults to 20.
scale (float, optional): _description_. Defaults to 7.5.
batch_size (int, optional): _description_. Defaults to 8.
ddim_eta (float, optional): _description_. Defaults to 0.0.
dtype (_type_, optional): _description_. Defaults to torch.float32.
device (str, optional): _description_. Defaults to "cuda".
camera (_type_, optional): camera info in tensor, shape: torch.Size([5, 16]) mean: 0.11, std: 0.49, min: -1.00, max: 1.00
num_frames (int, optional): _num of frames (views) to generate
pixel_control: whether to use pixel conditioning. Defaults to False, True when using pixel mode
transform: Compose(
Resize(size=(256, 256), interpolation=bilinear, max_size=None, antialias=warn)
ToTensor()
Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
)
"""
ip_raw = ip
if type(prompt) != list:
prompt = [prompt]
with torch.no_grad(), torch.autocast(device_type=torch.device(device).type, dtype=dtype):
c = model.get_learned_conditioning(prompt).to(
device
) # shape: torch.Size([1, 77, 1024]) mean: -0.17, std: 1.02, min: -7.50, max: 13.05
c_ = {"context": c.repeat(batch_size, 1, 1)} # batch_size
uc_ = {"context": uc.repeat(batch_size, 1, 1)}
if camera is not None:
c_["camera"] = uc_["camera"] = (
camera # shape: torch.Size([5, 16]) mean: 0.11, std: 0.49, min: -1.00, max: 1.00
)
c_["num_frames"] = uc_["num_frames"] = num_frames
if ip is not None:
ip_embed = model.get_learned_image_conditioning(ip).to(
device
) # shape: torch.Size([1, 257, 1280]) mean: 0.06, std: 0.53, min: -6.83, max: 11.12
ip_ = ip_embed.repeat(batch_size, 1, 1)
c_["ip"] = ip_
uc_["ip"] = torch.zeros_like(ip_)
if pixel_control:
assert camera is not None
ip = transform(ip).to(
device
) # shape: torch.Size([3, 256, 256]) mean: 0.33, std: 0.37, min: -1.00, max: 1.00
ip_img = model.get_first_stage_encoding(
model.encode_first_stage(ip[None, :, :, :])
) # shape: torch.Size([1, 4, 32, 32]) mean: 0.23, std: 0.77, min: -4.42, max: 3.55
c_["ip_img"] = ip_img
uc_["ip_img"] = torch.zeros_like(ip_img)
shape = [4, image_size // 8, image_size // 8] # [4, 32, 32]
if offset_noise:
ref = transform(ip_raw).to(device)
ref_latent = model.get_first_stage_encoding(model.encode_first_stage(ref[None, :, :, :]))
ref_mean = ref_latent.mean(dim=(-1, -2), keepdim=True)
time_steps = torch.randint(model.num_timesteps - 1, model.num_timesteps, (batch_size,), device=device)
x_T = model.q_sample(torch.ones([batch_size] + shape, device=device) * ref_mean, time_steps)
samples_ddim, _ = (
sampler.sample( # shape: torch.Size([5, 4, 32, 32]) mean: 0.29, std: 0.85, min: -3.38, max: 4.43
S=step,
conditioning=c_,
batch_size=batch_size,
shape=shape,
verbose=False,
unconditional_guidance_scale=scale,
unconditional_conditioning=uc_,
eta=ddim_eta,
x_T=x_T if offset_noise else None,
)
)
x_sample = model.decode_first_stage(samples_ddim)
x_sample = torch.clamp((x_sample + 1.0) / 2.0, min=0.0, max=1.0)
x_sample = 255.0 * x_sample.permute(0, 2, 3, 1).cpu().numpy()
return list(x_sample.astype(np.uint8))
def diffuse(self, t, ip, n_test=2):
set_seed(self.seed)
ip = do_resize_content(ip, self.resize_rate)
if self.random_background:
ip = add_random_background(ip)
images = []
for _ in range(n_test):
img = self.i2i(
self.model,
self.size,
t,
self.uc,
self.sampler,
ip=ip,
step=50,
scale=5,
batch_size=self.batch_size,
ddim_eta=0.0,
dtype=self.dtype,
device=self.device,
camera=self.camera,
num_frames=self.num_frames,
pixel_control=(self.mode == "pixel"),
transform=self.image_transform,
offset_noise=self.offset_noise,
)
img = np.concatenate(img, 1)
img = np.concatenate((img, ip.resize((self.size, self.size))), axis=1)
images.append(img)
set_seed() # unset random and numpy seed
return images
class ImageDreamDiffusionStage2:
def __init__(
self,
model,
device,
dtype,
num_frames,
camera_views,
ref_position,
random_background=False,
offset_noise=False,
resize_rate=1,
mode="pixel",
image_size=256,
seed=1234,
) -> None:
assert mode in ["pixel", "local"]
size = image_size
self.seed = seed
batch_size = max(4, num_frames)
neg_texts = "uniform low no texture ugly, boring, bad anatomy, blurry, pixelated, obscure, unnatural colors, poor lighting, dull, and unclear."
uc = model.get_learned_conditioning([neg_texts]).to(device)
sampler = DDIMSampler(model)
# pre-compute camera matrices
camera = [get_camera_for_index(i).squeeze() for i in camera_views]
if ref_position is not None:
camera[ref_position] = torch.zeros_like(camera[ref_position]) # set ref camera to zero
camera = torch.stack(camera)
camera = camera.repeat(batch_size // num_frames, 1).to(device)
self.image_transform = T.Compose(
[
T.Resize((size, size)),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
self.dtype = dtype
self.mode = mode
self.ref_position = ref_position
self.random_background = random_background
self.resize_rate = resize_rate
self.num_frames = num_frames
self.size = size
self.device = device
self.batch_size = batch_size
self.model = model
self.sampler = sampler
self.uc = uc
self.camera = camera
self.offset_noise = offset_noise
@staticmethod
def i2iStage2(
model,
image_size,
prompt,
uc,
sampler,
pixel_images,
ip=None,
step=20,
scale=5.0,
batch_size=8,
ddim_eta=0.0,
dtype=torch.float32,
device="cuda",
camera=None,
num_frames=4,
pixel_control=False,
transform=None,
offset_noise=False,
):
ip_raw = ip
if type(prompt) != list:
prompt = [prompt]
with torch.no_grad(), torch.autocast(device_type=torch.device(device).type, dtype=dtype):
c = model.get_learned_conditioning(prompt).to(
device
) # shape: torch.Size([1, 77, 1024]) mean: -0.17, std: 1.02, min: -7.50, max: 13.05
c_ = {"context": c.repeat(batch_size, 1, 1)} # batch_size
uc_ = {"context": uc.repeat(batch_size, 1, 1)}
if camera is not None:
c_["camera"] = uc_["camera"] = (
camera # shape: torch.Size([5, 16]) mean: 0.11, std: 0.49, min: -1.00, max: 1.00
)
c_["num_frames"] = uc_["num_frames"] = num_frames
if ip is not None:
ip_embed = model.get_learned_image_conditioning(ip).to(
device
) # shape: torch.Size([1, 257, 1280]) mean: 0.06, std: 0.53, min: -6.83, max: 11.12
ip_ = ip_embed.repeat(batch_size, 1, 1)
c_["ip"] = ip_
uc_["ip"] = torch.zeros_like(ip_)
if pixel_control:
assert camera is not None
transed_pixel_images = torch.stack([transform(i).to(device) for i in pixel_images])
latent_pixel_images = model.get_first_stage_encoding(model.encode_first_stage(transed_pixel_images))
c_["pixel_images"] = latent_pixel_images
uc_["pixel_images"] = torch.zeros_like(latent_pixel_images)
shape = [4, image_size // 8, image_size // 8] # [4, 32, 32]
if offset_noise:
ref = transform(ip_raw).to(device)
ref_latent = model.get_first_stage_encoding(model.encode_first_stage(ref[None, :, :, :]))
ref_mean = ref_latent.mean(dim=(-1, -2), keepdim=True)
time_steps = torch.randint(model.num_timesteps - 1, model.num_timesteps, (batch_size,), device=device)
x_T = model.q_sample(torch.ones([batch_size] + shape, device=device) * ref_mean, time_steps)
samples_ddim, _ = (
sampler.sample( # shape: torch.Size([5, 4, 32, 32]) mean: 0.29, std: 0.85, min: -3.38, max: 4.43
S=step,
conditioning=c_,
batch_size=batch_size,
shape=shape,
verbose=False,
unconditional_guidance_scale=scale,
unconditional_conditioning=uc_,
eta=ddim_eta,
x_T=x_T if offset_noise else None,
)
)
x_sample = model.decode_first_stage(samples_ddim)
x_sample = torch.clamp((x_sample + 1.0) / 2.0, min=0.0, max=1.0)
x_sample = 255.0 * x_sample.permute(0, 2, 3, 1).cpu().numpy()
return list(x_sample.astype(np.uint8))
@torch.no_grad()
def diffuse(self, t, ip, pixel_images, n_test=2):
set_seed(self.seed)
ip = do_resize_content(ip, self.resize_rate)
pixel_images = [do_resize_content(i, self.resize_rate) for i in pixel_images]
if self.random_background:
bg_color = np.random.rand() * 255
ip = add_random_background(ip, bg_color)
pixel_images = [add_random_background(i, bg_color) for i in pixel_images]
images = []
for _ in range(n_test):
img = self.i2iStage2(
self.model,
self.size,
t,
self.uc,
self.sampler,
pixel_images=pixel_images,
ip=ip,
step=50,
scale=5,
batch_size=self.batch_size,
ddim_eta=0.0,
dtype=self.dtype,
device=self.device,
camera=self.camera,
num_frames=self.num_frames,
pixel_control=(self.mode == "pixel"),
transform=self.image_transform,
offset_noise=self.offset_noise,
)
img = np.concatenate(img, 1)
img = np.concatenate(
(img, ip.resize((self.size, self.size)), *[i.resize((self.size, self.size)) for i in pixel_images]),
axis=1,
)
images.append(img)
set_seed() # unset random and numpy seed
return images
|