Spaces:
Running
on
A10G
Running
on
A10G
Linoy Tsaban
commited on
Commit
·
d754544
1
Parent(s):
d988f61
Update app.py
Browse filestesting interm yielding
app.py
CHANGED
@@ -5,10 +5,317 @@ from io import BytesIO
|
|
5 |
from diffusers import StableDiffusionPipeline
|
6 |
from diffusers import DDIMScheduler
|
7 |
from utils import *
|
8 |
-
from inversion_utils import *
|
9 |
from torch import autocast, inference_mode
|
10 |
import re
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):
|
13 |
|
14 |
# inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
|
@@ -34,7 +341,7 @@ def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta
|
|
34 |
def sample(wt, zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
35 |
|
36 |
# reverse process (via Zs and wT)
|
37 |
-
w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=
|
38 |
|
39 |
# vae decode image
|
40 |
with autocast("cuda"), inference_mode():
|
@@ -91,9 +398,76 @@ def edit(input_image,
|
|
91 |
wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
|
92 |
|
93 |
#
|
94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
-
return output
|
97 |
|
98 |
|
99 |
|
|
|
5 |
from diffusers import StableDiffusionPipeline
|
6 |
from diffusers import DDIMScheduler
|
7 |
from utils import *
|
8 |
+
# from inversion_utils import *
|
9 |
from torch import autocast, inference_mode
|
10 |
import re
|
11 |
|
12 |
+
|
13 |
+
############################################################################################################################################################################
|
14 |
+
import torch
|
15 |
+
import os
|
16 |
+
from tqdm import tqdm
|
17 |
+
from PIL import Image, ImageDraw ,ImageFont
|
18 |
+
from matplotlib import pyplot as plt
|
19 |
+
import torchvision.transforms as T
|
20 |
+
import os
|
21 |
+
import yaml
|
22 |
+
import numpy as np
|
23 |
+
import gradio as gr
|
24 |
+
|
25 |
+
|
26 |
+
def load_512(image_path, left=0, right=0, top=0, bottom=0, device=None):
|
27 |
+
if type(image_path) is str:
|
28 |
+
image = np.array(Image.open(image_path).convert('RGB'))[:, :, :3]
|
29 |
+
else:
|
30 |
+
image = image_path
|
31 |
+
h, w, c = image.shape
|
32 |
+
left = min(left, w-1)
|
33 |
+
right = min(right, w - left - 1)
|
34 |
+
top = min(top, h - left - 1)
|
35 |
+
bottom = min(bottom, h - top - 1)
|
36 |
+
image = image[top:h-bottom, left:w-right]
|
37 |
+
h, w, c = image.shape
|
38 |
+
if h < w:
|
39 |
+
offset = (w - h) // 2
|
40 |
+
image = image[:, offset:offset + h]
|
41 |
+
elif w < h:
|
42 |
+
offset = (h - w) // 2
|
43 |
+
image = image[offset:offset + w]
|
44 |
+
image = np.array(Image.fromarray(image).resize((512, 512)))
|
45 |
+
image = torch.from_numpy(image).float() / 127.5 - 1
|
46 |
+
image = image.permute(2, 0, 1).unsqueeze(0).to(device)
|
47 |
+
|
48 |
+
return image
|
49 |
+
|
50 |
+
|
51 |
+
def load_real_image(folder = "data/", img_name = None, idx = 0, img_size=512, device='cuda'):
|
52 |
+
from PIL import Image
|
53 |
+
from glob import glob
|
54 |
+
if img_name is not None:
|
55 |
+
path = os.path.join(folder, img_name)
|
56 |
+
else:
|
57 |
+
path = glob(folder + "*")[idx]
|
58 |
+
|
59 |
+
img = Image.open(path).resize((img_size,
|
60 |
+
img_size))
|
61 |
+
|
62 |
+
img = pil_to_tensor(img).to(device)
|
63 |
+
|
64 |
+
if img.shape[1]== 4:
|
65 |
+
img = img[:,:3,:,:]
|
66 |
+
return img
|
67 |
+
|
68 |
+
def mu_tilde(model, xt,x0, timestep):
|
69 |
+
"mu_tilde(x_t, x_0) DDPM paper eq. 7"
|
70 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
71 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
72 |
+
alpha_t = model.scheduler.alphas[timestep]
|
73 |
+
beta_t = 1 - alpha_t
|
74 |
+
alpha_bar = model.scheduler.alphas_cumprod[timestep]
|
75 |
+
return ((alpha_prod_t_prev ** 0.5 * beta_t) / (1-alpha_bar)) * x0 + ((alpha_t**0.5 *(1-alpha_prod_t_prev)) / (1- alpha_bar))*xt
|
76 |
+
|
77 |
+
def sample_xts_from_x0(model, x0, num_inference_steps=50):
|
78 |
+
"""
|
79 |
+
Samples from P(x_1:T|x_0)
|
80 |
+
"""
|
81 |
+
# torch.manual_seed(43256465436)
|
82 |
+
alpha_bar = model.scheduler.alphas_cumprod
|
83 |
+
sqrt_one_minus_alpha_bar = (1-alpha_bar) ** 0.5
|
84 |
+
alphas = model.scheduler.alphas
|
85 |
+
betas = 1 - alphas
|
86 |
+
variance_noise_shape = (
|
87 |
+
num_inference_steps,
|
88 |
+
model.unet.in_channels,
|
89 |
+
model.unet.sample_size,
|
90 |
+
model.unet.sample_size)
|
91 |
+
|
92 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
93 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
|
94 |
+
xts = torch.zeros(variance_noise_shape).to(x0.device)
|
95 |
+
for t in reversed(timesteps):
|
96 |
+
idx = t_to_idx[int(t)]
|
97 |
+
xts[idx] = x0 * (alpha_bar[t] ** 0.5) + torch.randn_like(x0) * sqrt_one_minus_alpha_bar[t]
|
98 |
+
xts = torch.cat([xts, x0 ],dim = 0)
|
99 |
+
|
100 |
+
return xts
|
101 |
+
|
102 |
+
def encode_text(model, prompts):
|
103 |
+
text_input = model.tokenizer(
|
104 |
+
prompts,
|
105 |
+
padding="max_length",
|
106 |
+
max_length=model.tokenizer.model_max_length,
|
107 |
+
truncation=True,
|
108 |
+
return_tensors="pt",
|
109 |
+
)
|
110 |
+
with torch.no_grad():
|
111 |
+
text_encoding = model.text_encoder(text_input.input_ids.to(model.device))[0]
|
112 |
+
return text_encoding
|
113 |
+
|
114 |
+
def forward_step(model, model_output, timestep, sample):
|
115 |
+
next_timestep = min(model.scheduler.config.num_train_timesteps - 2,
|
116 |
+
timestep + model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps)
|
117 |
+
|
118 |
+
# 2. compute alphas, betas
|
119 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
120 |
+
# alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep] if next_ltimestep >= 0 else self.scheduler.final_alpha_cumprod
|
121 |
+
|
122 |
+
beta_prod_t = 1 - alpha_prod_t
|
123 |
+
|
124 |
+
# 3. compute predicted original sample from predicted noise also called
|
125 |
+
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
126 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
127 |
+
|
128 |
+
# 5. TODO: simple noising implementatiom
|
129 |
+
next_sample = model.scheduler.add_noise(pred_original_sample,
|
130 |
+
model_output,
|
131 |
+
torch.LongTensor([next_timestep]))
|
132 |
+
return next_sample
|
133 |
+
|
134 |
+
|
135 |
+
def get_variance(model, timestep): #, prev_timestep):
|
136 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
137 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
138 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
139 |
+
beta_prod_t = 1 - alpha_prod_t
|
140 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
141 |
+
variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
|
142 |
+
return variance
|
143 |
+
|
144 |
+
def inversion_forward_process(model, x0,
|
145 |
+
etas = None,
|
146 |
+
prog_bar = False,
|
147 |
+
prompt = "",
|
148 |
+
cfg_scale = 3.5,
|
149 |
+
num_inference_steps=50, eps = None
|
150 |
+
):
|
151 |
+
|
152 |
+
if not prompt=="":
|
153 |
+
text_embeddings = encode_text(model, prompt)
|
154 |
+
uncond_embedding = encode_text(model, "")
|
155 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
156 |
+
variance_noise_shape = (
|
157 |
+
num_inference_steps,
|
158 |
+
model.unet.in_channels,
|
159 |
+
model.unet.sample_size,
|
160 |
+
model.unet.sample_size)
|
161 |
+
if etas is None or (type(etas) in [int, float] and etas == 0):
|
162 |
+
eta_is_zero = True
|
163 |
+
zs = None
|
164 |
+
else:
|
165 |
+
eta_is_zero = False
|
166 |
+
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
|
167 |
+
xts = sample_xts_from_x0(model, x0, num_inference_steps=num_inference_steps)
|
168 |
+
alpha_bar = model.scheduler.alphas_cumprod
|
169 |
+
zs = torch.zeros(size=variance_noise_shape, device=model.device)
|
170 |
+
|
171 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
|
172 |
+
xt = x0
|
173 |
+
op = tqdm(reversed(timesteps)) if prog_bar else reversed(timesteps)
|
174 |
+
|
175 |
+
for t in op:
|
176 |
+
idx = t_to_idx[int(t)]
|
177 |
+
# 1. predict noise residual
|
178 |
+
if not eta_is_zero:
|
179 |
+
xt = xts[idx][None]
|
180 |
+
|
181 |
+
with torch.no_grad():
|
182 |
+
out = model.unet.forward(xt, timestep = t, encoder_hidden_states = uncond_embedding)
|
183 |
+
if not prompt=="":
|
184 |
+
cond_out = model.unet.forward(xt, timestep=t, encoder_hidden_states = text_embeddings)
|
185 |
+
|
186 |
+
if not prompt=="":
|
187 |
+
## classifier free guidance
|
188 |
+
noise_pred = out.sample + cfg_scale * (cond_out.sample - out.sample)
|
189 |
+
else:
|
190 |
+
noise_pred = out.sample
|
191 |
+
|
192 |
+
if eta_is_zero:
|
193 |
+
# 2. compute more noisy image and set x_t -> x_t+1
|
194 |
+
xt = forward_step(model, noise_pred, t, xt)
|
195 |
+
|
196 |
+
else:
|
197 |
+
xtm1 = xts[idx+1][None]
|
198 |
+
# pred of x0
|
199 |
+
pred_original_sample = (xt - (1-alpha_bar[t]) ** 0.5 * noise_pred ) / alpha_bar[t] ** 0.5
|
200 |
+
|
201 |
+
# direction to xt
|
202 |
+
prev_timestep = t - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
203 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
204 |
+
|
205 |
+
variance = get_variance(model, t)
|
206 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - etas[idx] * variance ) ** (0.5) * noise_pred
|
207 |
+
|
208 |
+
mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
209 |
+
|
210 |
+
z = (xtm1 - mu_xt ) / ( etas[idx] * variance ** 0.5 )
|
211 |
+
zs[idx] = z
|
212 |
+
|
213 |
+
# correction to avoid error accumulation
|
214 |
+
xtm1 = mu_xt + ( etas[idx] * variance ** 0.5 )*z
|
215 |
+
xts[idx+1] = xtm1
|
216 |
+
|
217 |
+
if not zs is None:
|
218 |
+
zs[-1] = torch.zeros_like(zs[-1])
|
219 |
+
|
220 |
+
return xt, zs, xts
|
221 |
+
|
222 |
+
|
223 |
+
def reverse_step(model, model_output, timestep, sample, eta = 0, variance_noise=None):
|
224 |
+
# 1. get previous step value (=t-1)
|
225 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
226 |
+
# 2. compute alphas, betas
|
227 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
228 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
229 |
+
beta_prod_t = 1 - alpha_prod_t
|
230 |
+
# 3. compute predicted original sample from predicted noise also called
|
231 |
+
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
232 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
233 |
+
# 5. compute variance: "sigma_t(η)" -> see formula (16)
|
234 |
+
# σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
|
235 |
+
# variance = self.scheduler._get_variance(timestep, prev_timestep)
|
236 |
+
variance = get_variance(model, timestep) #, prev_timestep)
|
237 |
+
std_dev_t = eta * variance ** (0.5)
|
238 |
+
# Take care of asymetric reverse process (asyrp)
|
239 |
+
model_output_direction = model_output
|
240 |
+
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
241 |
+
# pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output_direction
|
242 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - eta * variance) ** (0.5) * model_output_direction
|
243 |
+
# 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
244 |
+
prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
245 |
+
# 8. Add noice if eta > 0
|
246 |
+
if eta > 0:
|
247 |
+
if variance_noise is None:
|
248 |
+
variance_noise = torch.randn(model_output.shape, device=model.device)
|
249 |
+
sigma_z = eta * variance ** (0.5) * variance_noise
|
250 |
+
prev_sample = prev_sample + sigma_z
|
251 |
+
|
252 |
+
return prev_sample
|
253 |
+
|
254 |
+
def inversion_reverse_process(model,
|
255 |
+
xT,
|
256 |
+
etas = 0,
|
257 |
+
prompts = "",
|
258 |
+
cfg_scales = None,
|
259 |
+
prog_bar = False,
|
260 |
+
zs = None,
|
261 |
+
controller=None,
|
262 |
+
asyrp = False
|
263 |
+
):
|
264 |
+
|
265 |
+
batch_size = len(prompts)
|
266 |
+
|
267 |
+
cfg_scales_tensor = torch.Tensor(cfg_scales).view(-1,1,1,1).to(model.device)
|
268 |
+
|
269 |
+
text_embeddings = encode_text(model, prompts)
|
270 |
+
uncond_embedding = encode_text(model, [""] * batch_size)
|
271 |
+
|
272 |
+
if etas is None: etas = 0
|
273 |
+
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
|
274 |
+
assert len(etas) == model.scheduler.num_inference_steps
|
275 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
276 |
+
|
277 |
+
xt = xT.expand(batch_size, -1, -1, -1)
|
278 |
+
op = tqdm(timesteps[-zs.shape[0]:]) if prog_bar else timesteps[-zs.shape[0]:]
|
279 |
+
|
280 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
|
281 |
+
|
282 |
+
for t in op:
|
283 |
+
idx = t_to_idx[int(t)]
|
284 |
+
## Unconditional embedding
|
285 |
+
with torch.no_grad():
|
286 |
+
uncond_out = model.unet.forward(xt, timestep = t,
|
287 |
+
encoder_hidden_states = uncond_embedding)
|
288 |
+
|
289 |
+
## Conditional embedding
|
290 |
+
if prompts:
|
291 |
+
with torch.no_grad():
|
292 |
+
cond_out = model.unet.forward(xt, timestep = t,
|
293 |
+
encoder_hidden_states = text_embeddings)
|
294 |
+
|
295 |
+
|
296 |
+
z = zs[idx] if not zs is None else None
|
297 |
+
z = z.expand(batch_size, -1, -1, -1)
|
298 |
+
if prompts:
|
299 |
+
## classifier free guidance
|
300 |
+
noise_pred = uncond_out.sample + cfg_scales_tensor * (cond_out.sample - uncond_out.sample)
|
301 |
+
else:
|
302 |
+
noise_pred = uncond_out.sample
|
303 |
+
# 2. compute less noisy image and set x_t -> x_t-1
|
304 |
+
xt = reverse_step(model, noise_pred, t, xt, eta = etas[idx], variance_noise = z)
|
305 |
+
|
306 |
+
# interm denoised img
|
307 |
+
with autocast("cuda"), inference_mode():
|
308 |
+
x0_dec = sd_pipe.vae.decode(1 / 0.18215 * xt).sample
|
309 |
+
if x0_dec.dim()<4:
|
310 |
+
x0_dec = x0_dec[None,:,:,:]
|
311 |
+
interm_img = image_grid(x0_dec)
|
312 |
+
yield interm_img
|
313 |
+
|
314 |
+
if controller is not None:
|
315 |
+
xt = controller.step_callback(xt)
|
316 |
+
return xt, zs
|
317 |
+
############################################################################################################################################################################
|
318 |
+
|
319 |
def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):
|
320 |
|
321 |
# inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
|
|
|
341 |
def sample(wt, zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
|
342 |
|
343 |
# reverse process (via Zs and wT)
|
344 |
+
w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=False, zs=zs[skip:])
|
345 |
|
346 |
# vae decode image
|
347 |
with autocast("cuda"), inference_mode():
|
|
|
398 |
wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
|
399 |
|
400 |
#
|
401 |
+
xT=wts[skip]
|
402 |
+
etas=eta
|
403 |
+
prompts=[prompt_tar]
|
404 |
+
cfg_scales=[cfg_scale_tar]
|
405 |
+
prog_bar=False
|
406 |
+
zs=zs[skip:]
|
407 |
+
|
408 |
+
|
409 |
+
batch_size = len(prompts)
|
410 |
+
|
411 |
+
cfg_scales_tensor = torch.Tensor(cfg_scales).view(-1,1,1,1).to(model.device)
|
412 |
+
|
413 |
+
text_embeddings = encode_text(model, prompts)
|
414 |
+
uncond_embedding = encode_text(model, [""] * batch_size)
|
415 |
+
|
416 |
+
if etas is None: etas = 0
|
417 |
+
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
|
418 |
+
assert len(etas) == model.scheduler.num_inference_steps
|
419 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
420 |
+
|
421 |
+
xt = xT.expand(batch_size, -1, -1, -1)
|
422 |
+
op = tqdm(timesteps[-zs.shape[0]:]) if prog_bar else timesteps[-zs.shape[0]:]
|
423 |
+
|
424 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
|
425 |
+
|
426 |
+
for t in op:
|
427 |
+
idx = t_to_idx[int(t)]
|
428 |
+
## Unconditional embedding
|
429 |
+
with torch.no_grad():
|
430 |
+
uncond_out = model.unet.forward(xt, timestep = t,
|
431 |
+
encoder_hidden_states = uncond_embedding)
|
432 |
+
|
433 |
+
## Conditional embedding
|
434 |
+
if prompts:
|
435 |
+
with torch.no_grad():
|
436 |
+
cond_out = model.unet.forward(xt, timestep = t,
|
437 |
+
encoder_hidden_states = text_embeddings)
|
438 |
+
|
439 |
+
|
440 |
+
z = zs[idx] if not zs is None else None
|
441 |
+
z = z.expand(batch_size, -1, -1, -1)
|
442 |
+
if prompts:
|
443 |
+
## classifier free guidance
|
444 |
+
noise_pred = uncond_out.sample + cfg_scales_tensor * (cond_out.sample - uncond_out.sample)
|
445 |
+
else:
|
446 |
+
noise_pred = uncond_out.sample
|
447 |
+
# 2. compute less noisy image and set x_t -> x_t-1
|
448 |
+
xt = reverse_step(model, noise_pred, t, xt, eta = etas[idx], variance_noise = z)
|
449 |
+
|
450 |
+
# interm denoised img
|
451 |
+
with autocast("cuda"), inference_mode():
|
452 |
+
x0_dec = sd_pipe.vae.decode(1 / 0.18215 * xt).sample
|
453 |
+
if x0_dec.dim()<4:
|
454 |
+
x0_dec = x0_dec[None,:,:,:]
|
455 |
+
interm_img = image_grid(x0_dec)
|
456 |
+
yield interm_img
|
457 |
+
|
458 |
+
return interm_img
|
459 |
+
|
460 |
+
# # vae decode image
|
461 |
+
# with autocast("cuda"), inference_mode():
|
462 |
+
# x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
|
463 |
+
# if x0_dec.dim()<4:
|
464 |
+
# x0_dec = x0_dec[None,:,:,:]
|
465 |
+
# img = image_grid(x0_dec)
|
466 |
+
# return img
|
467 |
+
|
468 |
+
# output = sample(wt, zs, wts, prompt_tar=tar_prompt)
|
469 |
|
470 |
+
# return output
|
471 |
|
472 |
|
473 |
|