Spaces:
Running
on
Zero
Running
on
Zero
IceClear
commited on
Commit
·
56d0abb
1
Parent(s):
f61c793
delete degrad
Browse files
README.md
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
---
|
2 |
title: SeedVR2-3B
|
3 |
-
emoji:
|
4 |
colorFrom: blue
|
5 |
colorTo: green
|
6 |
sdk: gradio
|
|
|
1 |
---
|
2 |
title: SeedVR2-3B
|
3 |
+
emoji: 🎥
|
4 |
colorFrom: blue
|
5 |
colorTo: green
|
6 |
sdk: gradio
|
projects/video_diffusion_sr/degradation_utils.py
DELETED
@@ -1,522 +0,0 @@
|
|
1 |
-
# Copyright (c) 2022 BasicSR: Xintao Wang and Liangbin Xie and Ke Yu and Kelvin C.K. Chan and Chen Change Loy and Chao Dong
|
2 |
-
# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates.
|
3 |
-
# SPDX-License-Identifier: Apache License, Version 2.0 (the "License")
|
4 |
-
#
|
5 |
-
# This file has been modified by ByteDance Ltd. and/or its affiliates. on 1st June 2025
|
6 |
-
#
|
7 |
-
# Original file was released under Apache License, Version 2.0 (the "License"), with the full license text
|
8 |
-
# available at http://www.apache.org/licenses/LICENSE-2.0.
|
9 |
-
#
|
10 |
-
# This modified file is released under the same license.
|
11 |
-
|
12 |
-
import io
|
13 |
-
import math
|
14 |
-
import random
|
15 |
-
from typing import Dict
|
16 |
-
import av
|
17 |
-
import numpy as np
|
18 |
-
import torch
|
19 |
-
from basicsr.data.degradations import (
|
20 |
-
circular_lowpass_kernel,
|
21 |
-
random_add_gaussian_noise_pt,
|
22 |
-
random_add_poisson_noise_pt,
|
23 |
-
random_mixed_kernels,
|
24 |
-
)
|
25 |
-
from basicsr.utils import DiffJPEG, USMSharp
|
26 |
-
from basicsr.utils.img_process_util import filter2D
|
27 |
-
from einops import rearrange
|
28 |
-
from torch import nn
|
29 |
-
from torch.nn import functional as F
|
30 |
-
|
31 |
-
|
32 |
-
def remove_prefix(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
|
33 |
-
for k in list(state_dict.keys()):
|
34 |
-
if k.startswith("_flops_wrap_module."):
|
35 |
-
v = state_dict.pop(k)
|
36 |
-
state_dict[k.replace("_flops_wrap_module.", "")] = v
|
37 |
-
if k.startswith("module."):
|
38 |
-
v = state_dict.pop(k)
|
39 |
-
state_dict[k.replace("module.", "")] = v
|
40 |
-
return state_dict
|
41 |
-
|
42 |
-
|
43 |
-
def clean_memory_bank(module: nn.Module):
|
44 |
-
if hasattr(module, "padding_bank"):
|
45 |
-
module.padding_bank = None
|
46 |
-
for child in module.children():
|
47 |
-
clean_memory_bank(child)
|
48 |
-
|
49 |
-
|
50 |
-
para_dic = {
|
51 |
-
"kernel_list": [
|
52 |
-
"iso",
|
53 |
-
"aniso",
|
54 |
-
"generalized_iso",
|
55 |
-
"generalized_aniso",
|
56 |
-
"plateau_iso",
|
57 |
-
"plateau_aniso",
|
58 |
-
],
|
59 |
-
"kernel_prob": [0.45, 0.25, 0.12, 0.03, 0.12, 0.03],
|
60 |
-
"sinc_prob": 0.1,
|
61 |
-
"blur_sigma": [0.2, 1.5],
|
62 |
-
"betag_range": [0.5, 2.0],
|
63 |
-
"betap_range": [1, 1.5],
|
64 |
-
"kernel_list2": [
|
65 |
-
"iso",
|
66 |
-
"aniso",
|
67 |
-
"generalized_iso",
|
68 |
-
"generalized_aniso",
|
69 |
-
"plateau_iso",
|
70 |
-
"plateau_aniso",
|
71 |
-
],
|
72 |
-
"kernel_prob2": [0.45, 0.25, 0.12, 0.03, 0.12, 0.03],
|
73 |
-
"sinc_prob2": 0.1,
|
74 |
-
"blur_sigma2": [0.2, 1.0],
|
75 |
-
"betag_range2": [0.5, 2.0],
|
76 |
-
"betap_range2": [1, 1.5],
|
77 |
-
"final_sinc_prob": 0.5,
|
78 |
-
}
|
79 |
-
|
80 |
-
degrade_dic = {
|
81 |
-
# "gt_usm": True, # USM the ground-truth
|
82 |
-
# the first degradation process
|
83 |
-
"resize_prob": [0.2, 0.7, 0.1], # up, down, keep
|
84 |
-
"resize_range": [0.3, 1.5],
|
85 |
-
"gaussian_noise_prob": 0.5,
|
86 |
-
"noise_range": [1, 15],
|
87 |
-
"poisson_scale_range": [0.05, 2],
|
88 |
-
"gray_noise_prob": 0.4,
|
89 |
-
"jpeg_range": [60, 95],
|
90 |
-
# the second degradation process
|
91 |
-
"second_blur_prob": 0.5,
|
92 |
-
"resize_prob2": [0.3, 0.4, 0.3], # up, down, keep
|
93 |
-
"resize_range2": [0.6, 1.2],
|
94 |
-
"gaussian_noise_prob2": 0.5,
|
95 |
-
"noise_range2": [1, 12],
|
96 |
-
"poisson_scale_range2": [0.05, 1.0],
|
97 |
-
"gray_noise_prob2": 0.4,
|
98 |
-
"jpeg_range2": [60, 95],
|
99 |
-
"queue_size": 180,
|
100 |
-
"scale": 4, # output size: ori_h // scale
|
101 |
-
"sharpen": False,
|
102 |
-
}
|
103 |
-
|
104 |
-
|
105 |
-
def set_para(para_dic):
|
106 |
-
# blur settings for the first degradation
|
107 |
-
# blur_kernel_size = opt['blur_kernel_size']
|
108 |
-
kernel_list = para_dic["kernel_list"]
|
109 |
-
kernel_prob = para_dic["kernel_prob"]
|
110 |
-
blur_sigma = para_dic["blur_sigma"]
|
111 |
-
betag_range = para_dic["betag_range"]
|
112 |
-
betap_range = para_dic["betap_range"]
|
113 |
-
sinc_prob = para_dic["sinc_prob"]
|
114 |
-
|
115 |
-
# blur settings for the second degradation
|
116 |
-
# blur_kernel_size2 = opt['blur_kernel_size2']
|
117 |
-
kernel_list2 = para_dic["kernel_list2"]
|
118 |
-
kernel_prob2 = para_dic["kernel_prob2"]
|
119 |
-
blur_sigma2 = para_dic["blur_sigma2"]
|
120 |
-
betag_range2 = para_dic["betag_range2"]
|
121 |
-
betap_range2 = para_dic["betap_range2"]
|
122 |
-
sinc_prob2 = para_dic["sinc_prob2"]
|
123 |
-
|
124 |
-
# a final sinc filter
|
125 |
-
final_sinc_prob = para_dic["final_sinc_prob"]
|
126 |
-
|
127 |
-
kernel_range = [2 * v + 1 for v in range(3, 11)] # kernel size ranges from 7 to 21
|
128 |
-
pulse_tensor = torch.zeros(
|
129 |
-
21, 21
|
130 |
-
).float() # convolving with pulse tensor brings no blurry effect
|
131 |
-
pulse_tensor[10, 10] = 1
|
132 |
-
kernel_size = random.choice(kernel_range)
|
133 |
-
if np.random.uniform() < sinc_prob:
|
134 |
-
# this sinc filter setting is for kernels ranging from [7, 21]
|
135 |
-
if kernel_size < 13:
|
136 |
-
omega_c = np.random.uniform(np.pi / 3, np.pi)
|
137 |
-
else:
|
138 |
-
omega_c = np.random.uniform(np.pi / 5, np.pi)
|
139 |
-
kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)
|
140 |
-
else:
|
141 |
-
kernel = random_mixed_kernels(
|
142 |
-
kernel_list,
|
143 |
-
kernel_prob,
|
144 |
-
kernel_size,
|
145 |
-
blur_sigma,
|
146 |
-
blur_sigma,
|
147 |
-
[-math.pi, math.pi],
|
148 |
-
betag_range,
|
149 |
-
betap_range,
|
150 |
-
noise_range=None,
|
151 |
-
)
|
152 |
-
# pad kernel
|
153 |
-
pad_size = (21 - kernel_size) // 2
|
154 |
-
kernel = np.pad(kernel, ((pad_size, pad_size), (pad_size, pad_size)))
|
155 |
-
|
156 |
-
# ------------------------ Generate kernels (used in the second degradation) -------------- #
|
157 |
-
kernel_size = random.choice(kernel_range)
|
158 |
-
if np.random.uniform() < sinc_prob2:
|
159 |
-
if kernel_size < 13:
|
160 |
-
omega_c = np.random.uniform(np.pi / 3, np.pi)
|
161 |
-
else:
|
162 |
-
omega_c = np.random.uniform(np.pi / 5, np.pi)
|
163 |
-
kernel2 = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)
|
164 |
-
else:
|
165 |
-
kernel2 = random_mixed_kernels(
|
166 |
-
kernel_list2,
|
167 |
-
kernel_prob2,
|
168 |
-
kernel_size,
|
169 |
-
blur_sigma2,
|
170 |
-
blur_sigma2,
|
171 |
-
[-math.pi, math.pi],
|
172 |
-
betag_range2,
|
173 |
-
betap_range2,
|
174 |
-
noise_range=None,
|
175 |
-
)
|
176 |
-
pad_size = (21 - kernel_size) // 2
|
177 |
-
kernel2 = np.pad(kernel2, ((pad_size, pad_size), (pad_size, pad_size)))
|
178 |
-
|
179 |
-
# ------------------------------------- sinc kernel ------------------------------------- #
|
180 |
-
if np.random.uniform() < final_sinc_prob:
|
181 |
-
kernel_size = random.choice(kernel_range)
|
182 |
-
omega_c = np.random.uniform(np.pi / 3, np.pi)
|
183 |
-
sinc_kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=21)
|
184 |
-
sinc_kernel = torch.FloatTensor(sinc_kernel)
|
185 |
-
else:
|
186 |
-
sinc_kernel = pulse_tensor
|
187 |
-
kernel = torch.FloatTensor(kernel)
|
188 |
-
kernel2 = torch.FloatTensor(kernel2)
|
189 |
-
return_d = {"kernel1": kernel, "kernel2": kernel2, "sinc_kernel": sinc_kernel}
|
190 |
-
return return_d
|
191 |
-
|
192 |
-
|
193 |
-
def print_stat(a):
|
194 |
-
print(
|
195 |
-
f"shape={a.shape}, min={a.min():.2f}, \
|
196 |
-
max={a.max():.2f}, var={a.var():.2f}, {a.flatten()[0]}"
|
197 |
-
)
|
198 |
-
|
199 |
-
|
200 |
-
@torch.no_grad()
|
201 |
-
def esr_blur_gpu(image, paras, usm_sharpener, jpeger, device="cpu"):
|
202 |
-
"""
|
203 |
-
input and output: image is a tensor with shape: b f c h w, range (-1, 1)
|
204 |
-
"""
|
205 |
-
video_length = image.shape[1]
|
206 |
-
image = rearrange(image, "b f c h w -> (b f) c h w").to(device)
|
207 |
-
image = (image + 1) * 0.5
|
208 |
-
if degrade_dic["sharpen"]:
|
209 |
-
gt_usm = usm_sharpener(image)
|
210 |
-
else:
|
211 |
-
gt_usm = image
|
212 |
-
ori_h, ori_w = image.size()[2:4]
|
213 |
-
# ----------------------- The first degradation process ----------------------- #
|
214 |
-
# blur
|
215 |
-
out = filter2D(gt_usm, paras["kernel1"].unsqueeze(0).to(device))
|
216 |
-
# random resize
|
217 |
-
updown_type = random.choices(["up", "down", "keep"], degrade_dic["resize_prob"])[0]
|
218 |
-
if updown_type == "up":
|
219 |
-
scale = np.random.uniform(1, degrade_dic["resize_range"][1])
|
220 |
-
elif updown_type == "down":
|
221 |
-
scale = np.random.uniform(degrade_dic["resize_range"][0], 1)
|
222 |
-
else:
|
223 |
-
scale = 1
|
224 |
-
mode = random.choice(["area", "bilinear", "bicubic"])
|
225 |
-
out = F.interpolate(out, scale_factor=scale, mode=mode)
|
226 |
-
# noise
|
227 |
-
gray_noise_prob = degrade_dic["gray_noise_prob"]
|
228 |
-
out = out.to(torch.float32)
|
229 |
-
if np.random.uniform() < degrade_dic["gaussian_noise_prob"]:
|
230 |
-
out = random_add_gaussian_noise_pt(
|
231 |
-
out,
|
232 |
-
# video_length=video_length,
|
233 |
-
sigma_range=degrade_dic["noise_range"],
|
234 |
-
clip=True,
|
235 |
-
rounds=False,
|
236 |
-
gray_prob=gray_noise_prob,
|
237 |
-
)
|
238 |
-
else:
|
239 |
-
out = random_add_poisson_noise_pt(
|
240 |
-
out,
|
241 |
-
# video_length=video_length,
|
242 |
-
scale_range=degrade_dic["poisson_scale_range"],
|
243 |
-
gray_prob=gray_noise_prob,
|
244 |
-
clip=True,
|
245 |
-
rounds=False,
|
246 |
-
)
|
247 |
-
# out = out.to(torch.bfloat16)
|
248 |
-
|
249 |
-
# JPEG compression
|
250 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*degrade_dic["jpeg_range"])
|
251 |
-
out = torch.clamp(out, 0, 1)
|
252 |
-
out = jpeger(out, quality=jpeg_p)
|
253 |
-
|
254 |
-
# Video compression 1
|
255 |
-
# print('Video compression 1')
|
256 |
-
|
257 |
-
# print_stat(out)
|
258 |
-
if video_length > 1:
|
259 |
-
out = video_compression(out, device=device)
|
260 |
-
# print('After video compression 1')
|
261 |
-
|
262 |
-
# print_stat(out)
|
263 |
-
|
264 |
-
# ----------------------- The second degradation process ----------------------- #
|
265 |
-
# blur
|
266 |
-
if np.random.uniform() < degrade_dic["second_blur_prob"]:
|
267 |
-
out = filter2D(out, paras["kernel2"].unsqueeze(0).to(device))
|
268 |
-
# random resize
|
269 |
-
updown_type = random.choices(["up", "down", "keep"], degrade_dic["resize_prob2"])[0]
|
270 |
-
if updown_type == "up":
|
271 |
-
scale = np.random.uniform(1, degrade_dic["resize_range2"][1])
|
272 |
-
elif updown_type == "down":
|
273 |
-
scale = np.random.uniform(degrade_dic["resize_range2"][0], 1)
|
274 |
-
else:
|
275 |
-
scale = 1
|
276 |
-
mode = random.choice(["area", "bilinear", "bicubic"])
|
277 |
-
out = F.interpolate(
|
278 |
-
out,
|
279 |
-
size=(
|
280 |
-
int(ori_h / degrade_dic["scale"] * scale),
|
281 |
-
int(ori_w / degrade_dic["scale"] * scale),
|
282 |
-
),
|
283 |
-
mode=mode,
|
284 |
-
)
|
285 |
-
# noise
|
286 |
-
gray_noise_prob = degrade_dic["gray_noise_prob2"]
|
287 |
-
out = out.to(torch.float32)
|
288 |
-
if np.random.uniform() < degrade_dic["gaussian_noise_prob2"]:
|
289 |
-
out = random_add_gaussian_noise_pt(
|
290 |
-
out,
|
291 |
-
# video_length=video_length,
|
292 |
-
sigma_range=degrade_dic["noise_range2"],
|
293 |
-
clip=True,
|
294 |
-
rounds=False,
|
295 |
-
gray_prob=gray_noise_prob,
|
296 |
-
)
|
297 |
-
else:
|
298 |
-
out = random_add_poisson_noise_pt(
|
299 |
-
out,
|
300 |
-
# video_length=video_length,
|
301 |
-
scale_range=degrade_dic["poisson_scale_range2"],
|
302 |
-
gray_prob=gray_noise_prob,
|
303 |
-
clip=True,
|
304 |
-
rounds=False,
|
305 |
-
)
|
306 |
-
# out = out.to(torch.bfloat16)
|
307 |
-
|
308 |
-
if np.random.uniform() < 0.5:
|
309 |
-
# resize back + the final sinc filter
|
310 |
-
mode = random.choice(["area", "bilinear", "bicubic"])
|
311 |
-
out = F.interpolate(
|
312 |
-
out, size=(ori_h // degrade_dic["scale"], ori_w // degrade_dic["scale"]), mode=mode
|
313 |
-
)
|
314 |
-
out = filter2D(out, paras["sinc_kernel"].unsqueeze(0).to(device))
|
315 |
-
# JPEG compression
|
316 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*degrade_dic["jpeg_range2"])
|
317 |
-
out = torch.clamp(out, 0, 1)
|
318 |
-
out = jpeger(out, quality=jpeg_p)
|
319 |
-
else:
|
320 |
-
# JPEG compression
|
321 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*degrade_dic["jpeg_range2"])
|
322 |
-
out = torch.clamp(out, 0, 1)
|
323 |
-
out = jpeger(out, quality=jpeg_p)
|
324 |
-
# resize back + the final sinc filter
|
325 |
-
mode = random.choice(["area", "bilinear", "bicubic"])
|
326 |
-
out = F.interpolate(
|
327 |
-
out, size=(ori_h // degrade_dic["scale"], ori_w // degrade_dic["scale"]), mode=mode
|
328 |
-
)
|
329 |
-
out = filter2D(out, paras["sinc_kernel"].unsqueeze(0).to(device))
|
330 |
-
|
331 |
-
# print('Video compression 2')
|
332 |
-
|
333 |
-
# print_stat(out)
|
334 |
-
if video_length > 1:
|
335 |
-
out = video_compression(out, device=device)
|
336 |
-
# print('After video compression 2')
|
337 |
-
|
338 |
-
# print_stat(out)
|
339 |
-
|
340 |
-
out = F.interpolate(out, size=(ori_h, ori_w), mode="bicubic")
|
341 |
-
blur_image = torch.clamp(out, 0, 1)
|
342 |
-
# blur_image = ColorJitter(0.1, 0.1, 0.1, 0.05)(blur_image) # 颜色数据增广
|
343 |
-
# (-1, 1)
|
344 |
-
blur_image = 2.0 * blur_image - 1
|
345 |
-
blur_image = rearrange(blur_image, "(b f) c h w->b f c h w", f=video_length)
|
346 |
-
return blur_image
|
347 |
-
|
348 |
-
|
349 |
-
def video_compression(video_in, device="cpu"):
|
350 |
-
# Shape: (t, c, h, w); channel order: RGB; image range: [0, 1], float32.
|
351 |
-
|
352 |
-
video_in = torch.clamp(video_in, 0, 1)
|
353 |
-
params = dict(
|
354 |
-
codec=["libx264", "h264", "mpeg4"],
|
355 |
-
codec_prob=[1 / 3.0, 1 / 3.0, 1 / 3.0],
|
356 |
-
bitrate=[1e4, 1e5],
|
357 |
-
) # 1e4, 1e5
|
358 |
-
codec = random.choices(params["codec"], params["codec_prob"])[0]
|
359 |
-
# print(f"use codec {codec}")
|
360 |
-
|
361 |
-
bitrate = params["bitrate"]
|
362 |
-
bitrate = np.random.randint(bitrate[0], bitrate[1] + 1)
|
363 |
-
|
364 |
-
h, w = video_in.shape[-2:]
|
365 |
-
video_in = F.interpolate(video_in, (h // 2 * 2, w // 2 * 2), mode="bilinear")
|
366 |
-
|
367 |
-
buf = io.BytesIO()
|
368 |
-
with av.open(buf, "w", "mp4") as container:
|
369 |
-
stream = container.add_stream(codec, rate=1)
|
370 |
-
stream.height = video_in.shape[-2]
|
371 |
-
stream.width = video_in.shape[-1]
|
372 |
-
stream.pix_fmt = "yuv420p"
|
373 |
-
stream.bit_rate = bitrate
|
374 |
-
|
375 |
-
for img in video_in: # img: C H W; 0-1
|
376 |
-
img_np = img.permute(1, 2, 0).contiguous() * 255.0
|
377 |
-
# 1 reference_np = reference.detach(). to (torch.float) .cpu() .numpy ()
|
378 |
-
img_np = img_np.detach().to(torch.float).cpu().numpy().astype(np.uint8)
|
379 |
-
frame = av.VideoFrame.from_ndarray(img_np, format="rgb24")
|
380 |
-
frame.pict_type = "NONE"
|
381 |
-
for packet in stream.encode(frame):
|
382 |
-
container.mux(packet)
|
383 |
-
|
384 |
-
# Flush stream
|
385 |
-
for packet in stream.encode():
|
386 |
-
container.mux(packet)
|
387 |
-
|
388 |
-
outputs = []
|
389 |
-
with av.open(buf, "r", "mp4") as container:
|
390 |
-
if container.streams.video:
|
391 |
-
for frame in container.decode(**{"video": 0}):
|
392 |
-
outputs.append(frame.to_rgb().to_ndarray().astype(np.float32))
|
393 |
-
|
394 |
-
video_in = torch.Tensor(np.array(outputs)).permute(0, 3, 1, 2).contiguous() # T C H W
|
395 |
-
video_in = torch.clamp(video_in / 255.0, 0, 1).to(device) # 0-1
|
396 |
-
return video_in
|
397 |
-
|
398 |
-
|
399 |
-
@torch.no_grad()
|
400 |
-
def my_esr_blur(images, device="cpu"):
|
401 |
-
"""
|
402 |
-
images is a list of tensor with shape: b f c h w, range (-1, 1)
|
403 |
-
"""
|
404 |
-
jpeger = DiffJPEG(differentiable=False).to(device)
|
405 |
-
usm_sharpener = USMSharp()
|
406 |
-
if degrade_dic["sharpen"]:
|
407 |
-
usm_sharpener = usm_sharpener.to(device)
|
408 |
-
paras = set_para(para_dic)
|
409 |
-
blur_image = [
|
410 |
-
esr_blur_gpu(image, paras, usm_sharpener, jpeger, device=device) for image in images
|
411 |
-
]
|
412 |
-
|
413 |
-
return blur_image
|
414 |
-
|
415 |
-
|
416 |
-
para_dic_latent = {
|
417 |
-
"kernel_list": [
|
418 |
-
"iso",
|
419 |
-
"aniso",
|
420 |
-
"generalized_iso",
|
421 |
-
"generalized_aniso",
|
422 |
-
"plateau_iso",
|
423 |
-
"plateau_aniso",
|
424 |
-
],
|
425 |
-
"kernel_prob": [0.45, 0.25, 0.12, 0.03, 0.12, 0.03],
|
426 |
-
"sinc_prob": 0.1,
|
427 |
-
"blur_sigma": [0.2, 1.5],
|
428 |
-
"betag_range": [0.5, 2.0],
|
429 |
-
"betap_range": [1, 1.5],
|
430 |
-
"kernel_list2": [
|
431 |
-
"iso",
|
432 |
-
"aniso",
|
433 |
-
"generalized_iso",
|
434 |
-
"generalized_aniso",
|
435 |
-
"plateau_iso",
|
436 |
-
"plateau_aniso",
|
437 |
-
],
|
438 |
-
"kernel_prob2": [0.45, 0.25, 0.12, 0.03, 0.12, 0.03],
|
439 |
-
"sinc_prob2": 0.1,
|
440 |
-
"blur_sigma2": [0.2, 1.0],
|
441 |
-
"betag_range2": [0.5, 2.0],
|
442 |
-
"betap_range2": [1, 1.5],
|
443 |
-
"final_sinc_prob": 0.5,
|
444 |
-
}
|
445 |
-
|
446 |
-
|
447 |
-
def set_para_latent(para_dic):
|
448 |
-
# blur settings for the first degradation
|
449 |
-
# blur_kernel_size = opt['blur_kernel_size']
|
450 |
-
kernel_list = para_dic["kernel_list"]
|
451 |
-
kernel_prob = para_dic["kernel_prob"]
|
452 |
-
blur_sigma = para_dic["blur_sigma"]
|
453 |
-
betag_range = para_dic["betag_range"]
|
454 |
-
betap_range = para_dic["betap_range"]
|
455 |
-
sinc_prob = para_dic["sinc_prob"]
|
456 |
-
|
457 |
-
# a final sinc filter
|
458 |
-
|
459 |
-
kernel_range = [2 * v + 1 for v in range(1, 11)] # kernel size ranges from 7 to 21
|
460 |
-
pulse_tensor = torch.zeros(
|
461 |
-
21, 21
|
462 |
-
).float() # convolving with pulse tensor brings no blurry effect
|
463 |
-
pulse_tensor[10, 10] = 1
|
464 |
-
kernel_size = random.choice(kernel_range)
|
465 |
-
if np.random.uniform() < sinc_prob:
|
466 |
-
# this sinc filter setting is for kernels ranging from [7, 21]
|
467 |
-
if kernel_size < 13:
|
468 |
-
omega_c = np.random.uniform(np.pi / 3, np.pi)
|
469 |
-
else:
|
470 |
-
omega_c = np.random.uniform(np.pi / 5, np.pi)
|
471 |
-
kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)
|
472 |
-
else:
|
473 |
-
kernel = random_mixed_kernels(
|
474 |
-
kernel_list,
|
475 |
-
kernel_prob,
|
476 |
-
kernel_size,
|
477 |
-
blur_sigma,
|
478 |
-
blur_sigma,
|
479 |
-
[-math.pi, math.pi],
|
480 |
-
betag_range,
|
481 |
-
betap_range,
|
482 |
-
noise_range=None,
|
483 |
-
)
|
484 |
-
# pad kernel
|
485 |
-
pad_size = (21 - kernel_size) // 2
|
486 |
-
kernel = np.pad(kernel, ((pad_size, pad_size), (pad_size, pad_size)))
|
487 |
-
kernel = torch.FloatTensor(kernel)
|
488 |
-
return_d = {"kernel1": kernel}
|
489 |
-
return return_d
|
490 |
-
|
491 |
-
|
492 |
-
@torch.no_grad()
|
493 |
-
def latent_blur_gpu(image, paras, device="cpu"):
|
494 |
-
"""
|
495 |
-
input and output: image is a tensor with shape: b f c h w, range (-1, 1)
|
496 |
-
"""
|
497 |
-
video_length = image.shape[1]
|
498 |
-
image = rearrange(image, "b f c h w -> (b f) c h w").to(device)
|
499 |
-
image = (image + 1) * 0.5
|
500 |
-
gt_usm = image
|
501 |
-
ori_h, ori_w = image.size()[2:4]
|
502 |
-
# ----------------------- The first degradation process ----------------------- #
|
503 |
-
# blur
|
504 |
-
out = filter2D(gt_usm, paras["kernel1"].unsqueeze(0).to(device))
|
505 |
-
blur_image = torch.clamp(out, 0, 1)
|
506 |
-
# blur_image = ColorJitter(0.1, 0.1, 0.1, 0.05)(blur_image) # 颜色数据增广
|
507 |
-
# (-1, 1)
|
508 |
-
blur_image = 2.0 * blur_image - 1
|
509 |
-
blur_image = rearrange(blur_image, "(b f) c h w->b f c h w", f=video_length)
|
510 |
-
return blur_image
|
511 |
-
|
512 |
-
|
513 |
-
@torch.no_grad()
|
514 |
-
def add_latent_blur(images, device="cpu"):
|
515 |
-
"""
|
516 |
-
images is a list of tensor with shape: b f c h w, range (-1, 1)
|
517 |
-
"""
|
518 |
-
paras = set_para_latent(para_dic_latent)
|
519 |
-
blur_image = [latent_blur_gpu(image, paras, device=device) for image in images]
|
520 |
-
print("apply blur to the latents")
|
521 |
-
|
522 |
-
return blur_image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
projects/video_diffusion_sr/infer.py
CHANGED
@@ -37,31 +37,11 @@ from common.distributed.meta_init_utils import (
|
|
37 |
# from common.fs import download
|
38 |
|
39 |
from models.dit_v2 import na
|
40 |
-
from projects.video_diffusion_sr.degradation_utils import my_esr_blur
|
41 |
-
|
42 |
|
43 |
class VideoDiffusionInfer():
|
44 |
def __init__(self, config: DictConfig):
|
45 |
self.config = config
|
46 |
|
47 |
-
@log_on_entry
|
48 |
-
def configure_blur(self):
|
49 |
-
# Create degradation.
|
50 |
-
def _blur_fn(x: List[torch.Tensor]):
|
51 |
-
if x[0].ndim == 4:
|
52 |
-
x = my_esr_blur(
|
53 |
-
[rearrange(i, "c f h w -> 1 f c h w") for i in x], device=get_device()
|
54 |
-
)
|
55 |
-
x = [rearrange(i, "1 f c h w -> c f h w") for i in x]
|
56 |
-
else:
|
57 |
-
x = my_esr_blur(
|
58 |
-
[rearrange(i, "c h w -> 1 1 c h w") for i in x], device=get_device()
|
59 |
-
)
|
60 |
-
x = [i[0, 0] for i in x]
|
61 |
-
return x
|
62 |
-
|
63 |
-
self.my_esr_blur = _blur_fn
|
64 |
-
|
65 |
def get_condition(self, latent: Tensor, latent_blur: Tensor, task: str) -> Tensor:
|
66 |
t, h, w, c = latent.shape
|
67 |
cond = torch.zeros([t, h, w, c + 1], device=latent.device, dtype=latent.dtype)
|
|
|
37 |
# from common.fs import download
|
38 |
|
39 |
from models.dit_v2 import na
|
|
|
|
|
40 |
|
41 |
class VideoDiffusionInfer():
|
42 |
def __init__(self, config: DictConfig):
|
43 |
self.config = config
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
def get_condition(self, latent: Tensor, latent_blur: Tensor, task: str) -> Tensor:
|
46 |
t, h, w, c = latent.shape
|
47 |
cond = torch.zeros([t, h, w, c + 1], device=latent.device, dtype=latent.dtype)
|