schirrmacher
commited on
Commit
•
58ed4f8
1
Parent(s):
54e7ec8
Upload folder using huggingface_hub
Browse files- README.md +8 -15
- final_image_example.png +3 -0
- ground_truth_example.png +3 -0
- layer_diffuse_example.png +3 -0
- util/ic-light.py +162 -85
README.md
CHANGED
@@ -11,28 +11,21 @@ pretty_name: Human Segmentation Dataset
|
|
11 |
|
12 |
This dataset was created **for developing the best fully open-source background remover** of images with humans. It was crafted with [LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse), a Stable Diffusion extension for generating transparent images. After creating segmented humans, [IC-Light](https://github.com/lllyasviel/IC-Light) was used for embedding them into realistic scenarios.
|
13 |
|
14 |
-
The dataset covers a diverse set of segmented humans: various skin tones, clothes, hair styles etc. Since Stable Diffusion is not perfect, the dataset contains images with flaws. Still the dataset is good enough for training background remover models.
|
15 |
|
16 |
-
|
17 |
-
|
18 |
-
I created more than 7.000 images with people and diverse backgrounds.
|
19 |
-
|
20 |
-
# Create Training Dataset
|
21 |
|
22 |
-
|
23 |
|
24 |
-
|
25 |
|
26 |
-
|
27 |
-
./create_dataset.sh
|
28 |
-
```
|
29 |
|
30 |
-
|
31 |
|
32 |
-
|
33 |
|
34 |
-
![](
|
35 |
-
![](example_ground_truth.png)
|
36 |
|
37 |
# Support
|
38 |
|
|
|
11 |
|
12 |
This dataset was created **for developing the best fully open-source background remover** of images with humans. It was crafted with [LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse), a Stable Diffusion extension for generating transparent images. After creating segmented humans, [IC-Light](https://github.com/lllyasviel/IC-Light) was used for embedding them into realistic scenarios.
|
13 |
|
14 |
+
The dataset covers a diverse set of segmented humans: various skin tones, clothes, hair styles etc. Since Stable Diffusion is not perfect, the dataset contains images with flaws. Still the dataset is good enough for training background remover models. I created more than 7.000 images with people and diverse backgrounds.
|
15 |
|
16 |
+
# Examples
|
|
|
|
|
|
|
|
|
17 |
|
18 |
+
[LayerDiffuse](https://github.com/layerdiffusion/LayerDiffuse) output:
|
19 |
|
20 |
+
![](layer_diffuse_example.png)
|
21 |
|
22 |
+
[IC-Light](https://github.com/lllyasviel/IC-Light) applied to segmented image:
|
|
|
|
|
23 |
|
24 |
+
![](final_image_example.png)
|
25 |
|
26 |
+
Ground truth:
|
27 |
|
28 |
+
![](ground_truth_example.png)
|
|
|
29 |
|
30 |
# Support
|
31 |
|
final_image_example.png
ADDED
Git LFS Details
|
ground_truth_example.png
ADDED
Git LFS Details
|
layer_diffuse_example.png
ADDED
Git LFS Details
|
util/ic-light.py
CHANGED
@@ -10,27 +10,46 @@ import cv2
|
|
10 |
from diffusers.utils import load_image
|
11 |
|
12 |
from PIL import Image, ImageFilter, ImageOps
|
13 |
-
from diffusers import
|
14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
16 |
from transformers import CLIPTextModel, CLIPTokenizer
|
17 |
from enum import Enum
|
|
|
18 |
# from torch.hub import download_url_to_file
|
19 |
|
20 |
|
21 |
# 'stablediffusionapi/realistic-vision-v51'
|
22 |
# 'runwayml/stable-diffusion-v1-5'
|
23 |
-
sd15_name =
|
24 |
tokenizer = CLIPTokenizer.from_pretrained(sd15_name, subfolder="tokenizer")
|
25 |
text_encoder = CLIPTextModel.from_pretrained(sd15_name, subfolder="text_encoder")
|
26 |
vae = AutoencoderKL.from_pretrained(sd15_name, subfolder="vae")
|
27 |
unet = UNet2DConditionModel.from_pretrained(sd15_name, subfolder="unet")
|
28 |
-
upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(
|
|
|
|
|
29 |
|
30 |
# Change UNet
|
31 |
|
32 |
with torch.no_grad():
|
33 |
-
new_conv_in = torch.nn.Conv2d(
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
new_conv_in.weight.zero_()
|
35 |
new_conv_in.weight[:, :4, :, :].copy_(unet.conv_in.weight)
|
36 |
new_conv_in.bias = unet.conv_in.bias
|
@@ -40,10 +59,10 @@ unet_original_forward = unet.forward
|
|
40 |
|
41 |
|
42 |
def hooked_unet_forward(sample, timestep, encoder_hidden_states, **kwargs):
|
43 |
-
c_concat = kwargs[
|
44 |
c_concat = torch.cat([c_concat] * (sample.shape[0] // c_concat.shape[0]), dim=0)
|
45 |
new_sample = torch.cat([sample, c_concat], dim=1)
|
46 |
-
kwargs[
|
47 |
return unet_original_forward(new_sample, timestep, encoder_hidden_states, **kwargs)
|
48 |
|
49 |
|
@@ -51,7 +70,7 @@ unet.forward = hooked_unet_forward
|
|
51 |
|
52 |
# Load
|
53 |
|
54 |
-
model_path =
|
55 |
# download_url_to_file(url='https://huggingface.co/lllyasviel/ic-light/resolve/main/iclight_sd15_fc.safetensors', dst=model_path)
|
56 |
sd_offset = sf.load_file(model_path)
|
57 |
sd_origin = unet.state_dict()
|
@@ -62,7 +81,7 @@ del sd_offset, sd_origin, sd_merged, keys
|
|
62 |
|
63 |
# Device
|
64 |
|
65 |
-
device = torch.device(
|
66 |
text_encoder = text_encoder.to(device=device, dtype=torch.float16)
|
67 |
vae = vae.to(device=device, dtype=torch.bfloat16)
|
68 |
unet = unet.to(device=device, dtype=torch.float16)
|
@@ -85,10 +104,7 @@ ddim_scheduler = DDIMScheduler(
|
|
85 |
)
|
86 |
|
87 |
euler_a_scheduler = EulerAncestralDiscreteScheduler(
|
88 |
-
num_train_timesteps=1000,
|
89 |
-
beta_start=0.00085,
|
90 |
-
beta_end=0.012,
|
91 |
-
steps_offset=1
|
92 |
)
|
93 |
|
94 |
dpmpp_2m_sde_karras_scheduler = DPMSolverMultistepScheduler(
|
@@ -97,7 +113,7 @@ dpmpp_2m_sde_karras_scheduler = DPMSolverMultistepScheduler(
|
|
97 |
beta_end=0.012,
|
98 |
algorithm_type="sde-dpmsolver++",
|
99 |
use_karras_sigmas=True,
|
100 |
-
steps_offset=1
|
101 |
)
|
102 |
|
103 |
# Pipelines
|
@@ -111,7 +127,7 @@ t2i_pipe = StableDiffusionPipeline(
|
|
111 |
safety_checker=None,
|
112 |
requires_safety_checker=False,
|
113 |
feature_extractor=None,
|
114 |
-
image_encoder=None
|
115 |
)
|
116 |
|
117 |
i2i_pipe = StableDiffusionImg2ImgPipeline(
|
@@ -123,7 +139,7 @@ i2i_pipe = StableDiffusionImg2ImgPipeline(
|
|
123 |
safety_checker=None,
|
124 |
requires_safety_checker=False,
|
125 |
feature_extractor=None,
|
126 |
-
image_encoder=None
|
127 |
)
|
128 |
|
129 |
|
@@ -139,7 +155,10 @@ def encode_prompt_inner(txt: str):
|
|
139 |
return x[:i] if len(x) >= i else x + [p] * (i - len(x))
|
140 |
|
141 |
tokens = tokenizer(txt, truncation=False, add_special_tokens=False)["input_ids"]
|
142 |
-
chunks = [
|
|
|
|
|
|
|
143 |
chunks = [pad(ck, id_pad, max_length) for ck in chunks]
|
144 |
|
145 |
token_ids = torch.tensor(chunks).to(device=device, dtype=torch.int64)
|
@@ -188,7 +207,9 @@ def pytorch2numpy(imgs, quant=True):
|
|
188 |
|
189 |
@torch.inference_mode()
|
190 |
def numpy2pytorch(imgs):
|
191 |
-
h =
|
|
|
|
|
192 |
h = h.movedim(-1, 1)
|
193 |
return h
|
194 |
|
@@ -213,14 +234,31 @@ def resize_without_crop(image, target_width, target_height):
|
|
213 |
resized_image = pil_image.resize((target_width, target_height), Image.LANCZOS)
|
214 |
return np.array(resized_image)
|
215 |
|
|
|
216 |
def remove_alpha_threshold(image, alpha_threshold=160):
|
217 |
# This function removes artifacts created by LayerDiffusion
|
218 |
mask = image[:, :, 3] < alpha_threshold
|
219 |
image[mask] = [0, 0, 0, 0]
|
220 |
return image
|
221 |
|
|
|
222 |
@torch.inference_mode()
|
223 |
-
def process(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
bg_source = BGSource(bg_source)
|
225 |
input_bg = None
|
226 |
|
@@ -243,56 +281,69 @@ def process(input_fg, prompt, image_width, image_height, num_samples, seed, step
|
|
243 |
image = np.tile(gradient, (1, image_width))
|
244 |
input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
|
245 |
else:
|
246 |
-
raise
|
247 |
|
248 |
rng = torch.Generator(device=device).manual_seed(int(seed))
|
249 |
|
250 |
fg = resize_and_center_crop(input_fg, image_width, image_height)
|
251 |
|
252 |
concat_conds = numpy2pytorch([fg]).to(device=vae.device, dtype=vae.dtype)
|
253 |
-
concat_conds =
|
|
|
|
|
254 |
|
255 |
-
conds, unconds = encode_prompt_pair(
|
|
|
|
|
256 |
|
257 |
if input_bg is None:
|
258 |
-
latents =
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
|
|
|
|
|
|
270 |
else:
|
271 |
bg = resize_and_center_crop(input_bg, image_width, image_height)
|
272 |
bg_latent = numpy2pytorch([bg]).to(device=vae.device, dtype=vae.dtype)
|
273 |
bg_latent = vae.encode(bg_latent).latent_dist.mode() * vae.config.scaling_factor
|
274 |
-
latents =
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
|
|
|
|
|
|
288 |
|
289 |
pixels = vae.decode(latents).sample
|
290 |
pixels = pytorch2numpy(pixels)
|
291 |
-
pixels = [
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
|
|
|
|
|
|
296 |
|
297 |
pixels = numpy2pytorch(pixels).to(device=vae.device, dtype=vae.dtype)
|
298 |
latents = vae.encode(pixels).latent_dist.mode() * vae.config.scaling_factor
|
@@ -302,22 +353,27 @@ def process(input_fg, prompt, image_width, image_height, num_samples, seed, step
|
|
302 |
|
303 |
fg = resize_and_center_crop(input_fg, image_width, image_height)
|
304 |
concat_conds = numpy2pytorch([fg]).to(device=vae.device, dtype=vae.dtype)
|
305 |
-
concat_conds =
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
321 |
|
322 |
pixels = vae.decode(latents).sample
|
323 |
|
@@ -335,16 +391,18 @@ def augment(image):
|
|
335 |
else:
|
336 |
target_height, target_width = 512 * 2, 640 * 2
|
337 |
|
338 |
-
left_right_padding = (
|
|
|
|
|
339 |
|
340 |
original = cv2.copyMakeBorder(
|
341 |
-
original,
|
342 |
-
top=max(target_height, image_height) - min(target_height, image_height),
|
343 |
bottom=0,
|
344 |
-
left=left_right_padding,
|
345 |
-
right=left_right_padding,
|
346 |
-
borderType=cv2.BORDER_CONSTANT,
|
347 |
-
value=(0, 0, 0)
|
348 |
)
|
349 |
|
350 |
transform = A.Compose(
|
@@ -363,6 +421,7 @@ def augment(image):
|
|
363 |
|
364 |
return transform(image=original)["image"]
|
365 |
|
|
|
366 |
class BGSource(Enum):
|
367 |
NONE = "None"
|
368 |
LEFT = "Left Light"
|
@@ -380,8 +439,7 @@ prompts = [
|
|
380 |
"sunshine, cafe, chilled",
|
381 |
"exhibition, paintings",
|
382 |
"beach",
|
383 |
-
"winter, snow"
|
384 |
-
"forrest, cloudy",
|
385 |
"party, people",
|
386 |
"cozy living room, sofa, shelf",
|
387 |
"mountains",
|
@@ -392,7 +450,7 @@ prompts = [
|
|
392 |
"appartment, soft light",
|
393 |
"garden",
|
394 |
"school",
|
395 |
-
"art exhibition with paintings in background"
|
396 |
]
|
397 |
|
398 |
os.makedirs(ground_truth_dir, exist_ok=True)
|
@@ -402,7 +460,9 @@ all_images = os.listdir(input_dir)
|
|
402 |
random.shuffle(all_images)
|
403 |
|
404 |
for filename in all_images:
|
405 |
-
if filename.lower().endswith(
|
|
|
|
|
406 |
|
407 |
letters = string.ascii_lowercase
|
408 |
random_string = "".join(random.choice(letters) for i in range(13))
|
@@ -418,7 +478,9 @@ for filename in all_images:
|
|
418 |
image = np.array(image)
|
419 |
|
420 |
image_augmented = augment(image)
|
421 |
-
Image.fromarray(image_augmented).getchannel("A").save(
|
|
|
|
|
422 |
|
423 |
image_augmented = image_augmented[:, :, :3]
|
424 |
|
@@ -427,7 +489,7 @@ for filename in all_images:
|
|
427 |
image_height, image_width, _ = image_augmented.shape
|
428 |
|
429 |
num_samples = 1
|
430 |
-
seed = random.randint(1,123456789012345678901234567890)
|
431 |
steps = 25
|
432 |
constant_prompt = "details, high quality"
|
433 |
prompt = random.choice(prompts)
|
@@ -437,7 +499,22 @@ for filename in all_images:
|
|
437 |
highres_denoise = 0.7
|
438 |
lowres_denoise = 0.5
|
439 |
bg_source = BGSource.NONE
|
440 |
-
|
441 |
-
results = process(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
442 |
result_image = Image.fromarray(results[0])
|
443 |
-
result_image.save(os.path.join(image_dir, random_filename))
|
|
|
10 |
from diffusers.utils import load_image
|
11 |
|
12 |
from PIL import Image, ImageFilter, ImageOps
|
13 |
+
from diffusers import (
|
14 |
+
StableDiffusionPipeline,
|
15 |
+
StableDiffusionImg2ImgPipeline,
|
16 |
+
StableDiffusionLatentUpscalePipeline,
|
17 |
+
)
|
18 |
+
from diffusers import (
|
19 |
+
AutoencoderKL,
|
20 |
+
UNet2DConditionModel,
|
21 |
+
DDIMScheduler,
|
22 |
+
EulerAncestralDiscreteScheduler,
|
23 |
+
DPMSolverMultistepScheduler,
|
24 |
+
)
|
25 |
from diffusers.models.attention_processor import AttnProcessor2_0
|
26 |
from transformers import CLIPTextModel, CLIPTokenizer
|
27 |
from enum import Enum
|
28 |
+
|
29 |
# from torch.hub import download_url_to_file
|
30 |
|
31 |
|
32 |
# 'stablediffusionapi/realistic-vision-v51'
|
33 |
# 'runwayml/stable-diffusion-v1-5'
|
34 |
+
sd15_name = "stablediffusionapi/realistic-vision-v51"
|
35 |
tokenizer = CLIPTokenizer.from_pretrained(sd15_name, subfolder="tokenizer")
|
36 |
text_encoder = CLIPTextModel.from_pretrained(sd15_name, subfolder="text_encoder")
|
37 |
vae = AutoencoderKL.from_pretrained(sd15_name, subfolder="vae")
|
38 |
unet = UNet2DConditionModel.from_pretrained(sd15_name, subfolder="unet")
|
39 |
+
upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(
|
40 |
+
"stabilityai/sd-x2-latent-upscaler", torch_dtype=torch.float16
|
41 |
+
)
|
42 |
|
43 |
# Change UNet
|
44 |
|
45 |
with torch.no_grad():
|
46 |
+
new_conv_in = torch.nn.Conv2d(
|
47 |
+
8,
|
48 |
+
unet.conv_in.out_channels,
|
49 |
+
unet.conv_in.kernel_size,
|
50 |
+
unet.conv_in.stride,
|
51 |
+
unet.conv_in.padding,
|
52 |
+
)
|
53 |
new_conv_in.weight.zero_()
|
54 |
new_conv_in.weight[:, :4, :, :].copy_(unet.conv_in.weight)
|
55 |
new_conv_in.bias = unet.conv_in.bias
|
|
|
59 |
|
60 |
|
61 |
def hooked_unet_forward(sample, timestep, encoder_hidden_states, **kwargs):
|
62 |
+
c_concat = kwargs["cross_attention_kwargs"]["concat_conds"].to(sample)
|
63 |
c_concat = torch.cat([c_concat] * (sample.shape[0] // c_concat.shape[0]), dim=0)
|
64 |
new_sample = torch.cat([sample, c_concat], dim=1)
|
65 |
+
kwargs["cross_attention_kwargs"] = {}
|
66 |
return unet_original_forward(new_sample, timestep, encoder_hidden_states, **kwargs)
|
67 |
|
68 |
|
|
|
70 |
|
71 |
# Load
|
72 |
|
73 |
+
model_path = "./models/iclight_sd15_fc.safetensors"
|
74 |
# download_url_to_file(url='https://huggingface.co/lllyasviel/ic-light/resolve/main/iclight_sd15_fc.safetensors', dst=model_path)
|
75 |
sd_offset = sf.load_file(model_path)
|
76 |
sd_origin = unet.state_dict()
|
|
|
81 |
|
82 |
# Device
|
83 |
|
84 |
+
device = torch.device("cuda")
|
85 |
text_encoder = text_encoder.to(device=device, dtype=torch.float16)
|
86 |
vae = vae.to(device=device, dtype=torch.bfloat16)
|
87 |
unet = unet.to(device=device, dtype=torch.float16)
|
|
|
104 |
)
|
105 |
|
106 |
euler_a_scheduler = EulerAncestralDiscreteScheduler(
|
107 |
+
num_train_timesteps=1000, beta_start=0.00085, beta_end=0.012, steps_offset=1
|
|
|
|
|
|
|
108 |
)
|
109 |
|
110 |
dpmpp_2m_sde_karras_scheduler = DPMSolverMultistepScheduler(
|
|
|
113 |
beta_end=0.012,
|
114 |
algorithm_type="sde-dpmsolver++",
|
115 |
use_karras_sigmas=True,
|
116 |
+
steps_offset=1,
|
117 |
)
|
118 |
|
119 |
# Pipelines
|
|
|
127 |
safety_checker=None,
|
128 |
requires_safety_checker=False,
|
129 |
feature_extractor=None,
|
130 |
+
image_encoder=None,
|
131 |
)
|
132 |
|
133 |
i2i_pipe = StableDiffusionImg2ImgPipeline(
|
|
|
139 |
safety_checker=None,
|
140 |
requires_safety_checker=False,
|
141 |
feature_extractor=None,
|
142 |
+
image_encoder=None,
|
143 |
)
|
144 |
|
145 |
|
|
|
155 |
return x[:i] if len(x) >= i else x + [p] * (i - len(x))
|
156 |
|
157 |
tokens = tokenizer(txt, truncation=False, add_special_tokens=False)["input_ids"]
|
158 |
+
chunks = [
|
159 |
+
[id_start] + tokens[i : i + chunk_length] + [id_end]
|
160 |
+
for i in range(0, len(tokens), chunk_length)
|
161 |
+
]
|
162 |
chunks = [pad(ck, id_pad, max_length) for ck in chunks]
|
163 |
|
164 |
token_ids = torch.tensor(chunks).to(device=device, dtype=torch.int64)
|
|
|
207 |
|
208 |
@torch.inference_mode()
|
209 |
def numpy2pytorch(imgs):
|
210 |
+
h = (
|
211 |
+
torch.from_numpy(np.stack(imgs, axis=0)).float() / 127.0 - 1.0
|
212 |
+
) # so that 127 must be strictly 0.0
|
213 |
h = h.movedim(-1, 1)
|
214 |
return h
|
215 |
|
|
|
234 |
resized_image = pil_image.resize((target_width, target_height), Image.LANCZOS)
|
235 |
return np.array(resized_image)
|
236 |
|
237 |
+
|
238 |
def remove_alpha_threshold(image, alpha_threshold=160):
|
239 |
# This function removes artifacts created by LayerDiffusion
|
240 |
mask = image[:, :, 3] < alpha_threshold
|
241 |
image[mask] = [0, 0, 0, 0]
|
242 |
return image
|
243 |
|
244 |
+
|
245 |
@torch.inference_mode()
|
246 |
+
def process(
|
247 |
+
input_fg,
|
248 |
+
prompt,
|
249 |
+
image_width,
|
250 |
+
image_height,
|
251 |
+
num_samples,
|
252 |
+
seed,
|
253 |
+
steps,
|
254 |
+
a_prompt,
|
255 |
+
n_prompt,
|
256 |
+
cfg,
|
257 |
+
highres_scale,
|
258 |
+
highres_denoise,
|
259 |
+
lowres_denoise,
|
260 |
+
bg_source,
|
261 |
+
):
|
262 |
bg_source = BGSource(bg_source)
|
263 |
input_bg = None
|
264 |
|
|
|
281 |
image = np.tile(gradient, (1, image_width))
|
282 |
input_bg = np.stack((image,) * 3, axis=-1).astype(np.uint8)
|
283 |
else:
|
284 |
+
raise "Wrong initial latent!"
|
285 |
|
286 |
rng = torch.Generator(device=device).manual_seed(int(seed))
|
287 |
|
288 |
fg = resize_and_center_crop(input_fg, image_width, image_height)
|
289 |
|
290 |
concat_conds = numpy2pytorch([fg]).to(device=vae.device, dtype=vae.dtype)
|
291 |
+
concat_conds = (
|
292 |
+
vae.encode(concat_conds).latent_dist.mode() * vae.config.scaling_factor
|
293 |
+
)
|
294 |
|
295 |
+
conds, unconds = encode_prompt_pair(
|
296 |
+
positive_prompt=prompt + ", " + a_prompt, negative_prompt=n_prompt
|
297 |
+
)
|
298 |
|
299 |
if input_bg is None:
|
300 |
+
latents = (
|
301 |
+
t2i_pipe(
|
302 |
+
prompt_embeds=conds,
|
303 |
+
negative_prompt_embeds=unconds,
|
304 |
+
width=image_width,
|
305 |
+
height=image_height,
|
306 |
+
num_inference_steps=steps,
|
307 |
+
num_images_per_prompt=num_samples,
|
308 |
+
generator=rng,
|
309 |
+
output_type="latent",
|
310 |
+
guidance_scale=cfg,
|
311 |
+
cross_attention_kwargs={"concat_conds": concat_conds},
|
312 |
+
).images.to(vae.dtype)
|
313 |
+
/ vae.config.scaling_factor
|
314 |
+
)
|
315 |
else:
|
316 |
bg = resize_and_center_crop(input_bg, image_width, image_height)
|
317 |
bg_latent = numpy2pytorch([bg]).to(device=vae.device, dtype=vae.dtype)
|
318 |
bg_latent = vae.encode(bg_latent).latent_dist.mode() * vae.config.scaling_factor
|
319 |
+
latents = (
|
320 |
+
i2i_pipe(
|
321 |
+
image=bg_latent,
|
322 |
+
strength=lowres_denoise,
|
323 |
+
prompt_embeds=conds,
|
324 |
+
negative_prompt_embeds=unconds,
|
325 |
+
width=image_width,
|
326 |
+
height=image_height,
|
327 |
+
num_inference_steps=int(round(steps / lowres_denoise)),
|
328 |
+
num_images_per_prompt=num_samples,
|
329 |
+
generator=rng,
|
330 |
+
output_type="latent",
|
331 |
+
guidance_scale=cfg,
|
332 |
+
cross_attention_kwargs={"concat_conds": concat_conds},
|
333 |
+
).images.to(vae.dtype)
|
334 |
+
/ vae.config.scaling_factor
|
335 |
+
)
|
336 |
|
337 |
pixels = vae.decode(latents).sample
|
338 |
pixels = pytorch2numpy(pixels)
|
339 |
+
pixels = [
|
340 |
+
resize_without_crop(
|
341 |
+
image=p,
|
342 |
+
target_width=int(round(image_width * highres_scale / 64.0) * 64),
|
343 |
+
target_height=int(round(image_height * highres_scale / 64.0) * 64),
|
344 |
+
)
|
345 |
+
for p in pixels
|
346 |
+
]
|
347 |
|
348 |
pixels = numpy2pytorch(pixels).to(device=vae.device, dtype=vae.dtype)
|
349 |
latents = vae.encode(pixels).latent_dist.mode() * vae.config.scaling_factor
|
|
|
353 |
|
354 |
fg = resize_and_center_crop(input_fg, image_width, image_height)
|
355 |
concat_conds = numpy2pytorch([fg]).to(device=vae.device, dtype=vae.dtype)
|
356 |
+
concat_conds = (
|
357 |
+
vae.encode(concat_conds).latent_dist.mode() * vae.config.scaling_factor
|
358 |
+
)
|
359 |
+
|
360 |
+
latents = (
|
361 |
+
i2i_pipe(
|
362 |
+
image=latents,
|
363 |
+
strength=highres_denoise,
|
364 |
+
prompt_embeds=conds,
|
365 |
+
negative_prompt_embeds=unconds,
|
366 |
+
width=image_width,
|
367 |
+
height=image_height,
|
368 |
+
num_inference_steps=int(round(steps / highres_denoise)),
|
369 |
+
num_images_per_prompt=num_samples,
|
370 |
+
generator=rng,
|
371 |
+
output_type="latent",
|
372 |
+
guidance_scale=cfg,
|
373 |
+
cross_attention_kwargs={"concat_conds": concat_conds},
|
374 |
+
).images.to(vae.dtype)
|
375 |
+
/ vae.config.scaling_factor
|
376 |
+
)
|
377 |
|
378 |
pixels = vae.decode(latents).sample
|
379 |
|
|
|
391 |
else:
|
392 |
target_height, target_width = 512 * 2, 640 * 2
|
393 |
|
394 |
+
left_right_padding = (
|
395 |
+
max(target_width, image_width) - min(target_width, image_width)
|
396 |
+
) // 2
|
397 |
|
398 |
original = cv2.copyMakeBorder(
|
399 |
+
original,
|
400 |
+
top=max(target_height, image_height) - min(target_height, image_height),
|
401 |
bottom=0,
|
402 |
+
left=left_right_padding,
|
403 |
+
right=left_right_padding,
|
404 |
+
borderType=cv2.BORDER_CONSTANT,
|
405 |
+
value=(0, 0, 0),
|
406 |
)
|
407 |
|
408 |
transform = A.Compose(
|
|
|
421 |
|
422 |
return transform(image=original)["image"]
|
423 |
|
424 |
+
|
425 |
class BGSource(Enum):
|
426 |
NONE = "None"
|
427 |
LEFT = "Left Light"
|
|
|
439 |
"sunshine, cafe, chilled",
|
440 |
"exhibition, paintings",
|
441 |
"beach",
|
442 |
+
"winter, snow" "forrest, cloudy",
|
|
|
443 |
"party, people",
|
444 |
"cozy living room, sofa, shelf",
|
445 |
"mountains",
|
|
|
450 |
"appartment, soft light",
|
451 |
"garden",
|
452 |
"school",
|
453 |
+
"art exhibition with paintings in background",
|
454 |
]
|
455 |
|
456 |
os.makedirs(ground_truth_dir, exist_ok=True)
|
|
|
460 |
random.shuffle(all_images)
|
461 |
|
462 |
for filename in all_images:
|
463 |
+
if filename.lower().endswith(
|
464 |
+
(".png", ".jpg", ".jpeg")
|
465 |
+
): # Check if the file is an image
|
466 |
|
467 |
letters = string.ascii_lowercase
|
468 |
random_string = "".join(random.choice(letters) for i in range(13))
|
|
|
478 |
image = np.array(image)
|
479 |
|
480 |
image_augmented = augment(image)
|
481 |
+
Image.fromarray(image_augmented).getchannel("A").save(
|
482 |
+
os.path.join(ground_truth_dir, random_filename)
|
483 |
+
)
|
484 |
|
485 |
image_augmented = image_augmented[:, :, :3]
|
486 |
|
|
|
489 |
image_height, image_width, _ = image_augmented.shape
|
490 |
|
491 |
num_samples = 1
|
492 |
+
seed = random.randint(1, 123456789012345678901234567890)
|
493 |
steps = 25
|
494 |
constant_prompt = "details, high quality"
|
495 |
prompt = random.choice(prompts)
|
|
|
499 |
highres_denoise = 0.7
|
500 |
lowres_denoise = 0.5
|
501 |
bg_source = BGSource.NONE
|
502 |
+
|
503 |
+
results = process(
|
504 |
+
image_augmented,
|
505 |
+
constant_prompt,
|
506 |
+
image_width,
|
507 |
+
image_height,
|
508 |
+
num_samples,
|
509 |
+
seed,
|
510 |
+
steps,
|
511 |
+
prompt,
|
512 |
+
n_prompt,
|
513 |
+
cfg,
|
514 |
+
highres_scale,
|
515 |
+
highres_denoise,
|
516 |
+
lowres_denoise,
|
517 |
+
bg_source,
|
518 |
+
)
|
519 |
result_image = Image.fromarray(results[0])
|
520 |
+
result_image.save(os.path.join(image_dir, random_filename))
|