Update app_test.py
Browse files- app_test.py +178 -10
app_test.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
import
|
2 |
import huggingface_hub
|
3 |
|
4 |
huggingface_hub.snapshot_download(
|
@@ -11,18 +11,186 @@ huggingface_hub.snapshot_download(
|
|
11 |
local_dir_use_symlinks=False,
|
12 |
)
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
-
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
-
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
24 |
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
|
|
|
1 |
+
import spaces
|
2 |
import huggingface_hub
|
3 |
|
4 |
huggingface_hub.snapshot_download(
|
|
|
11 |
local_dir_use_symlinks=False,
|
12 |
)
|
13 |
|
14 |
+
import gradio as gr
|
15 |
+
from diffusers import StableDiffusionXLControlNetInpaintPipeline, ControlNetModel
|
16 |
+
from rembg import remove
|
17 |
+
from PIL import Image
|
18 |
+
import torch
|
19 |
+
from ip_adapter import IPAdapterXL
|
20 |
+
from ip_adapter.utils import register_cross_attention_hook, get_net_attn_map, attnmaps2images
|
21 |
+
from PIL import Image, ImageChops, ImageEnhance
|
22 |
+
import numpy as np
|
23 |
+
|
24 |
import os
|
25 |
+
import glob
|
26 |
+
import torch
|
27 |
+
import cv2
|
28 |
+
import argparse
|
29 |
+
|
30 |
+
import DPT.util.io
|
31 |
+
|
32 |
+
from torchvision.transforms import Compose
|
33 |
+
|
34 |
+
from DPT.dpt.models import DPTDepthModel
|
35 |
+
from DPT.dpt.midas_net import MidasNet_large
|
36 |
+
from DPT.dpt.transforms import Resize, NormalizeImage, PrepareForNet
|
37 |
+
|
38 |
+
"""
|
39 |
+
Get ZeST Ready
|
40 |
+
"""
|
41 |
+
base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
|
42 |
+
image_encoder_path = "models/image_encoder"
|
43 |
+
ip_ckpt = "sdxl_models/ip-adapter_sdxl_vit-h.bin"
|
44 |
+
controlnet_path = "diffusers/controlnet-depth-sdxl-1.0"
|
45 |
+
device = "cuda"
|
46 |
+
torch.cuda.empty_cache()
|
47 |
+
|
48 |
+
# load SDXL pipeline
|
49 |
+
controlnet = ControlNetModel.from_pretrained(controlnet_path, variant="fp16", use_safetensors=True, torch_dtype=torch.float16).to(device)
|
50 |
+
pipe = StableDiffusionXLControlNetInpaintPipeline.from_pretrained(
|
51 |
+
base_model_path,
|
52 |
+
controlnet=controlnet,
|
53 |
+
use_safetensors=True,
|
54 |
+
torch_dtype=torch.float16,
|
55 |
+
add_watermarker=False,
|
56 |
+
).to(device)
|
57 |
+
pipe.unet = register_cross_attention_hook(pipe.unet)
|
58 |
+
|
59 |
+
ip_model = IPAdapterXL(pipe, image_encoder_path, ip_ckpt, device)
|
60 |
+
|
61 |
+
|
62 |
+
"""
|
63 |
+
Get Depth Model Ready
|
64 |
+
"""
|
65 |
+
model_path = "DPT/weights/dpt_hybrid-midas-501f0c75.pt"
|
66 |
+
net_w = net_h = 384
|
67 |
+
model = DPTDepthModel(
|
68 |
+
path=model_path,
|
69 |
+
backbone="vitb_rn50_384",
|
70 |
+
non_negative=True,
|
71 |
+
enable_attention_hooks=False,
|
72 |
+
)
|
73 |
+
normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
|
74 |
+
|
75 |
+
transform = Compose(
|
76 |
+
[
|
77 |
+
Resize(
|
78 |
+
net_w,
|
79 |
+
net_h,
|
80 |
+
resize_target=None,
|
81 |
+
keep_aspect_ratio=True,
|
82 |
+
ensure_multiple_of=32,
|
83 |
+
resize_method="minimal",
|
84 |
+
image_interpolation_method=cv2.INTER_CUBIC,
|
85 |
+
),
|
86 |
+
normalization,
|
87 |
+
PrepareForNet(),
|
88 |
+
]
|
89 |
+
)
|
90 |
+
|
91 |
+
model.eval()
|
92 |
+
|
93 |
+
@spaces.GPU()
|
94 |
+
def greet(input_image, material_exemplar):
|
95 |
+
|
96 |
+
"""
|
97 |
+
Compute depth map from input_image
|
98 |
+
"""
|
99 |
+
|
100 |
+
img = np.array(input_image)
|
101 |
+
|
102 |
+
img_input = transform({"image": img})["image"]
|
103 |
+
|
104 |
+
# compute
|
105 |
+
with torch.no_grad():
|
106 |
+
sample = torch.from_numpy(img_input).unsqueeze(0)
|
107 |
+
|
108 |
+
# if optimize == True and device == torch.device("cuda"):
|
109 |
+
# sample = sample.to(memory_format=torch.channels_last)
|
110 |
+
# sample = sample.half()
|
111 |
|
112 |
+
prediction = model.forward(sample)
|
113 |
+
prediction = (
|
114 |
+
torch.nn.functional.interpolate(
|
115 |
+
prediction.unsqueeze(1),
|
116 |
+
size=img.shape[:2],
|
117 |
+
mode="bicubic",
|
118 |
+
align_corners=False,
|
119 |
+
)
|
120 |
+
.squeeze()
|
121 |
+
.cpu()
|
122 |
+
.numpy()
|
123 |
+
)
|
124 |
+
|
125 |
+
depth_min = prediction.min()
|
126 |
+
depth_max = prediction.max()
|
127 |
+
bits = 2
|
128 |
+
max_val = (2 ** (8 * bits)) - 1
|
129 |
|
130 |
+
if depth_max - depth_min > np.finfo("float").eps:
|
131 |
+
out = max_val * (prediction - depth_min) / (depth_max - depth_min)
|
132 |
+
else:
|
133 |
+
out = np.zeros(prediction.shape, dtype=depth.dtype)
|
134 |
+
|
135 |
+
out = (out / 256).astype('uint8')
|
136 |
+
depth_map = Image.fromarray(out).resize((1024, 1024))
|
137 |
+
|
138 |
+
|
139 |
+
"""
|
140 |
+
Process foreground decolored image
|
141 |
+
"""
|
142 |
+
rm_bg = remove(input_image)
|
143 |
+
target_mask = rm_bg.convert("RGB").point(lambda x: 0 if x < 1 else 255).convert('L').convert('RGB')
|
144 |
+
mask_target_img = ImageChops.lighter(input_image, target_mask)
|
145 |
+
invert_target_mask = ImageChops.invert(target_mask)
|
146 |
+
gray_target_image = input_image.convert('L').convert('RGB')
|
147 |
+
gray_target_image = ImageEnhance.Brightness(gray_target_image)
|
148 |
+
factor = 1.0 # Try adjusting this to get the desired brightness
|
149 |
+
gray_target_image = gray_target_image.enhance(factor)
|
150 |
+
grayscale_img = ImageChops.darker(gray_target_image, target_mask)
|
151 |
+
img_black_mask = ImageChops.darker(input_image, invert_target_mask)
|
152 |
+
grayscale_init_img = ImageChops.lighter(img_black_mask, grayscale_img)
|
153 |
+
init_img = grayscale_init_img
|
154 |
+
|
155 |
+
"""
|
156 |
+
Process material exemplar and resize all images
|
157 |
+
"""
|
158 |
+
ip_image = material_exemplar.resize((1024, 1024))
|
159 |
+
init_img = init_img.resize((1024,1024))
|
160 |
+
mask = target_mask.resize((1024, 1024))
|
161 |
+
|
162 |
+
|
163 |
+
num_samples = 1
|
164 |
+
images = ip_model.generate(pil_image=ip_image, image=init_img, control_image=depth_map, mask_image=mask, controlnet_conditioning_scale=0.9, num_samples=num_samples, num_inference_steps=30, seed=42)
|
165 |
+
|
166 |
+
return images[0]
|
167 |
|
168 |
+
css = """
|
169 |
+
#col-container{
|
170 |
+
margin: 0 auto;
|
171 |
+
max-width: 960px;
|
172 |
+
}
|
173 |
+
"""
|
174 |
|
175 |
+
with gr.Blocks(css=css) as demo:
|
176 |
+
with gr.Column(elem_id="col-container"):
|
177 |
+
gr.Markdown("""
|
178 |
+
# ZeST: Zero-Shot Material Transfer from a Single Image
|
179 |
+
<p>Upload two images -- input image and material exemplar. (both 1024*1024 for better results) <br />
|
180 |
+
ZeST extracts the material from the exemplar and cast it onto the input image following the original lighting cues.</p>
|
181 |
+
""")
|
182 |
+
with gr.Row():
|
183 |
+
with gr.Column():
|
184 |
+
with gr.Row():
|
185 |
+
input_image = gr.Image(type="pil", label="input image")
|
186 |
+
input_image2 = gr.Image(type="pil", label = "material examplar")
|
187 |
+
submit_btn = gr.Button("Submit")
|
188 |
+
gr.Examples(
|
189 |
+
examples = [["demo_assets/input_imgs/pumpkin.png", "demo_assets/material_exemplars/cup_glaze.png"]],
|
190 |
+
inputs = [input_image, input_image2]
|
191 |
+
)
|
192 |
+
with gr.Column():
|
193 |
+
output_image = gr.Image(label="transfer result")
|
194 |
+
submit_btn.click(fn=greet, inputs=[input_image, input_image2], outputs=[output_image])
|
195 |
|
196 |
+
demo.queue().launch()
|