Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,238 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys
|
2 |
+
sys.path.append('./')
|
3 |
+
import gradio as gr
|
4 |
+
import torch
|
5 |
+
from PIL import Image
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from transformers import CLIPImageProcessor
|
8 |
+
|
9 |
+
# Add necessary imports and initialize the model as in your code...
|
10 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Literal
|
11 |
+
from ip_adapter.ip_adapter import Resampler
|
12 |
+
import matplotlib.pyplot as plt
|
13 |
+
|
14 |
+
|
15 |
+
import torch.utils.data as data
|
16 |
+
import torchvision
|
17 |
+
import numpy as np
|
18 |
+
import torch
|
19 |
+
import torch.nn.functional as F
|
20 |
+
from accelerate.logging import get_logger
|
21 |
+
from accelerate.utils import set_seed
|
22 |
+
from torchvision import transforms
|
23 |
+
|
24 |
+
from diffusers import AutoencoderKL, DDPMScheduler
|
25 |
+
from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModelWithProjection,CLIPTextModelWithProjection, CLIPTextModel,
|
26 |
+
|
27 |
+
|
28 |
+
from src.unet_hacked_tryon import UNet2DConditionModel
|
29 |
+
from src.unet_hacked_garmnet import UNet2DConditionModel as UNet2DConditionModel_ref
|
30 |
+
from src.tryon_pipeline import StableDiffusionXLInpaintPipeline as TryonPipeline
|
31 |
+
# Define a class to hold configuration arguments
|
32 |
+
class Args:
|
33 |
+
def __init__(self):
|
34 |
+
self.pretrained_model_name_or_path = "yisol/IDM-VTON"
|
35 |
+
self.width = 768
|
36 |
+
self.height = 1024
|
37 |
+
self.num_inference_steps = 10
|
38 |
+
self.seed = 42
|
39 |
+
self.guidance_scale = 2.0
|
40 |
+
self.mixed_precision = None
|
41 |
+
|
42 |
+
# Determine the device to be used for computations (CUDA if available)
|
43 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
44 |
+
|
45 |
+
logger = get_logger(__name__, log_level="INFO")
|
46 |
+
|
47 |
+
def pil_to_tensor(images):
|
48 |
+
images = np.array(images).astype(np.float32) / 255.0
|
49 |
+
images = torch.from_numpy(images.transpose(2, 0, 1))
|
50 |
+
return images
|
51 |
+
|
52 |
+
|
53 |
+
|
54 |
+
args = Args()
|
55 |
+
|
56 |
+
# Define the data type for model weights
|
57 |
+
weight_dtype = torch.float16
|
58 |
+
|
59 |
+
if args.seed is not None:
|
60 |
+
set_seed(args.seed)
|
61 |
+
|
62 |
+
|
63 |
+
# Load scheduler, tokenizer and models.
|
64 |
+
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
|
65 |
+
vae = AutoencoderKL.from_pretrained(
|
66 |
+
args.pretrained_model_name_or_path,
|
67 |
+
subfolder="vae",
|
68 |
+
torch_dtype=torch.float16,
|
69 |
+
)
|
70 |
+
unet = UNet2DConditionModel.from_pretrained(
|
71 |
+
args.pretrained_model_name_or_path,
|
72 |
+
subfolder="unet",
|
73 |
+
torch_dtype=torch.float16,
|
74 |
+
)
|
75 |
+
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
76 |
+
args.pretrained_model_name_or_path,
|
77 |
+
subfolder="image_encoder",
|
78 |
+
torch_dtype=torch.float16,
|
79 |
+
)
|
80 |
+
unet_encoder = UNet2DConditionModel_ref.from_pretrained(
|
81 |
+
args.pretrained_model_name_or_path,
|
82 |
+
subfolder="unet_encoder",
|
83 |
+
torch_dtype=torch.float16,
|
84 |
+
)
|
85 |
+
text_encoder_one = CLIPTextModel.from_pretrained(
|
86 |
+
args.pretrained_model_name_or_path,
|
87 |
+
subfolder="text_encoder",
|
88 |
+
torch_dtype=torch.float16,
|
89 |
+
)
|
90 |
+
text_encoder_two = CLIPTextModelWithProjection.from_pretrained(
|
91 |
+
args.pretrained_model_name_or_path,
|
92 |
+
subfolder="text_encoder_2",
|
93 |
+
torch_dtype=torch.float16,
|
94 |
+
)
|
95 |
+
tokenizer_one = AutoTokenizer.from_pretrained(
|
96 |
+
args.pretrained_model_name_or_path,
|
97 |
+
subfolder="tokenizer",
|
98 |
+
revision=None,
|
99 |
+
use_fast=False,
|
100 |
+
)
|
101 |
+
tokenizer_two = AutoTokenizer.from_pretrained(
|
102 |
+
args.pretrained_model_name_or_path,
|
103 |
+
subfolder="tokenizer_2",
|
104 |
+
revision=None,
|
105 |
+
use_fast=False,
|
106 |
+
)
|
107 |
+
# Freeze vae and text_encoder and set unet to trainable
|
108 |
+
unet.requires_grad_(False)
|
109 |
+
vae.requires_grad_(False)
|
110 |
+
image_encoder.requires_grad_(False)
|
111 |
+
unet_encoder.requires_grad_(False)
|
112 |
+
text_encoder_one.requires_grad_(False)
|
113 |
+
text_encoder_two.requires_grad_(False)
|
114 |
+
unet_encoder.to(device, weight_dtype)
|
115 |
+
unet.eval()
|
116 |
+
unet_encoder.eval()
|
117 |
+
|
118 |
+
pipe = TryonPipeline.from_pretrained(
|
119 |
+
args.pretrained_model_name_or_path,
|
120 |
+
unet=unet,
|
121 |
+
vae=vae,
|
122 |
+
feature_extractor= CLIPImageProcessor(),
|
123 |
+
text_encoder = text_encoder_one,
|
124 |
+
text_encoder_2 = text_encoder_two,
|
125 |
+
tokenizer = tokenizer_one,
|
126 |
+
tokenizer_2 = tokenizer_two,
|
127 |
+
scheduler = noise_scheduler,
|
128 |
+
image_encoder=image_encoder,
|
129 |
+
unet_encoder = unet_encoder,
|
130 |
+
torch_dtype=torch.float16,
|
131 |
+
).to(device)
|
132 |
+
# pipe.enable_sequential_cpu_offload()
|
133 |
+
# pipe.enable_model_cpu_offload()
|
134 |
+
# pipe.enable_vae_slicing()
|
135 |
+
|
136 |
+
# Function to generate the image based on inputs
|
137 |
+
def generate_virtual_try_on(person_image, cloth_image, mask_image, pose_image,cloth_des):
|
138 |
+
# Prepare the input images as tensors
|
139 |
+
person_image = person_image.resize((args.width, args.height))
|
140 |
+
cloth_image = cloth_image.resize((args.width, args.height))
|
141 |
+
mask_image = mask_image.resize((args.width, args.height))
|
142 |
+
pose_image = pose_image.resize((args.width, args.height))
|
143 |
+
# Define transformations
|
144 |
+
transform = transforms.Compose([
|
145 |
+
transforms.ToTensor(),
|
146 |
+
transforms.Normalize([0.5], [0.5]),
|
147 |
+
])
|
148 |
+
guidance_scale=2.0
|
149 |
+
seed=42
|
150 |
+
|
151 |
+
to_tensor = transforms.ToTensor()
|
152 |
+
|
153 |
+
person_tensor = transform(person_image).unsqueeze(0).to(device) # Add batch dimension
|
154 |
+
cloth_pure = transform(cloth_image).unsqueeze(0).to(device)
|
155 |
+
mask_tensor = to_tensor(mask_image)[:1].unsqueeze(0).to(device) # Keep only one channel
|
156 |
+
pose_tensor = transform(pose_image).unsqueeze(0).to(device)
|
157 |
+
|
158 |
+
# Prepare text prompts
|
159 |
+
prompt = ["A person wearing the cloth"+cloth_des] # Example prompt
|
160 |
+
negative_prompt = ["monochrome, lowres, bad anatomy, worst quality, low quality"]
|
161 |
+
|
162 |
+
# Encode prompts
|
163 |
+
with torch.inference_mode():
|
164 |
+
(
|
165 |
+
prompt_embeds,
|
166 |
+
negative_prompt_embeds,
|
167 |
+
pooled_prompt_embeds,
|
168 |
+
negative_pooled_prompt_embeds,
|
169 |
+
) = pipe.encode_prompt(
|
170 |
+
prompt,
|
171 |
+
num_images_per_prompt=1,
|
172 |
+
do_classifier_free_guidance=True,
|
173 |
+
negative_prompt=negative_prompt,
|
174 |
+
)
|
175 |
+
prompt_cloth = ["a photo of"+cloth_des]
|
176 |
+
with torch.inference_mode():
|
177 |
+
(
|
178 |
+
prompt_embeds_c,
|
179 |
+
_,
|
180 |
+
_,
|
181 |
+
_,
|
182 |
+
) = pipe.encode_prompt(
|
183 |
+
prompt_cloth,
|
184 |
+
num_images_per_prompt=1,
|
185 |
+
do_classifier_free_guidance=False,
|
186 |
+
negative_prompt=negative_prompt,
|
187 |
+
)
|
188 |
+
|
189 |
+
# Encode garment using IP-Adapter
|
190 |
+
clip_processor = CLIPImageProcessor()
|
191 |
+
image_embeds = clip_processor(images=cloth_image, return_tensors="pt").pixel_values.to(device)
|
192 |
+
|
193 |
+
# Generate the image
|
194 |
+
generator = torch.Generator(pipe.device).manual_seed(seed) if seed is not None else None
|
195 |
+
|
196 |
+
with torch.no_grad():
|
197 |
+
images = pipe(
|
198 |
+
prompt_embeds=prompt_embeds,
|
199 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
200 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
201 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
202 |
+
num_inference_steps=args.num_inference_steps,
|
203 |
+
generator=generator,
|
204 |
+
strength=1.0,
|
205 |
+
pose_img=pose_tensor,
|
206 |
+
text_embeds_cloth=prompt_embeds_c,
|
207 |
+
cloth=cloth_pure,
|
208 |
+
mask_image=mask_tensor,
|
209 |
+
image=(person_tensor + 1.0) / 2.0,
|
210 |
+
height=args.height,
|
211 |
+
width=args.width,
|
212 |
+
guidance_scale=guidance_scale,
|
213 |
+
ip_adapter_image=image_embeds,
|
214 |
+
)[0]
|
215 |
+
|
216 |
+
# Convert output image to PIL format for display
|
217 |
+
generated_image = transforms.ToPILImage()(images[0])
|
218 |
+
return generated_image
|
219 |
+
|
220 |
+
# Create Gradio interface
|
221 |
+
iface = gr.Interface(
|
222 |
+
fn=generate_virtual_try_on,
|
223 |
+
inputs=[
|
224 |
+
gr.Image(type="pil", label="Person Image"),
|
225 |
+
gr.Image(type="pil", label="Cloth Image"),
|
226 |
+
gr.Image(type="pil", label="Mask Image"),
|
227 |
+
gr.Image(type="pil", label="Pose Image"),
|
228 |
+
gr.Textbox(label="cloth_des"), # Add text input
|
229 |
+
|
230 |
+
|
231 |
+
|
232 |
+
|
233 |
+
],
|
234 |
+
outputs=gr.Image(type="pil", label="Generated Image"),
|
235 |
+
)
|
236 |
+
|
237 |
+
# Launch the interface
|
238 |
+
iface.launch()
|