Spaces:
Runtime error
Runtime error
restyle launch codes
Browse files- app.py +5 -308
- launch/__init__.py +0 -0
- launch/image_generation.py +127 -0
- launch/model_generation.py +183 -0
- launch/utils.py +14 -0
app.py
CHANGED
@@ -1,319 +1,16 @@
|
|
1 |
-
import os
|
2 |
-
import shutil
|
3 |
-
import tempfile
|
4 |
-
|
5 |
import gradio as gr
|
6 |
-
import numpy as np
|
7 |
-
import rembg
|
8 |
-
import spaces
|
9 |
-
import torch
|
10 |
-
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler, StableDiffusionXLPipeline, EulerDiscreteScheduler
|
11 |
-
from einops import rearrange
|
12 |
-
from huggingface_hub import hf_hub_download
|
13 |
-
from omegaconf import OmegaConf
|
14 |
-
from PIL import Image
|
15 |
-
from pytorch_lightning import seed_everything
|
16 |
-
from torchvision.transforms import v2
|
17 |
-
from safetensors.torch import load_file
|
18 |
-
|
19 |
-
from src.utils.camera_util import (FOV_to_intrinsics, get_circular_camera_poses,
|
20 |
-
get_zero123plus_input_cameras)
|
21 |
-
from src.utils.infer_util import (remove_background, resize_foreground)
|
22 |
-
from src.utils.mesh_util import save_glb, save_obj
|
23 |
-
from src.utils.train_util import instantiate_from_config
|
24 |
-
|
25 |
-
|
26 |
-
def find_cuda():
|
27 |
-
cuda_home = os.environ.get('CUDA_HOME') or os.environ.get('CUDA_PATH')
|
28 |
-
if cuda_home and os.path.exists(cuda_home):
|
29 |
-
return cuda_home
|
30 |
-
|
31 |
-
nvcc_path = shutil.which('nvcc')
|
32 |
-
if nvcc_path:
|
33 |
-
cuda_path = os.path.dirname(os.path.dirname(nvcc_path))
|
34 |
-
return cuda_path
|
35 |
-
|
36 |
-
return None
|
37 |
-
|
38 |
-
|
39 |
-
def get_render_cameras(batch_size=1, M=120, radius=2.5, elevation=10.0, is_flexicubes=False):
|
40 |
-
c2ws = get_circular_camera_poses(M=M, radius=radius, elevation=elevation)
|
41 |
-
if is_flexicubes:
|
42 |
-
cameras = torch.linalg.inv(c2ws)
|
43 |
-
cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1, 1)
|
44 |
-
else:
|
45 |
-
extrinsics = c2ws.flatten(-2)
|
46 |
-
intrinsics = FOV_to_intrinsics(50.0).unsqueeze(
|
47 |
-
0).repeat(M, 1, 1).float().flatten(-2)
|
48 |
-
cameras = torch.cat([extrinsics, intrinsics], dim=-1)
|
49 |
-
cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1)
|
50 |
-
return cameras
|
51 |
-
|
52 |
-
|
53 |
-
def check_input_image(input_image):
|
54 |
-
if input_image is None:
|
55 |
-
raise gr.Error("No image selected!")
|
56 |
-
|
57 |
-
|
58 |
-
def preprocess(input_image):
|
59 |
-
rembg_session = rembg.new_session()
|
60 |
-
|
61 |
-
input_image = remove_background(input_image, rembg_session)
|
62 |
-
input_image = resize_foreground(input_image, 0.85)
|
63 |
-
|
64 |
-
return input_image
|
65 |
-
|
66 |
-
|
67 |
-
def generate_prompt(subject, style, color_scheme, angle, lighting_type, additional_details):
|
68 |
-
prompt = f"A 3D cartoon render of {subject}, featuring the entire body and shape, on a transparent background. The style should be {style}, with {color_scheme} colors, emphasizing the essential features and lines. The pose should clearly showcase the full form of the {subject} from a {angle} perspective. Lighting is {lighting_type}, highlighting the volume and depth of the subject. {additional_details}. Output as a high-resolution PNG with no background."
|
69 |
-
return prompt
|
70 |
-
|
71 |
-
|
72 |
-
@spaces.GPU
|
73 |
-
def generate_image(subject, style, color_scheme, angle, lighting_type, additional_details):
|
74 |
-
checkpoint = "sdxl_lightning_8step_unet.safetensors"
|
75 |
-
num_inference_steps = 8
|
76 |
-
|
77 |
-
pipe.scheduler = EulerDiscreteScheduler.from_config(
|
78 |
-
pipe.scheduler.config, timestep_spacing="trailing")
|
79 |
-
pipe.unet.load_state_dict(
|
80 |
-
load_file(hf_hub_download(repo, checkpoint), device="cuda"))
|
81 |
-
|
82 |
-
prompt = generate_prompt(subject, style, color_scheme,
|
83 |
-
angle, lighting_type, additional_details)
|
84 |
-
results = pipe(
|
85 |
-
prompt, num_inference_steps=num_inference_steps, guidance_scale=0)
|
86 |
-
return results.images[0]
|
87 |
-
|
88 |
-
|
89 |
-
@spaces.GPU
|
90 |
-
def generate_mvs(input_image, sample_steps, sample_seed):
|
91 |
-
seed_everything(sample_seed)
|
92 |
-
|
93 |
-
z123_image = pipeline(
|
94 |
-
input_image, num_inference_steps=sample_steps).images[0]
|
95 |
-
|
96 |
-
show_image = np.asarray(z123_image, dtype=np.uint8)
|
97 |
-
show_image = torch.from_numpy(show_image)
|
98 |
-
show_image = rearrange(
|
99 |
-
show_image, '(n h) (m w) c -> (n m) h w c', n=3, m=2)
|
100 |
-
show_image = rearrange(
|
101 |
-
show_image, '(n m) h w c -> (n h) (m w) c', n=2, m=3)
|
102 |
-
show_image = Image.fromarray(show_image.numpy())
|
103 |
|
104 |
-
|
|
|
105 |
|
106 |
|
107 |
-
@spaces.GPU
|
108 |
-
def make3d(images):
|
109 |
-
global model
|
110 |
-
if IS_FLEXICUBES:
|
111 |
-
model.init_flexicubes_geometry(device, use_renderer=False)
|
112 |
-
model = model.eval()
|
113 |
-
|
114 |
-
images = np.asarray(images, dtype=np.float32) / 255.0
|
115 |
-
images = torch.from_numpy(images).permute(2, 0, 1).contiguous().float()
|
116 |
-
images = rearrange(images, 'c (n h) (m w) -> (n m) c h w', n=3, m=2)
|
117 |
-
|
118 |
-
input_cameras = get_zero123plus_input_cameras(
|
119 |
-
batch_size=1, radius=4.0).to(device)
|
120 |
-
render_cameras = get_render_cameras(
|
121 |
-
batch_size=1, radius=2.5, is_flexicubes=IS_FLEXICUBES).to(device)
|
122 |
-
|
123 |
-
images = images.unsqueeze(0).to(device)
|
124 |
-
images = v2.functional.resize(
|
125 |
-
images, (320, 320), interpolation=3, antialias=True).clamp(0, 1)
|
126 |
-
|
127 |
-
mesh_fpath = tempfile.NamedTemporaryFile(suffix=f".obj", delete=False).name
|
128 |
-
print(mesh_fpath)
|
129 |
-
mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
|
130 |
-
mesh_dirname = os.path.dirname(mesh_fpath)
|
131 |
-
mesh_glb_fpath = os.path.join(mesh_dirname, f"{mesh_basename}.glb")
|
132 |
-
|
133 |
-
with torch.no_grad():
|
134 |
-
planes = model.forward_planes(images, input_cameras)
|
135 |
-
mesh_out = model.extract_mesh(
|
136 |
-
planes, use_texture_map=False, **infer_config)
|
137 |
-
|
138 |
-
vertices, faces, vertex_colors = mesh_out
|
139 |
-
vertices = vertices[:, [1, 2, 0]]
|
140 |
-
|
141 |
-
save_glb(vertices, faces, vertex_colors, mesh_glb_fpath)
|
142 |
-
save_obj(vertices, faces, vertex_colors, mesh_fpath)
|
143 |
-
|
144 |
-
print(f"Mesh saved to {mesh_fpath}")
|
145 |
-
|
146 |
-
return mesh_fpath, mesh_glb_fpath
|
147 |
-
|
148 |
-
|
149 |
-
# Configuration
|
150 |
-
cuda_path = find_cuda()
|
151 |
-
config_path = 'configs/instant-mesh-large.yaml'
|
152 |
-
config = OmegaConf.load(config_path)
|
153 |
-
config_name = os.path.basename(config_path).replace('.yaml', '')
|
154 |
-
model_config = config.model_config
|
155 |
-
infer_config = config.infer_config
|
156 |
-
|
157 |
-
IS_FLEXICUBES = config_name.startswith('instant-mesh')
|
158 |
-
device = torch.device('cuda')
|
159 |
-
|
160 |
-
# Load diffusion model
|
161 |
-
print('Loading diffusion model ...')
|
162 |
-
pipeline = DiffusionPipeline.from_pretrained(
|
163 |
-
"sudo-ai/zero123plus-v1.2",
|
164 |
-
custom_pipeline="zero123plus",
|
165 |
-
torch_dtype=torch.float16,
|
166 |
-
)
|
167 |
-
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
|
168 |
-
pipeline.scheduler.config, timestep_spacing='trailing'
|
169 |
-
)
|
170 |
-
|
171 |
-
unet_ckpt_path = hf_hub_download(
|
172 |
-
repo_id="TencentARC/InstantMesh", filename="diffusion_pytorch_model.bin", repo_type="model")
|
173 |
-
state_dict = torch.load(unet_ckpt_path, map_location='cpu')
|
174 |
-
pipeline.unet.load_state_dict(state_dict, strict=True)
|
175 |
-
|
176 |
-
pipeline = pipeline.to(device)
|
177 |
-
|
178 |
-
# Load reconstruction model
|
179 |
-
print('Loading reconstruction model ...')
|
180 |
-
model_ckpt_path = hf_hub_download(
|
181 |
-
repo_id="TencentARC/InstantMesh", filename="instant_mesh_large.ckpt", repo_type="model")
|
182 |
-
model = instantiate_from_config(model_config)
|
183 |
-
state_dict = torch.load(model_ckpt_path, map_location='cpu')['state_dict']
|
184 |
-
state_dict = {k[14:]: v for k, v in state_dict.items() if k.startswith(
|
185 |
-
'lrm_generator.') and 'source_camera' not in k}
|
186 |
-
model.load_state_dict(state_dict, strict=True)
|
187 |
-
|
188 |
-
model = model.to(device)
|
189 |
-
|
190 |
-
# Load StableDiffusionXL model
|
191 |
-
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
192 |
-
repo = "ByteDance/SDXL-Lightning"
|
193 |
-
|
194 |
-
pipe = StableDiffusionXLPipeline.from_pretrained(
|
195 |
-
base, torch_dtype=torch.float16, variant="fp16").to("cuda")
|
196 |
-
|
197 |
-
print('Loading Finished!')
|
198 |
-
|
199 |
with gr.Blocks() as demo:
|
200 |
with gr.Group():
|
201 |
with gr.Tab("Generate Image and Remove Background"):
|
202 |
-
|
203 |
-
subject = gr.Textbox(label='Subject', scale=2)
|
204 |
-
style = gr.Dropdown(
|
205 |
-
label='Style',
|
206 |
-
choices=['Pixar-like', 'Disney-esque', 'Anime-inspired'],
|
207 |
-
value='Pixar-like',
|
208 |
-
multiselect=False,
|
209 |
-
scale=2
|
210 |
-
)
|
211 |
-
color_scheme = gr.Dropdown(
|
212 |
-
label='Color Scheme',
|
213 |
-
choices=['Vibrant', 'Pastel',
|
214 |
-
'Monochromatic', 'Black and White'],
|
215 |
-
value='Vibrant',
|
216 |
-
multiselect=False,
|
217 |
-
scale=2
|
218 |
-
)
|
219 |
-
angle = gr.Dropdown(
|
220 |
-
label='Angle',
|
221 |
-
choices=['Front', 'Side', 'Three-quarter'],
|
222 |
-
value='Front',
|
223 |
-
multiselect=False,
|
224 |
-
scale=2
|
225 |
-
)
|
226 |
-
lighting_type = gr.Dropdown(
|
227 |
-
label='Lighting Type',
|
228 |
-
choices=['Bright and Even',
|
229 |
-
'Dramatic Shadows', 'Soft and Warm'],
|
230 |
-
value='Bright and Even',
|
231 |
-
multiselect=False,
|
232 |
-
scale=2
|
233 |
-
)
|
234 |
-
additional_details = gr.Textbox(
|
235 |
-
label='Additional Details', scale=2)
|
236 |
-
submit_prompt = gr.Button(
|
237 |
-
'Generate Image', scale=1, variant='primary')
|
238 |
-
|
239 |
-
with gr.Row(variant="panel"):
|
240 |
-
with gr.Column():
|
241 |
-
with gr.Row():
|
242 |
-
input_image = gr.Image(
|
243 |
-
label="Input Image",
|
244 |
-
image_mode="RGBA",
|
245 |
-
sources="upload",
|
246 |
-
type="pil",
|
247 |
-
elem_id="content_image",
|
248 |
-
)
|
249 |
-
processed_image = gr.Image(
|
250 |
-
label="Processed Image",
|
251 |
-
image_mode="RGBA",
|
252 |
-
type="pil",
|
253 |
-
interactive=False
|
254 |
-
)
|
255 |
-
with gr.Row():
|
256 |
-
submit_process = gr.Button(
|
257 |
-
"Remove Background", elem_id="process", variant="primary")
|
258 |
-
with gr.Row(variant="panel"):
|
259 |
-
gr.Examples(
|
260 |
-
examples=[os.path.join("examples", img_name) for img_name in sorted(
|
261 |
-
os.listdir("examples"))],
|
262 |
-
inputs=[input_image],
|
263 |
-
label="Examples",
|
264 |
-
cache_examples=False,
|
265 |
-
examples_per_page=16
|
266 |
-
)
|
267 |
|
268 |
with gr.Tab("Generate 3D Model"):
|
269 |
-
|
270 |
-
|
271 |
-
with gr.Column():
|
272 |
-
mv_show_images = gr.Image(
|
273 |
-
label="Generated Multi-views",
|
274 |
-
type="pil",
|
275 |
-
width=379,
|
276 |
-
interactive=False
|
277 |
-
)
|
278 |
-
with gr.Row():
|
279 |
-
with gr.Group():
|
280 |
-
sample_seed = gr.Number(
|
281 |
-
value=42, label="Seed Value", precision=0)
|
282 |
-
sample_steps = gr.Slider(
|
283 |
-
label="Sample Steps", minimum=30, maximum=75, value=75, step=5)
|
284 |
-
with gr.Row():
|
285 |
-
submit_mesh = gr.Button(
|
286 |
-
"Generate 3D Model", elem_id="generate", variant="primary")
|
287 |
-
with gr.Row():
|
288 |
-
with gr.Tab("OBJ"):
|
289 |
-
output_model_obj = gr.Model3D(
|
290 |
-
label="Output Model (OBJ Format)",
|
291 |
-
interactive=False,
|
292 |
-
)
|
293 |
-
gr.Markdown(
|
294 |
-
"Note: Downloaded .obj model will be flipped. Export .glb instead or manually flip it before usage.")
|
295 |
-
with gr.Tab("GLB"):
|
296 |
-
output_model_glb = gr.Model3D(
|
297 |
-
label="Output Model (GLB Format)",
|
298 |
-
interactive=False,
|
299 |
-
)
|
300 |
-
gr.Markdown(
|
301 |
-
"Note: The model shown here has a darker appearance. Download to get correct results.")
|
302 |
-
with gr.Row():
|
303 |
-
gr.Markdown(
|
304 |
-
'''Try a different <b>seed value</b> if the result is unsatisfying (Default: 42).''')
|
305 |
-
|
306 |
-
mv_images = gr.State()
|
307 |
-
|
308 |
-
submit_prompt.click(fn=generate_image, inputs=[subject, style, color_scheme, angle, lighting_type, additional_details], outputs=input_image).success(
|
309 |
-
fn=preprocess, inputs=[input_image], outputs=[processed_image]
|
310 |
-
)
|
311 |
-
submit_process.click(fn=check_input_image, inputs=[input_image]).success(
|
312 |
-
fn=preprocess, inputs=[input_image], outputs=[processed_image],
|
313 |
-
)
|
314 |
-
submit_mesh.click(fn=generate_mvs, inputs=[processed_image, sample_steps, sample_seed], outputs=[mv_images, mv_show_images]).success(
|
315 |
-
fn=make3d, inputs=[mv_images], outputs=[
|
316 |
-
output_model_obj, output_model_glb]
|
317 |
-
)
|
318 |
|
319 |
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
+
from launch.image_generation import image_generation_ui
|
4 |
+
from launch.model_generation import model_generation_ui
|
5 |
|
6 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
with gr.Blocks() as demo:
|
8 |
with gr.Group():
|
9 |
with gr.Tab("Generate Image and Remove Background"):
|
10 |
+
input_image, processed_image = image_generation_ui()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
with gr.Tab("Generate 3D Model"):
|
13 |
+
output_model_obj, output_model_glb = model_generation_ui(
|
14 |
+
processed_image)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
demo.launch()
|
launch/__init__.py
ADDED
File without changes
|
launch/image_generation.py
ADDED
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import gradio as gr
|
4 |
+
import rembg
|
5 |
+
import spaces
|
6 |
+
import torch
|
7 |
+
from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler
|
8 |
+
from huggingface_hub import hf_hub_download
|
9 |
+
from safetensors.torch import load_file
|
10 |
+
|
11 |
+
from src.utils.infer_util import (remove_background, resize_foreground)
|
12 |
+
|
13 |
+
|
14 |
+
# Load StableDiffusionXL model
|
15 |
+
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
16 |
+
repo = "ByteDance/SDXL-Lightning"
|
17 |
+
|
18 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(
|
19 |
+
base, torch_dtype=torch.float16, variant="fp16").to("cuda")
|
20 |
+
|
21 |
+
|
22 |
+
def generate_prompt(subject, style, color_scheme, angle, lighting_type, additional_details):
|
23 |
+
return f"A 3D cartoon render of {subject}, featuring the entire body and shape, on a transparent background. The style should be {style}, with {color_scheme} colors, emphasizing the essential features and lines. The pose should clearly showcase the full form of the {subject} from a {angle} perspective. Lighting is {lighting_type}, highlighting the volume and depth of the subject. {additional_details}. Output as a high-resolution PNG with no background."
|
24 |
+
|
25 |
+
|
26 |
+
@spaces.GPU
|
27 |
+
def generate_image(subject, style, color_scheme, angle, lighting_type, additional_details):
|
28 |
+
checkpoint = "sdxl_lightning_8step_unet.safetensors"
|
29 |
+
num_inference_steps = 8
|
30 |
+
|
31 |
+
pipe.scheduler = EulerDiscreteScheduler.from_config(
|
32 |
+
pipe.scheduler.config, timestep_spacing="trailing")
|
33 |
+
pipe.unet.load_state_dict(
|
34 |
+
load_file(hf_hub_download(repo, checkpoint), device="cuda"))
|
35 |
+
|
36 |
+
prompt = generate_prompt(subject, style, color_scheme,
|
37 |
+
angle, lighting_type, additional_details)
|
38 |
+
results = pipe(
|
39 |
+
prompt, num_inference_steps=num_inference_steps, guidance_scale=0)
|
40 |
+
return results.images[0]
|
41 |
+
|
42 |
+
|
43 |
+
def check_input_image(input_image):
|
44 |
+
if input_image is None:
|
45 |
+
raise gr.Error("No image selected!")
|
46 |
+
|
47 |
+
|
48 |
+
def preprocess(input_image):
|
49 |
+
rembg_session = rembg.new_session()
|
50 |
+
|
51 |
+
input_image = remove_background(input_image, rembg_session)
|
52 |
+
input_image = resize_foreground(input_image, 0.85)
|
53 |
+
|
54 |
+
return input_image
|
55 |
+
|
56 |
+
|
57 |
+
def image_generation_ui():
|
58 |
+
with gr.Row():
|
59 |
+
subject = gr.Textbox(label='Subject', scale=2)
|
60 |
+
style = gr.Dropdown(
|
61 |
+
label='Style',
|
62 |
+
choices=['Pixar-like', 'Disney-esque', 'Anime-inspired'],
|
63 |
+
value='Pixar-like',
|
64 |
+
multiselect=False,
|
65 |
+
scale=2
|
66 |
+
)
|
67 |
+
color_scheme = gr.Dropdown(
|
68 |
+
label='Color Scheme',
|
69 |
+
choices=['Vibrant', 'Pastel', 'Monochromatic', 'Black and White'],
|
70 |
+
value='Vibrant',
|
71 |
+
multiselect=False,
|
72 |
+
scale=2
|
73 |
+
)
|
74 |
+
angle = gr.Dropdown(
|
75 |
+
label='Angle',
|
76 |
+
choices=['Front', 'Side', 'Three-quarter'],
|
77 |
+
value='Front',
|
78 |
+
multiselect=False,
|
79 |
+
scale=2
|
80 |
+
)
|
81 |
+
lighting_type = gr.Dropdown(
|
82 |
+
label='Lighting Type',
|
83 |
+
choices=['Bright and Even', 'Dramatic Shadows', 'Soft and Warm'],
|
84 |
+
value='Bright and Even',
|
85 |
+
multiselect=False,
|
86 |
+
scale=2
|
87 |
+
)
|
88 |
+
additional_details = gr.Textbox(label='Additional Details', scale=2)
|
89 |
+
submit_prompt = gr.Button('Generate Image', scale=1, variant='primary')
|
90 |
+
|
91 |
+
with gr.Row(variant="panel"):
|
92 |
+
with gr.Column():
|
93 |
+
with gr.Row():
|
94 |
+
input_image = gr.Image(
|
95 |
+
label="Input Image",
|
96 |
+
image_mode="RGBA",
|
97 |
+
sources="upload",
|
98 |
+
type="pil",
|
99 |
+
elem_id="content_image",
|
100 |
+
)
|
101 |
+
processed_image = gr.Image(
|
102 |
+
label="Processed Image",
|
103 |
+
image_mode="RGBA",
|
104 |
+
type="pil",
|
105 |
+
interactive=False
|
106 |
+
)
|
107 |
+
with gr.Row():
|
108 |
+
submit_process = gr.Button(
|
109 |
+
"Remove Background", elem_id="process", variant="primary")
|
110 |
+
with gr.Row(variant="panel"):
|
111 |
+
gr.Examples(
|
112 |
+
examples=[os.path.join("examples", img_name)
|
113 |
+
for img_name in sorted(os.listdir("examples"))],
|
114 |
+
inputs=[input_image],
|
115 |
+
label="Examples",
|
116 |
+
cache_examples=False,
|
117 |
+
examples_per_page=16
|
118 |
+
)
|
119 |
+
|
120 |
+
submit_prompt.click(fn=generate_image, inputs=[subject, style, color_scheme, angle, lighting_type, additional_details], outputs=input_image).success(
|
121 |
+
fn=preprocess, inputs=[input_image], outputs=[processed_image]
|
122 |
+
)
|
123 |
+
submit_process.click(fn=check_input_image, inputs=[input_image]).success(
|
124 |
+
fn=preprocess, inputs=[input_image], outputs=[processed_image],
|
125 |
+
)
|
126 |
+
|
127 |
+
return input_image, processed_image
|
launch/model_generation.py
ADDED
@@ -0,0 +1,183 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import tempfile
|
3 |
+
|
4 |
+
import gradio as gr
|
5 |
+
import numpy as np
|
6 |
+
from launch.utils import find_cuda
|
7 |
+
import spaces
|
8 |
+
import torch
|
9 |
+
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler
|
10 |
+
from einops import rearrange
|
11 |
+
from huggingface_hub import hf_hub_download
|
12 |
+
from omegaconf import OmegaConf
|
13 |
+
from PIL import Image
|
14 |
+
from pytorch_lightning import seed_everything
|
15 |
+
from torchvision.transforms import v2
|
16 |
+
|
17 |
+
from src.utils.camera_util import (FOV_to_intrinsics, get_circular_camera_poses,
|
18 |
+
get_zero123plus_input_cameras)
|
19 |
+
from src.utils.mesh_util import save_glb, save_obj
|
20 |
+
from src.utils.train_util import instantiate_from_config
|
21 |
+
|
22 |
+
# Configuration
|
23 |
+
cuda_path = find_cuda()
|
24 |
+
config_path = 'configs/instant-mesh-large.yaml'
|
25 |
+
config = OmegaConf.load(config_path)
|
26 |
+
config_name = os.path.basename(config_path).replace('.yaml', '')
|
27 |
+
model_config = config.model_config
|
28 |
+
infer_config = config.infer_config
|
29 |
+
|
30 |
+
IS_FLEXICUBES = config_name.startswith('instant-mesh')
|
31 |
+
device = torch.device('cuda')
|
32 |
+
|
33 |
+
# Load diffusion model
|
34 |
+
print('Loading diffusion model ...')
|
35 |
+
pipeline = DiffusionPipeline.from_pretrained(
|
36 |
+
"sudo-ai/zero123plus-v1.2",
|
37 |
+
custom_pipeline="zero123plus",
|
38 |
+
torch_dtype=torch.float16,
|
39 |
+
)
|
40 |
+
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
|
41 |
+
pipeline.scheduler.config, timestep_spacing='trailing'
|
42 |
+
)
|
43 |
+
|
44 |
+
unet_ckpt_path = hf_hub_download(
|
45 |
+
repo_id="TencentARC/InstantMesh", filename="diffusion_pytorch_model.bin", repo_type="model")
|
46 |
+
state_dict = torch.load(unet_ckpt_path, map_location='cpu')
|
47 |
+
pipeline.unet.load_state_dict(state_dict, strict=True)
|
48 |
+
|
49 |
+
pipeline = pipeline.to(device)
|
50 |
+
|
51 |
+
# Load reconstruction model
|
52 |
+
print('Loading reconstruction model ...')
|
53 |
+
model_ckpt_path = hf_hub_download(
|
54 |
+
repo_id="TencentARC/InstantMesh", filename="instant_mesh_large.ckpt", repo_type="model")
|
55 |
+
model = instantiate_from_config(model_config)
|
56 |
+
state_dict = torch.load(model_ckpt_path, map_location='cpu')['state_dict']
|
57 |
+
state_dict = {k[14:]: v for k, v in state_dict.items() if k.startswith(
|
58 |
+
'lrm_generator.') and 'source_camera' not in k}
|
59 |
+
model.load_state_dict(state_dict, strict=True)
|
60 |
+
|
61 |
+
model = model.to(device)
|
62 |
+
|
63 |
+
|
64 |
+
def get_render_cameras(batch_size=1, M=120, radius=2.5, elevation=10.0, is_flexicubes=False):
|
65 |
+
c2ws = get_circular_camera_poses(M=M, radius=radius, elevation=elevation)
|
66 |
+
if is_flexicubes:
|
67 |
+
cameras = torch.linalg.inv(c2ws)
|
68 |
+
cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1, 1)
|
69 |
+
else:
|
70 |
+
extrinsics = c2ws.flatten(-2)
|
71 |
+
intrinsics = FOV_to_intrinsics(50.0).unsqueeze(
|
72 |
+
0).repeat(M, 1, 1).float().flatten(-2)
|
73 |
+
cameras = torch.cat([extrinsics, intrinsics], dim=-1)
|
74 |
+
cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1)
|
75 |
+
return cameras
|
76 |
+
|
77 |
+
|
78 |
+
@spaces.GPU
|
79 |
+
def generate_mvs(input_image, sample_steps, sample_seed):
|
80 |
+
seed_everything(sample_seed)
|
81 |
+
|
82 |
+
z123_image = pipeline(
|
83 |
+
input_image, num_inference_steps=sample_steps).images[0]
|
84 |
+
|
85 |
+
show_image = np.asarray(z123_image, dtype=np.uint8)
|
86 |
+
show_image = torch.from_numpy(show_image)
|
87 |
+
show_image = rearrange(
|
88 |
+
show_image, '(n h) (m w) c -> (n m) h w c', n=3, m=2)
|
89 |
+
show_image = rearrange(
|
90 |
+
show_image, '(n m) h w c -> (n h) (m w) c', n=2, m=3)
|
91 |
+
show_image = Image.fromarray(show_image.numpy())
|
92 |
+
|
93 |
+
return z123_image, show_image
|
94 |
+
|
95 |
+
|
96 |
+
@spaces.GPU
|
97 |
+
def make3d(images):
|
98 |
+
global model
|
99 |
+
if IS_FLEXICUBES:
|
100 |
+
model.init_flexicubes_geometry(device, use_renderer=False)
|
101 |
+
model = model.eval()
|
102 |
+
|
103 |
+
images = np.asarray(images, dtype=np.float32) / 255.0
|
104 |
+
images = torch.from_numpy(images).permute(2, 0, 1).contiguous().float()
|
105 |
+
images = rearrange(images, 'c (n h) (m w) -> (n m) c h w', n=3, m=2)
|
106 |
+
|
107 |
+
input_cameras = get_zero123plus_input_cameras(
|
108 |
+
batch_size=1, radius=4.0).to(device)
|
109 |
+
render_cameras = get_render_cameras(
|
110 |
+
batch_size=1, radius=2.5, is_flexicubes=IS_FLEXICUBES).to(device)
|
111 |
+
|
112 |
+
images = images.unsqueeze(0).to(device)
|
113 |
+
images = v2.functional.resize(
|
114 |
+
images, (320, 320), interpolation=3, antialias=True).clamp(0, 1)
|
115 |
+
|
116 |
+
mesh_fpath = tempfile.NamedTemporaryFile(suffix=f".obj", delete=False).name
|
117 |
+
print(mesh_fpath)
|
118 |
+
mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
|
119 |
+
mesh_dirname = os.path.dirname(mesh_fpath)
|
120 |
+
mesh_glb_fpath = os.path.join(mesh_dirname, f"{mesh_basename}.glb")
|
121 |
+
|
122 |
+
with torch.no_grad():
|
123 |
+
planes = model.forward_planes(images, input_cameras)
|
124 |
+
mesh_out = model.extract_mesh(
|
125 |
+
planes, use_texture_map=False, **infer_config)
|
126 |
+
|
127 |
+
vertices, faces, vertex_colors = mesh_out
|
128 |
+
vertices = vertices[:, [1, 2, 0]]
|
129 |
+
|
130 |
+
save_glb(vertices, faces, vertex_colors, mesh_glb_fpath)
|
131 |
+
save_obj(vertices, faces, vertex_colors, mesh_fpath)
|
132 |
+
|
133 |
+
print(f"Mesh saved to {mesh_fpath}")
|
134 |
+
|
135 |
+
return mesh_fpath, mesh_glb_fpath
|
136 |
+
|
137 |
+
|
138 |
+
def model_generation_ui(processed_image):
|
139 |
+
with gr.Column():
|
140 |
+
with gr.Row():
|
141 |
+
with gr.Column():
|
142 |
+
mv_show_images = gr.Image(
|
143 |
+
label="Generated Multi-views",
|
144 |
+
type="pil",
|
145 |
+
width=379,
|
146 |
+
interactive=False
|
147 |
+
)
|
148 |
+
with gr.Row():
|
149 |
+
with gr.Group():
|
150 |
+
sample_seed = gr.Number(
|
151 |
+
value=42, label="Seed Value", precision=0)
|
152 |
+
sample_steps = gr.Slider(
|
153 |
+
label="Sample Steps", minimum=30, maximum=75, value=75, step=5)
|
154 |
+
with gr.Row():
|
155 |
+
submit_mesh = gr.Button(
|
156 |
+
"Generate 3D Model", elem_id="generate", variant="primary")
|
157 |
+
with gr.Row():
|
158 |
+
with gr.Tab("OBJ"):
|
159 |
+
output_model_obj = gr.Model3D(
|
160 |
+
label="Output Model (OBJ Format)",
|
161 |
+
interactive=False,
|
162 |
+
)
|
163 |
+
gr.Markdown(
|
164 |
+
"Note: Downloaded .obj model will be flipped. Export .glb instead or manually flip it before usage.")
|
165 |
+
with gr.Tab("GLB"):
|
166 |
+
output_model_glb = gr.Model3D(
|
167 |
+
label="Output Model (GLB Format)",
|
168 |
+
interactive=False,
|
169 |
+
)
|
170 |
+
gr.Markdown(
|
171 |
+
"Note: The model shown here has a darker appearance. Download to get correct results.")
|
172 |
+
with gr.Row():
|
173 |
+
gr.Markdown(
|
174 |
+
'''Try a different <b>seed value</b> if the result is unsatisfying (Default: 42).''')
|
175 |
+
|
176 |
+
mv_images = gr.State()
|
177 |
+
|
178 |
+
submit_mesh.click(fn=generate_mvs, inputs=[processed_image, sample_steps, sample_seed], outputs=[mv_images, mv_show_images]).success(
|
179 |
+
fn=make3d, inputs=[mv_images], outputs=[
|
180 |
+
output_model_obj, output_model_glb]
|
181 |
+
)
|
182 |
+
|
183 |
+
return output_model_obj, output_model_glb
|
launch/utils.py
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import shutil
|
3 |
+
|
4 |
+
def find_cuda():
|
5 |
+
cuda_home = os.environ.get('CUDA_HOME') or os.environ.get('CUDA_PATH')
|
6 |
+
if cuda_home and os.path.exists(cuda_home):
|
7 |
+
return cuda_home
|
8 |
+
|
9 |
+
nvcc_path = shutil.which('nvcc')
|
10 |
+
if nvcc_path:
|
11 |
+
cuda_path = os.path.dirname(os.path.dirname(nvcc_path))
|
12 |
+
return cuda_path
|
13 |
+
|
14 |
+
return None
|