Spaces:
Runtime error
Runtime error
Commit
·
5d94b0f
1
Parent(s):
4533120
update space
Browse files
app.py
CHANGED
|
@@ -62,158 +62,6 @@ class Blocks(gr.Blocks):
|
|
| 62 |
|
| 63 |
return config
|
| 64 |
|
| 65 |
-
@torch.no_grad()
|
| 66 |
-
def optimize_all(xm, models, initial_noise, noise_start_t, diffusion, latent_model, device, prompt, instruction, rand_seed):
|
| 67 |
-
state = {}
|
| 68 |
-
out_gen_1, out_gen_2, out_gen_3, out_gen_4, state = generate_3d_with_shap_e(xm, diffusion, latent_model, device, prompt, rand_seed, state)
|
| 69 |
-
edited_1, edited_2, edited_3, edited_4, state = _3d_editing(xm, models, diffusion, initial_noise, noise_start_t, device, instruction, rand_seed, state)
|
| 70 |
-
print(state)
|
| 71 |
-
return out_gen_1, out_gen_2, out_gen_3, out_gen_4, edited_1, edited_2, edited_3, edited_4
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
@spaces.GPU()
|
| 75 |
-
@torch.no_grad()
|
| 76 |
-
def generate_3d_with_shap_e(xm, diffusion, latent_model, device, prompt, rand_seed, state):
|
| 77 |
-
print("Check if I can use partial")
|
| 78 |
-
set_seed(rand_seed)
|
| 79 |
-
batch_size = 4
|
| 80 |
-
guidance_scale = 15.0
|
| 81 |
-
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.0]).to(device)
|
| 82 |
-
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1.0]).to(device)
|
| 83 |
-
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 84 |
-
|
| 85 |
-
print("prompt: ", prompt, "rand_seed: ", rand_seed, "state:", state)
|
| 86 |
-
latents = sample_latents(
|
| 87 |
-
batch_size=batch_size,
|
| 88 |
-
model=latent_model,
|
| 89 |
-
diffusion=diffusion,
|
| 90 |
-
guidance_scale=guidance_scale,
|
| 91 |
-
model_kwargs=dict(texts=[prompt] * batch_size),
|
| 92 |
-
progress=True,
|
| 93 |
-
clip_denoised=True,
|
| 94 |
-
use_fp16=True,
|
| 95 |
-
use_karras=True,
|
| 96 |
-
karras_steps=64,
|
| 97 |
-
sigma_min=1e-3,
|
| 98 |
-
sigma_max=160,
|
| 99 |
-
s_churn=0,
|
| 100 |
-
)
|
| 101 |
-
prompt_hash = str(hashlib.sha256((prompt + '_' + str(rand_seed)).encode('utf-8')).hexdigest())
|
| 102 |
-
mesh_path = []
|
| 103 |
-
output_path = './logs'
|
| 104 |
-
os.makedirs(os.path.join(output_path, 'source'), exist_ok=True)
|
| 105 |
-
state['latent'] = []
|
| 106 |
-
state['prompt'] = prompt
|
| 107 |
-
state['rand_seed_1'] = rand_seed
|
| 108 |
-
for i, latent in enumerate(latents):
|
| 109 |
-
|
| 110 |
-
output_path_tmp = os.path.join(output_path, 'source', '{}_{}.obj'.format(prompt_hash, i))
|
| 111 |
-
t_obj = decode_latent_mesh(xm, latent).tri_mesh()
|
| 112 |
-
with open(output_path_tmp, 'w') as f:
|
| 113 |
-
t_obj.write_obj(f)
|
| 114 |
-
|
| 115 |
-
mesh = trimesh.load_mesh(output_path_tmp)
|
| 116 |
-
angle = np.radians(180)
|
| 117 |
-
axis = [0, 1, 0]
|
| 118 |
-
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 119 |
-
mesh.apply_transform(rotation_matrix)
|
| 120 |
-
angle = np.radians(90)
|
| 121 |
-
axis = [1, 0, 0]
|
| 122 |
-
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 123 |
-
mesh.apply_transform(rotation_matrix)
|
| 124 |
-
output_path_tmp = os.path.join(output_path, 'source', '{}_{}.obj'.format(prompt_hash, i))
|
| 125 |
-
mesh.export(output_path_tmp)
|
| 126 |
-
state['latent'].append(latent.clone().detach().cpu())
|
| 127 |
-
mesh_path.append(output_path_tmp)
|
| 128 |
-
del latents
|
| 129 |
-
return mesh_path[0], mesh_path[1], mesh_path[2], mesh_path[3], state
|
| 130 |
-
|
| 131 |
-
@spaces.GPU()
|
| 132 |
-
@torch.no_grad()
|
| 133 |
-
def _3d_editing(xm, models, diffusion, initial_noise, start_t, device, instruction, rand_seed, state):
|
| 134 |
-
set_seed(rand_seed)
|
| 135 |
-
mesh_path = []
|
| 136 |
-
prompt = state['prompt']
|
| 137 |
-
rand_seed_1 = state['rand_seed_1']
|
| 138 |
-
print("prompt: ", prompt, "rand_seed: ", rand_seed, "instruction:", instruction, "state:", state)
|
| 139 |
-
prompt_hash = str(hashlib.sha256((prompt + '_' + str(rand_seed_1) + '_' + instruction + '_' + str(rand_seed)).encode('utf-8')).hexdigest())
|
| 140 |
-
if 'santa' in instruction:
|
| 141 |
-
e_type = 'santa_hat'
|
| 142 |
-
elif 'rainbow' in instruction:
|
| 143 |
-
e_type = 'rainbow'
|
| 144 |
-
elif 'gold' in instruction:
|
| 145 |
-
e_type = 'golden'
|
| 146 |
-
elif 'lego' in instruction:
|
| 147 |
-
e_type = 'lego'
|
| 148 |
-
elif 'wooden' in instruction:
|
| 149 |
-
e_type = 'wooden'
|
| 150 |
-
elif 'cyber' in instruction:
|
| 151 |
-
e_type = 'cyber'
|
| 152 |
-
|
| 153 |
-
model = load_model('text300M', device=device)
|
| 154 |
-
with torch.no_grad():
|
| 155 |
-
new_proj = nn.Linear(1024 * 2, 1024, device=device, dtype=model.wrapped.input_proj.weight.dtype)
|
| 156 |
-
new_proj.weight = nn.Parameter(torch.zeros_like(new_proj.weight))
|
| 157 |
-
new_proj.weight[:, :1024].copy_(model.wrapped.input_proj.weight) #
|
| 158 |
-
new_proj.bias = nn.Parameter(torch.zeros_like(new_proj.bias))
|
| 159 |
-
new_proj.bias[:1024].copy_(model.wrapped.input_proj.bias)
|
| 160 |
-
model.wrapped.input_proj = new_proj
|
| 161 |
-
|
| 162 |
-
ckp = torch.load(hf_hub_download(repo_id='silentchen/Shap_Editor', subfolder='single', filename='{}.pt'.format(e_type)), map_location='cpu')
|
| 163 |
-
model.load_state_dict(ckp['model'])
|
| 164 |
-
|
| 165 |
-
noise_initial = initial_noise[e_type].to(device)
|
| 166 |
-
noise_start_t = start_t[e_type]
|
| 167 |
-
general_save_path = './logs/edited'
|
| 168 |
-
os.makedirs(general_save_path, exist_ok=True)
|
| 169 |
-
for i, latent in enumerate(state['latent']):
|
| 170 |
-
latent = latent.to(device)
|
| 171 |
-
text_embeddings_clip = model.cached_model_kwargs(1, dict(texts=[instruction]))
|
| 172 |
-
print("shape of latent: ", latent.clone().unsqueeze(0).shape, "instruction: ", instruction)
|
| 173 |
-
ref_latent = latent.clone().unsqueeze(0)
|
| 174 |
-
t_1 = torch.randint(noise_start_t, noise_start_t + 1, (1,), device=device).long()
|
| 175 |
-
|
| 176 |
-
noise_input = diffusion.q_sample(ref_latent, t_1, noise=noise_initial)
|
| 177 |
-
out_1 = diffusion.p_mean_variance(model, noise_input, t_1, clip_denoised=True,
|
| 178 |
-
model_kwargs=text_embeddings_clip,
|
| 179 |
-
condition_latents=ref_latent)
|
| 180 |
-
|
| 181 |
-
updated_latents = out_1['pred_xstart']
|
| 182 |
-
|
| 183 |
-
if 'santa' in instruction:
|
| 184 |
-
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.25]).to(device)
|
| 185 |
-
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1]).to(device)
|
| 186 |
-
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 187 |
-
|
| 188 |
-
else:
|
| 189 |
-
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.0]).to(device)
|
| 190 |
-
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1.0]).to(device)
|
| 191 |
-
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 192 |
-
|
| 193 |
-
for latent_idx, updated_latent in enumerate(updated_latents):
|
| 194 |
-
output_path = os.path.join(general_save_path, '{}_{}.obj'.format(prompt_hash, i))
|
| 195 |
-
|
| 196 |
-
t = decode_latent_mesh(xm, updated_latent).tri_mesh()
|
| 197 |
-
with open(output_path, 'w') as f:
|
| 198 |
-
t.write_obj(f)
|
| 199 |
-
mesh = trimesh.load_mesh(output_path)
|
| 200 |
-
|
| 201 |
-
angle = np.radians(180)
|
| 202 |
-
axis = [0, 1, 0]
|
| 203 |
-
|
| 204 |
-
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 205 |
-
mesh.apply_transform(rotation_matrix)
|
| 206 |
-
angle = np.radians(90)
|
| 207 |
-
axis = [1, 0, 0]
|
| 208 |
-
|
| 209 |
-
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 210 |
-
mesh.apply_transform(rotation_matrix)
|
| 211 |
-
|
| 212 |
-
output_path = os.path.join(general_save_path, '{}_{}.obj'.format(prompt_hash, i))
|
| 213 |
-
mesh.export(output_path)
|
| 214 |
-
mesh_path.append(output_path)
|
| 215 |
-
return mesh_path[0], mesh_path[1], mesh_path[2], mesh_path[3], state
|
| 216 |
-
|
| 217 |
def main():
|
| 218 |
|
| 219 |
css = """
|
|
@@ -320,6 +168,161 @@ def main():
|
|
| 320 |
initial_noise[editing_type] = noise_initial
|
| 321 |
noise_start_t[editing_type] = ckp['t_start']
|
| 322 |
models[editing_type] = tmp_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 323 |
|
| 324 |
del models
|
| 325 |
models = None
|
|
@@ -388,13 +391,13 @@ def main():
|
|
| 388 |
rand_seed = gr.Slider(minimum=0, maximum=1000, step=1, value=445, label="Random seed")
|
| 389 |
|
| 390 |
gen_btn.click(
|
| 391 |
-
fn=
|
| 392 |
inputs=[prompt, rand_seed, state],
|
| 393 |
outputs=[out_gen_1, out_gen_2, out_gen_3, out_gen_4, state],
|
| 394 |
queue=False)
|
| 395 |
|
| 396 |
apply_btn.click(
|
| 397 |
-
fn=
|
| 398 |
inputs=[
|
| 399 |
editing_choice[0], rand_seed, state
|
| 400 |
],
|
|
@@ -416,7 +419,7 @@ def main():
|
|
| 416 |
],
|
| 417 |
inputs=[prompt, editing_choice[0], rand_seed],
|
| 418 |
outputs=[out_gen_1, out_gen_2, out_gen_3, out_gen_4, edited_1, edited_2, edited_3, edited_4],
|
| 419 |
-
fn=
|
| 420 |
cache_examples=True,
|
| 421 |
)
|
| 422 |
|
|
|
|
| 62 |
|
| 63 |
return config
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
def main():
|
| 66 |
|
| 67 |
css = """
|
|
|
|
| 168 |
initial_noise[editing_type] = noise_initial
|
| 169 |
noise_start_t[editing_type] = ckp['t_start']
|
| 170 |
models[editing_type] = tmp_model
|
| 171 |
+
@torch.no_grad()
|
| 172 |
+
def optimize_all(prompt, instruction,
|
| 173 |
+
rand_seed):
|
| 174 |
+
print("Optimizing all")
|
| 175 |
+
state = {}
|
| 176 |
+
out_gen_1, out_gen_2, out_gen_3, out_gen_4, state = generate_3d_with_shap_e(prompt, rand_seed, state)
|
| 177 |
+
edited_1, edited_2, edited_3, edited_4, state = _3d_editing(instruction, rand_seed, state)
|
| 178 |
+
print(state)
|
| 179 |
+
return out_gen_1, out_gen_2, out_gen_3, out_gen_4, edited_1, edited_2, edited_3, edited_4
|
| 180 |
+
|
| 181 |
+
@spaces.GPU()
|
| 182 |
+
@torch.no_grad()
|
| 183 |
+
def generate_3d_with_shap_e(prompt, rand_seed, state):
|
| 184 |
+
print("Check if I can use partial")
|
| 185 |
+
set_seed(rand_seed)
|
| 186 |
+
batch_size = 4
|
| 187 |
+
guidance_scale = 15.0
|
| 188 |
+
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.0]).to(device)
|
| 189 |
+
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1.0]).to(device)
|
| 190 |
+
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 191 |
+
|
| 192 |
+
print("prompt: ", prompt, "rand_seed: ", rand_seed, "state:", state)
|
| 193 |
+
latents = sample_latents(
|
| 194 |
+
batch_size=batch_size,
|
| 195 |
+
model=latent_model,
|
| 196 |
+
diffusion=diffusion,
|
| 197 |
+
guidance_scale=guidance_scale,
|
| 198 |
+
model_kwargs=dict(texts=[prompt] * batch_size),
|
| 199 |
+
progress=True,
|
| 200 |
+
clip_denoised=True,
|
| 201 |
+
use_fp16=True,
|
| 202 |
+
use_karras=True,
|
| 203 |
+
karras_steps=64,
|
| 204 |
+
sigma_min=1e-3,
|
| 205 |
+
sigma_max=160,
|
| 206 |
+
s_churn=0,
|
| 207 |
+
)
|
| 208 |
+
prompt_hash = str(hashlib.sha256((prompt + '_' + str(rand_seed)).encode('utf-8')).hexdigest())
|
| 209 |
+
mesh_path = []
|
| 210 |
+
output_path = './logs'
|
| 211 |
+
os.makedirs(os.path.join(output_path, 'source'), exist_ok=True)
|
| 212 |
+
state['latent'] = []
|
| 213 |
+
state['prompt'] = prompt
|
| 214 |
+
state['rand_seed_1'] = rand_seed
|
| 215 |
+
for i, latent in enumerate(latents):
|
| 216 |
+
output_path_tmp = os.path.join(output_path, 'source', '{}_{}.obj'.format(prompt_hash, i))
|
| 217 |
+
t_obj = decode_latent_mesh(xm, latent).tri_mesh()
|
| 218 |
+
with open(output_path_tmp, 'w') as f:
|
| 219 |
+
t_obj.write_obj(f)
|
| 220 |
+
|
| 221 |
+
mesh = trimesh.load_mesh(output_path_tmp)
|
| 222 |
+
angle = np.radians(180)
|
| 223 |
+
axis = [0, 1, 0]
|
| 224 |
+
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 225 |
+
mesh.apply_transform(rotation_matrix)
|
| 226 |
+
angle = np.radians(90)
|
| 227 |
+
axis = [1, 0, 0]
|
| 228 |
+
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 229 |
+
mesh.apply_transform(rotation_matrix)
|
| 230 |
+
output_path_tmp = os.path.join(output_path, 'source', '{}_{}.obj'.format(prompt_hash, i))
|
| 231 |
+
mesh.export(output_path_tmp)
|
| 232 |
+
state['latent'].append(latent.clone().detach().cpu())
|
| 233 |
+
mesh_path.append(output_path_tmp)
|
| 234 |
+
del latents
|
| 235 |
+
return mesh_path[0], mesh_path[1], mesh_path[2], mesh_path[3], state
|
| 236 |
+
|
| 237 |
+
@spaces.GPU()
|
| 238 |
+
@torch.no_grad()
|
| 239 |
+
def _3d_editing(instruction, rand_seed, state):
|
| 240 |
+
set_seed(rand_seed)
|
| 241 |
+
mesh_path = []
|
| 242 |
+
prompt = state['prompt']
|
| 243 |
+
rand_seed_1 = state['rand_seed_1']
|
| 244 |
+
print("prompt: ", prompt, "rand_seed: ", rand_seed, "instruction:", instruction, "state:", state)
|
| 245 |
+
prompt_hash = str(hashlib.sha256(
|
| 246 |
+
(prompt + '_' + str(rand_seed_1) + '_' + instruction + '_' + str(rand_seed)).encode('utf-8')).hexdigest())
|
| 247 |
+
if 'santa' in instruction:
|
| 248 |
+
e_type = 'santa_hat'
|
| 249 |
+
elif 'rainbow' in instruction:
|
| 250 |
+
e_type = 'rainbow'
|
| 251 |
+
elif 'gold' in instruction:
|
| 252 |
+
e_type = 'golden'
|
| 253 |
+
elif 'lego' in instruction:
|
| 254 |
+
e_type = 'lego'
|
| 255 |
+
elif 'wooden' in instruction:
|
| 256 |
+
e_type = 'wooden'
|
| 257 |
+
elif 'cyber' in instruction:
|
| 258 |
+
e_type = 'cyber'
|
| 259 |
+
|
| 260 |
+
model = load_model('text300M', device=device)
|
| 261 |
+
with torch.no_grad():
|
| 262 |
+
new_proj = nn.Linear(1024 * 2, 1024, device=device, dtype=model.wrapped.input_proj.weight.dtype)
|
| 263 |
+
new_proj.weight = nn.Parameter(torch.zeros_like(new_proj.weight))
|
| 264 |
+
new_proj.weight[:, :1024].copy_(model.wrapped.input_proj.weight) #
|
| 265 |
+
new_proj.bias = nn.Parameter(torch.zeros_like(new_proj.bias))
|
| 266 |
+
new_proj.bias[:1024].copy_(model.wrapped.input_proj.bias)
|
| 267 |
+
model.wrapped.input_proj = new_proj
|
| 268 |
+
|
| 269 |
+
ckp = torch.load(
|
| 270 |
+
hf_hub_download(repo_id='silentchen/Shap_Editor', subfolder='single', filename='{}.pt'.format(e_type)),
|
| 271 |
+
map_location='cpu')
|
| 272 |
+
model.load_state_dict(ckp['model'])
|
| 273 |
+
|
| 274 |
+
noise_initial = initial_noise[e_type].to(device)
|
| 275 |
+
noise_start_t = noise_start_t[e_type]
|
| 276 |
+
general_save_path = './logs/edited'
|
| 277 |
+
os.makedirs(general_save_path, exist_ok=True)
|
| 278 |
+
for i, latent in enumerate(state['latent']):
|
| 279 |
+
latent = latent.to(device)
|
| 280 |
+
text_embeddings_clip = model.cached_model_kwargs(1, dict(texts=[instruction]))
|
| 281 |
+
print("shape of latent: ", latent.clone().unsqueeze(0).shape, "instruction: ", instruction)
|
| 282 |
+
ref_latent = latent.clone().unsqueeze(0)
|
| 283 |
+
t_1 = torch.randint(noise_start_t, noise_start_t + 1, (1,), device=device).long()
|
| 284 |
+
|
| 285 |
+
noise_input = diffusion.q_sample(ref_latent, t_1, noise=noise_initial)
|
| 286 |
+
out_1 = diffusion.p_mean_variance(model, noise_input, t_1, clip_denoised=True,
|
| 287 |
+
model_kwargs=text_embeddings_clip,
|
| 288 |
+
condition_latents=ref_latent)
|
| 289 |
+
|
| 290 |
+
updated_latents = out_1['pred_xstart']
|
| 291 |
+
|
| 292 |
+
if 'santa' in instruction:
|
| 293 |
+
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.25]).to(device)
|
| 294 |
+
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1]).to(device)
|
| 295 |
+
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 296 |
+
|
| 297 |
+
else:
|
| 298 |
+
xm.renderer.volume.bbox_max = torch.tensor([1.0, 1.0, 1.0]).to(device)
|
| 299 |
+
xm.renderer.volume.bbox_min = torch.tensor([-1.0, -1.0, -1.0]).to(device)
|
| 300 |
+
xm.renderer.volume.bbox = torch.stack([xm.renderer.volume.bbox_min, xm.renderer.volume.bbox_max])
|
| 301 |
+
|
| 302 |
+
for latent_idx, updated_latent in enumerate(updated_latents):
|
| 303 |
+
output_path = os.path.join(general_save_path, '{}_{}.obj'.format(prompt_hash, i))
|
| 304 |
+
|
| 305 |
+
t = decode_latent_mesh(xm, updated_latent).tri_mesh()
|
| 306 |
+
with open(output_path, 'w') as f:
|
| 307 |
+
t.write_obj(f)
|
| 308 |
+
mesh = trimesh.load_mesh(output_path)
|
| 309 |
+
|
| 310 |
+
angle = np.radians(180)
|
| 311 |
+
axis = [0, 1, 0]
|
| 312 |
+
|
| 313 |
+
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 314 |
+
mesh.apply_transform(rotation_matrix)
|
| 315 |
+
angle = np.radians(90)
|
| 316 |
+
axis = [1, 0, 0]
|
| 317 |
+
|
| 318 |
+
rotation_matrix = trimesh.transformations.rotation_matrix(angle, axis)
|
| 319 |
+
mesh.apply_transform(rotation_matrix)
|
| 320 |
+
|
| 321 |
+
output_path = os.path.join(general_save_path, '{}_{}.obj'.format(prompt_hash, i))
|
| 322 |
+
mesh.export(output_path)
|
| 323 |
+
mesh_path.append(output_path)
|
| 324 |
+
return mesh_path[0], mesh_path[1], mesh_path[2], mesh_path[3], state
|
| 325 |
+
|
| 326 |
|
| 327 |
del models
|
| 328 |
models = None
|
|
|
|
| 391 |
rand_seed = gr.Slider(minimum=0, maximum=1000, step=1, value=445, label="Random seed")
|
| 392 |
|
| 393 |
gen_btn.click(
|
| 394 |
+
fn=generate_3d_with_shap_e,
|
| 395 |
inputs=[prompt, rand_seed, state],
|
| 396 |
outputs=[out_gen_1, out_gen_2, out_gen_3, out_gen_4, state],
|
| 397 |
queue=False)
|
| 398 |
|
| 399 |
apply_btn.click(
|
| 400 |
+
fn=_3d_editing,
|
| 401 |
inputs=[
|
| 402 |
editing_choice[0], rand_seed, state
|
| 403 |
],
|
|
|
|
| 419 |
],
|
| 420 |
inputs=[prompt, editing_choice[0], rand_seed],
|
| 421 |
outputs=[out_gen_1, out_gen_2, out_gen_3, out_gen_4, edited_1, edited_2, edited_3, edited_4],
|
| 422 |
+
fn=optimize_all,
|
| 423 |
cache_examples=True,
|
| 424 |
)
|
| 425 |
|