Spaces:
Running
Running
Commit
·
db894f7
1
Parent(s):
690b53e
fix gradio image compression
Browse files- app.py +15 -12
- trellis/pipelines/trellis_image_to_3d.py +3 -1
app.py
CHANGED
|
@@ -19,7 +19,7 @@ from trellis.utils import render_utils, postprocessing_utils
|
|
| 19 |
MAX_SEED = np.iinfo(np.int32).max
|
| 20 |
|
| 21 |
|
| 22 |
-
def preprocess_image(image: Image.Image) -> Image.Image:
|
| 23 |
"""
|
| 24 |
Preprocess the input image.
|
| 25 |
|
|
@@ -27,9 +27,11 @@ def preprocess_image(image: Image.Image) -> Image.Image:
|
|
| 27 |
image (Image.Image): The input image.
|
| 28 |
|
| 29 |
Returns:
|
|
|
|
| 30 |
Image.Image: The preprocessed image.
|
| 31 |
"""
|
| 32 |
-
|
|
|
|
| 33 |
|
| 34 |
|
| 35 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult, model_id: str) -> dict:
|
|
@@ -74,12 +76,12 @@ def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
|
|
| 74 |
|
| 75 |
|
| 76 |
@spaces.GPU
|
| 77 |
-
def image_to_3d(image:
|
| 78 |
"""
|
| 79 |
Convert an image to a 3D model.
|
| 80 |
|
| 81 |
Args:
|
| 82 |
-
image (
|
| 83 |
seed (int): The random seed.
|
| 84 |
randomize_seed (bool): Whether to randomize the seed.
|
| 85 |
ss_guidance_strength (float): The guidance strength for sparse structure generation.
|
|
@@ -93,9 +95,9 @@ def image_to_3d(image: Image.Image, seed: int, randomize_seed: bool, ss_guidance
|
|
| 93 |
"""
|
| 94 |
if randomize_seed:
|
| 95 |
seed = np.random.randint(0, MAX_SEED)
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
formats=["gaussian", "mesh"],
|
| 100 |
preprocess_image=False,
|
| 101 |
sparse_structure_sampler_params={
|
|
@@ -181,6 +183,9 @@ with gr.Blocks() as demo:
|
|
| 181 |
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
| 182 |
model_output = LitModel3D(label="Extracted GLB", exposure=20.0, height=300)
|
| 183 |
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
|
|
|
|
|
|
|
|
|
| 184 |
|
| 185 |
# Example images at the bottom of the page
|
| 186 |
with gr.Row():
|
|
@@ -191,23 +196,21 @@ with gr.Blocks() as demo:
|
|
| 191 |
],
|
| 192 |
inputs=[image_prompt],
|
| 193 |
fn=lambda image: preprocess_image(image),
|
| 194 |
-
outputs=[image_prompt],
|
| 195 |
run_on_click=True,
|
| 196 |
examples_per_page=64,
|
| 197 |
)
|
| 198 |
|
| 199 |
-
model = gr.State()
|
| 200 |
-
|
| 201 |
# Handlers
|
| 202 |
image_prompt.upload(
|
| 203 |
preprocess_image,
|
| 204 |
inputs=[image_prompt],
|
| 205 |
-
outputs=[image_prompt],
|
| 206 |
)
|
| 207 |
|
| 208 |
generate_btn.click(
|
| 209 |
image_to_3d,
|
| 210 |
-
inputs=[
|
| 211 |
outputs=[model, video_output],
|
| 212 |
).then(
|
| 213 |
activate_button,
|
|
|
|
| 19 |
MAX_SEED = np.iinfo(np.int32).max
|
| 20 |
|
| 21 |
|
| 22 |
+
def preprocess_image(image: Image.Image) -> Tuple[np.array, Image.Image]:
|
| 23 |
"""
|
| 24 |
Preprocess the input image.
|
| 25 |
|
|
|
|
| 27 |
image (Image.Image): The input image.
|
| 28 |
|
| 29 |
Returns:
|
| 30 |
+
np.array: The preprocessed image.
|
| 31 |
Image.Image: The preprocessed image.
|
| 32 |
"""
|
| 33 |
+
processed_image = pipeline.preprocess_image(image)
|
| 34 |
+
return np.array(processed_image), processed_image
|
| 35 |
|
| 36 |
|
| 37 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult, model_id: str) -> dict:
|
|
|
|
| 76 |
|
| 77 |
|
| 78 |
@spaces.GPU
|
| 79 |
+
def image_to_3d(image: np.array, seed: int, randomize_seed: bool, ss_guidance_strength: float, ss_sampling_steps: int, slat_guidance_strength: float, slat_sampling_steps: int) -> Tuple[dict, str]:
|
| 80 |
"""
|
| 81 |
Convert an image to a 3D model.
|
| 82 |
|
| 83 |
Args:
|
| 84 |
+
image (np.array): The input image.
|
| 85 |
seed (int): The random seed.
|
| 86 |
randomize_seed (bool): Whether to randomize the seed.
|
| 87 |
ss_guidance_strength (float): The guidance strength for sparse structure generation.
|
|
|
|
| 95 |
"""
|
| 96 |
if randomize_seed:
|
| 97 |
seed = np.random.randint(0, MAX_SEED)
|
| 98 |
+
outputs = pipeline.run(
|
| 99 |
+
Image.fromarray(image),
|
| 100 |
+
seed=seed,
|
| 101 |
formats=["gaussian", "mesh"],
|
| 102 |
preprocess_image=False,
|
| 103 |
sparse_structure_sampler_params={
|
|
|
|
| 183 |
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
| 184 |
model_output = LitModel3D(label="Extracted GLB", exposure=20.0, height=300)
|
| 185 |
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
| 186 |
+
|
| 187 |
+
image = gr.State()
|
| 188 |
+
model = gr.State()
|
| 189 |
|
| 190 |
# Example images at the bottom of the page
|
| 191 |
with gr.Row():
|
|
|
|
| 196 |
],
|
| 197 |
inputs=[image_prompt],
|
| 198 |
fn=lambda image: preprocess_image(image),
|
| 199 |
+
outputs=[image, image_prompt],
|
| 200 |
run_on_click=True,
|
| 201 |
examples_per_page=64,
|
| 202 |
)
|
| 203 |
|
|
|
|
|
|
|
| 204 |
# Handlers
|
| 205 |
image_prompt.upload(
|
| 206 |
preprocess_image,
|
| 207 |
inputs=[image_prompt],
|
| 208 |
+
outputs=[image, image_prompt],
|
| 209 |
)
|
| 210 |
|
| 211 |
generate_btn.click(
|
| 212 |
image_to_3d,
|
| 213 |
+
inputs=[image, seed, randomize_seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
|
| 214 |
outputs=[model, video_output],
|
| 215 |
).then(
|
| 216 |
activate_button,
|
trellis/pipelines/trellis_image_to_3d.py
CHANGED
|
@@ -254,10 +254,11 @@ class TrellisImageTo3DPipeline(Pipeline):
|
|
| 254 |
return slat
|
| 255 |
|
| 256 |
@torch.no_grad()
|
| 257 |
-
def
|
| 258 |
self,
|
| 259 |
image: Image.Image,
|
| 260 |
num_samples: int = 1,
|
|
|
|
| 261 |
sparse_structure_sampler_params: dict = {},
|
| 262 |
slat_sampler_params: dict = {},
|
| 263 |
formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
|
|
@@ -276,6 +277,7 @@ class TrellisImageTo3DPipeline(Pipeline):
|
|
| 276 |
if preprocess_image:
|
| 277 |
image = self.preprocess_image(image)
|
| 278 |
cond = self.get_cond([image])
|
|
|
|
| 279 |
coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
|
| 280 |
slat = self.sample_slat(cond, coords, slat_sampler_params)
|
| 281 |
return self.decode_slat(slat, formats)
|
|
|
|
| 254 |
return slat
|
| 255 |
|
| 256 |
@torch.no_grad()
|
| 257 |
+
def run(
|
| 258 |
self,
|
| 259 |
image: Image.Image,
|
| 260 |
num_samples: int = 1,
|
| 261 |
+
seed: int = 42,
|
| 262 |
sparse_structure_sampler_params: dict = {},
|
| 263 |
slat_sampler_params: dict = {},
|
| 264 |
formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
|
|
|
|
| 277 |
if preprocess_image:
|
| 278 |
image = self.preprocess_image(image)
|
| 279 |
cond = self.get_cond([image])
|
| 280 |
+
torch.manual_seed(seed)
|
| 281 |
coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
|
| 282 |
slat = self.sample_slat(cond, coords, slat_sampler_params)
|
| 283 |
return self.decode_slat(slat, formats)
|