Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
!pip install "huggingface_hub[hf_transfer]"
|
2 |
+
!pip install -U "huggingface_hub[cli]"
|
3 |
+
!pip install gradio trimesh scipy
|
4 |
+
!HF_HUB_ENABLE_HF_TRANSFER=1
|
5 |
+
!git clone https://github.com/PaulBorneP/MESA.git
|
6 |
+
!cd MESA
|
7 |
+
!mkdir weights
|
8 |
+
!huggingface-cli download NewtNewt/MESA --local-dir weights
|
9 |
+
|
10 |
+
import torch
|
11 |
+
from MESA.pipeline_terrain import TerrainDiffusionPipeline
|
12 |
+
import sys
|
13 |
+
import gradio as gr
|
14 |
+
import numpy as np
|
15 |
+
import trimesh
|
16 |
+
import tempfile
|
17 |
+
import torch
|
18 |
+
from scipy.spatial import Delaunay
|
19 |
+
|
20 |
+
sys.path.append('MESA/')
|
21 |
+
|
22 |
+
pipe = TerrainDiffusionPipeline.from_pretrained("./weights", torch_dtype=torch.float16)
|
23 |
+
pipe.to("cuda")
|
24 |
+
|
25 |
+
def generate_terrain(prompt, num_inference_steps, guidance_scale, seed, crop_size, prefix):
|
26 |
+
"""Generates terrain data (RGB and elevation) from a text prompt."""
|
27 |
+
if prefix and not prefix.endswith(' '):
|
28 |
+
prefix += ' ' # Ensure prefix ends with a space
|
29 |
+
|
30 |
+
full_prompt = prefix + prompt
|
31 |
+
generator = torch.Generator("cuda").manual_seed(seed)
|
32 |
+
image, dem = pipe(full_prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale, generator=generator)
|
33 |
+
|
34 |
+
# Center crop the image and dem
|
35 |
+
h, w, c = image[0].shape
|
36 |
+
start_h = (h - crop_size) // 2
|
37 |
+
start_w = (w - crop_size) // 2
|
38 |
+
end_h = start_h + crop_size
|
39 |
+
end_w = start_w + crop_size
|
40 |
+
|
41 |
+
cropped_image = image[0][start_h:end_h, start_w:end_w, :]
|
42 |
+
cropped_dem = dem[0][start_h:end_h, start_w:end_w, :]
|
43 |
+
|
44 |
+
return (255 * cropped_image).astype(np.uint8), 500*cropped_dem.mean(-1)
|
45 |
+
|
46 |
+
def create_3d_mesh(rgb, elevation):
|
47 |
+
"""Creates a 3D mesh from RGB and elevation data."""
|
48 |
+
x, y = np.meshgrid(np.arange(elevation.shape[1]), np.arange(elevation.shape[0]))
|
49 |
+
points = np.stack([x.flatten(), y.flatten()], axis=-1)
|
50 |
+
tri = Delaunay(points)
|
51 |
+
|
52 |
+
vertices = np.stack([x.flatten(), y.flatten(), elevation.flatten()], axis=-1)
|
53 |
+
faces = tri.simplices
|
54 |
+
|
55 |
+
mesh = trimesh.Trimesh(vertices=vertices, faces=faces, vertex_colors=rgb.reshape(-1, 3))
|
56 |
+
|
57 |
+
return mesh
|
58 |
+
|
59 |
+
def generate_and_display(prompt, num_inference_steps, guidance_scale, seed, crop_size, prefix):
|
60 |
+
"""Generates terrain and displays it as a 3D model."""
|
61 |
+
rgb, elevation = generate_terrain(prompt, num_inference_steps, guidance_scale, seed, crop_size, prefix)
|
62 |
+
mesh = create_3d_mesh(rgb, elevation)
|
63 |
+
|
64 |
+
with tempfile.NamedTemporaryFile(suffix=".obj", delete=False) as temp_file:
|
65 |
+
mesh.export(temp_file.name)
|
66 |
+
file_path = temp_file.name
|
67 |
+
|
68 |
+
return file_path
|
69 |
+
|
70 |
+
theme = gr.themes.Soft(primary_hue="red", secondary_hue="red", font=['arial'])
|
71 |
+
|
72 |
+
with gr.Blocks(theme=theme) as demo:
|
73 |
+
with gr.Column(elem_classes="header"):
|
74 |
+
gr.Markdown("# MESA: Text-Driven Terrain Generation Using Latent Diffusion and Global Copernicus Data")
|
75 |
+
gr.Markdown("### Paul Borne–Pons, Mikolaj Czerkawski, Rosalie Martin, Romain Rouffet")
|
76 |
+
gr.Markdown('[[GitHub](https://github.com/PaulBorneP/MESA)] [[Model](https://huggingface.co/NewtNewt/MESA)] [[Dataset](https://huggingface.co/datasets/Major-TOM/Core-DEM)]')
|
77 |
+
|
78 |
+
# Abstract Section
|
79 |
+
with gr.Column(elem_classes="abstract"):
|
80 |
+
gr.Markdown("MESA is a novel generative model based on latent denoising diffusion capable of generating 2.5D representations of terrain based on the text prompt conditioning supplied via natural language. The model produces two co-registered modalities of optical and depth maps.") # Replace with your abstract text
|
81 |
+
gr.Markdown("This is a test version of the demo app. Please be aware that MESA supports primarily complex, mountainous terrains as opposed to flat land")
|
82 |
+
gr.Markdown("The generated image is quite large, so for the full resolution (768) it might take a while to load the surface")
|
83 |
+
|
84 |
+
with gr.Row():
|
85 |
+
prompt_input = gr.Textbox(lines=2, placeholder="Enter a terrain description...")
|
86 |
+
generate_button = gr.Button("Generate Terrain", variant="primary")
|
87 |
+
|
88 |
+
model_output = gr.Model3D(
|
89 |
+
camera_position=[90, 180, 512]
|
90 |
+
)
|
91 |
+
|
92 |
+
with gr.Accordion("Advanced Options", open=False) as advanced_options:
|
93 |
+
num_inference_steps_slider = gr.Slider(minimum=10, maximum=1000, step=10, value=50, label="Inference Steps")
|
94 |
+
guidance_scale_slider = gr.Slider(minimum=1.0, maximum=15.0, step=0.5, value=7.5, label="Guidance Scale")
|
95 |
+
seed_number = gr.Number(value=6378, label="Seed")
|
96 |
+
crop_size_slider = gr.Slider(minimum=128, maximum=768, step=64, value=512, label="Crop Size")
|
97 |
+
prefix_textbox = gr.Textbox(label="Prompt Prefix", value="A Sentinel-2 image of ")
|
98 |
+
|
99 |
+
generate_button.click(
|
100 |
+
fn=generate_and_display,
|
101 |
+
inputs=[prompt_input, num_inference_steps_slider, guidance_scale_slider, seed_number, crop_size_slider, prefix_textbox],
|
102 |
+
outputs=model_output,
|
103 |
+
)
|
104 |
+
|
105 |
+
if __name__ == "__main__":
|
106 |
+
demo.launch(debug=True,
|
107 |
+
share=True)
|