Spaces:
Sleeping
Sleeping
Create app.py
Browse filesTODO: make sure source_camero have the right shape and value
TODO: instead of outputting .obj file -> directly output a 3d model
app.py
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import gradio as gr
|
3 |
+
import os
|
4 |
+
import numpy as np
|
5 |
+
import trimesh
|
6 |
+
import mcubes
|
7 |
+
from torchvision.utils import save_image
|
8 |
+
from PIL import Image
|
9 |
+
from transformers import AutoModel, AutoConfig
|
10 |
+
from rembg import remove, new_session
|
11 |
+
from functools import partial
|
12 |
+
from kiui.op import recenter
|
13 |
+
import kiui
|
14 |
+
|
15 |
+
|
16 |
+
# we load the pre-trained model from HF
|
17 |
+
class LRMGeneratorWrapper:
|
18 |
+
def __init__(self):
|
19 |
+
self.config = AutoConfig.from_pretrained("jadechoghari/custom-llrm", trust_remote_code=True)
|
20 |
+
self.model = AutoModel.from_pretrained("jadechoghari/custom-llrm", trust_remote_code=True)
|
21 |
+
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
22 |
+
self.model.to(self.device)
|
23 |
+
self.model.eval()
|
24 |
+
|
25 |
+
def forward(self, image, camera):
|
26 |
+
return self.model(image, camera)
|
27 |
+
|
28 |
+
model_wrapper = LRMGeneratorWrapper()
|
29 |
+
|
30 |
+
|
31 |
+
def preprocess_image(image, source_size):
|
32 |
+
session = new_session("isnet-general-use")
|
33 |
+
rembg_remove = partial(remove, session=session)
|
34 |
+
image = np.array(image)
|
35 |
+
image = rembg_remove(image)
|
36 |
+
mask = rembg_remove(image, only_mask=True)
|
37 |
+
image = recenter(image, mask, border_ratio=0.20)
|
38 |
+
image = torch.tensor(image).permute(2, 0, 1).unsqueeze(0) / 255.0
|
39 |
+
if image.shape[1] == 4:
|
40 |
+
image = image[:, :3, ...] * image[:, 3:, ...] + (1 - image[:, 3:, ...])
|
41 |
+
image = torch.nn.functional.interpolate(image, size=(source_size, source_size), mode='bicubic', align_corners=True)
|
42 |
+
image = torch.clamp(image, 0, 1)
|
43 |
+
return image
|
44 |
+
|
45 |
+
#Ref: https://github.com/jadechoghari/vfusion3d/blob/main/lrm/inferrer.py
|
46 |
+
def generate_mesh(image, source_size=512, render_size=384, mesh_size=512, export_mesh=True):
|
47 |
+
image = preprocess_image(image, source_size).to(model_wrapper.device)
|
48 |
+
|
49 |
+
# TODO: make sure source_camero have the right shape and value
|
50 |
+
source_camera = torch.tensor([[0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]], dtype=torch.float32).to(model_wrapper.device)
|
51 |
+
|
52 |
+
render_camera = torch.tensor([[0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]], dtype=torch.float32).to(model_wrapper.device)
|
53 |
+
|
54 |
+
with torch.no_grad():
|
55 |
+
planes = model_wrapper.forward(image, source_camera)
|
56 |
+
|
57 |
+
if export_mesh:
|
58 |
+
grid_out = model_wrapper.model.synthesizer.forward_grid(planes=planes, grid_size=mesh_size)
|
59 |
+
|
60 |
+
vtx, faces = mcubes.marching_cubes(grid_out['sigma'].float().squeeze(0).squeeze(-1).cpu().numpy(), 1.0)
|
61 |
+
vtx = vtx / (mesh_size - 1) * 2 - 1
|
62 |
+
vtx_tensor = torch.tensor(vtx, dtype=torch.float32, device=model_wrapper.device).unsqueeze(0)
|
63 |
+
vtx_colors = model_wrapper.model.synthesizer.forward_points(planes, vtx_tensor)['rgb'].float().squeeze(0).cpu().numpy()
|
64 |
+
|
65 |
+
vtx_colors = (vtx_colors * 255).astype(np.uint8)
|
66 |
+
mesh = trimesh.Trimesh(vertices=vtx, faces=faces, vertex_colors=vtx_colors)
|
67 |
+
|
68 |
+
mesh_path = "awesome_mesh.obj"
|
69 |
+
mesh.export(mesh_path, 'obj')
|
70 |
+
return mesh_path
|
71 |
+
|
72 |
+
# TODO: instead of outputting .obj file -> directly output a 3d model
|
73 |
+
def gradio_interface(image):
|
74 |
+
mesh_file = generate_mesh(image)
|
75 |
+
print("Generated Mesh File Path:", mesh_file)
|
76 |
+
return mesh_file
|
77 |
+
|
78 |
+
|
79 |
+
gr.Interface(
|
80 |
+
fn=gradio_interface,
|
81 |
+
inputs=gr.Image(type="pil", label="Input Image"),
|
82 |
+
outputs=gr.File(label="Awesome 3D Mesh (.obj)"),
|
83 |
+
title="3D Mesh Generator by FacebookAI",
|
84 |
+
description="Upload an image and generate a 3D mesh (.obj) file using VFusion3D by FacebookAI"
|
85 |
+
).launch()
|