File size: 8,419 Bytes
b4754db |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 |
import os
from os.path import basename, splitext, join
import tempfile
import gradio as gr
import numpy as np
from PIL import Image
import torch
import cv2
from torchvision.transforms.functional import to_tensor, to_pil_image
from torch import Tensor
from genstereo import GenStereo, AdaptiveFusionLayer
import ssl
from huggingface_hub import hf_hub_download
from extern.DAM2.depth_anything_v2.dpt import DepthAnythingV2
ssl._create_default_https_context = ssl._create_unverified_context
IMAGE_SIZE = 512
DEVICE = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'
CHECKPOINT_NAME = 'genstereo'
def download_models():
models = [
{
'repo': 'stabilityai/sd-vae-ft-mse',
'sub': None,
'dst': 'checkpoints/sd-vae-ft-mse',
'files': ['config.json', 'diffusion_pytorch_model.safetensors'],
'token': None
},
{
'repo': 'lambdalabs/sd-image-variations-diffusers',
'sub': 'image_encoder',
'dst': 'checkpoints',
'files': ['config.json', 'pytorch_model.bin'],
'token': None
},
{
'repo': 'FQiao/GenStereo',
'sub': None,
'dst': 'checkpoints/genstereo',
'files': ['config.json', 'denoising_unet.pth', 'fusion_layer.pth', 'pose_guider.pth', 'reference_unet.pth'],
'token': None
},
{
'repo': 'depth-anything/Depth-Anything-V2-Large',
'sub': None,
'dst': 'checkpoints',
'files': [f'depth_anything_v2_vitl.pth'],
'token': None
}
]
for model in models:
for file in model['files']:
hf_hub_download(
repo_id=model['repo'],
subfolder=model['sub'],
filename=file,
local_dir=model['dst'],
token=model['token']
)
# Setup.
download_models()
# DepthAnythingV2
if 'dam2' not in globals():
model_configs = {
'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]},
'vitb': {'encoder': 'vitb', 'features': 128, 'out_channels': [96, 192, 384, 768]},
'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]},
}
encoder = 'vitl'
encoder_size_map = {'vits': 'Small', 'vitb': 'Base', 'vitl': 'Large'}
if encoder not in encoder_size_map:
raise ValueError(f"Unsupported encoder: {encoder}. Supported: {list(encoder_size_map.keys())}")
dam2 = DepthAnythingV2(**model_configs[encoder])
dam2_checkpoint = f'checkpoints/depth_anything_v2_{encoder}.pth'
dam2.load_state_dict(torch.load(dam2_checkpoint, map_location='cpu'))
dam2 = dam2.to(DEVICE).eval()
# GenStereo
if 'genstereo' not in globals():
genwarp_cfg = dict(
pretrained_model_path='checkpoints',
checkpoint_name=CHECKPOINT_NAME,
half_precision_weights=True
)
genstereo = GenStereo(cfg=genwarp_cfg, device=DEVICE)
# Adaptive Fusion
if 'fusion_model' not in globals():
fusion_model = AdaptiveFusionLayer()
fusion_checkpoint = join('checkpoints', CHECKPOINT_NAME, 'fusion_layer.pth')
fusion_model.load_state_dict(torch.load(fusion_checkpoint))
fusion_model = fusion_model.to(DEVICE).eval()
# Crop the image to the shorter side.
def crop(img: Image) -> Image:
W, H = img.size
if W < H:
left, right = 0, W
top, bottom = np.ceil((H - W) / 2.), np.floor((H - W) / 2.) + W
else:
left, right = np.ceil((W - H) / 2.), np.floor((W - H) / 2.) + H
top, bottom = 0, H
return img.crop((left, top, right, bottom))
# Gradio app
with tempfile.TemporaryDirectory() as tmpdir:
with gr.Blocks(
title='StereoGen Demo',
css='img {display: inline;}'
) as demo:
# Internal states.
src_image = gr.State()
src_depth = gr.State()
proj_mtx = gr.State()
src_view_mtx = gr.State()
# Blocks.
gr.Markdown(
"""
# StereoGen: Towards Open-World Generation of Stereo Images and Unsupervised Matching
[](https://qjizhi.github.io/genstereo)
[](https://huggingface.co/spaces/FQiao/GenStereo)
[](https://github.com/Qjizhi/GenStereo)
[](https://huggingface.co/FQiao/GenStereo/tree/main)
[]()
## Introduction
This is an official demo for the paper "[Towards Open-World Generation of Stereo Images and Unsupervised Matching](https://qjizhi.github.io/genstereo)". Given an arbitrary reference image, GenStereo can generate the corresponding right-view image.
## How to Use
1. Upload a reference image to "Left Image"
- You can also select an image from "Examples"
3. Hit "Generate a right image" button and check the result
"""
)
file = gr.File(label='Left', file_types=['image'])
examples = gr.Examples(
examples=['./assets/COCO_val2017_000000070229.jpg',
'./assets/COCO_val2017_000000092839.jpg',
'./assets/KITTI2015_000003_10.png',
'./assets/KITTI2015_000147_10.png'],
inputs=file
)
with gr.Row():
image_widget = gr.Image(
label='Depth', type='filepath',
interactive=False
)
depth_widget = gr.Image(label='Estimated Depth', type='pil')
# Add scale factor slider
scale_slider = gr.Slider(
label='Scale Factor',
minimum=1.0,
maximum=30.0,
value=15.0,
step=0.1,
)
button = gr.Button('Generate a right image', size='lg', variant='primary')
with gr.Row():
warped_widget = gr.Image(
label='Warped Image', type='pil', interactive=False
)
gen_widget = gr.Image(
label='Generated Right', type='pil', interactive=False
)
def normalize_disp(disp):
return (disp - disp.min()) / (disp.max() - disp.min())
# Callbacks
def cb_mde(image_file: str):
if not image_file:
# Return None if no image is provided (e.g., when file is cleared).
return None, None, None, None
image = crop(Image.open(image_file).convert('RGB')) # Load image using PIL
image = image.resize((IMAGE_SIZE, IMAGE_SIZE))
image_bgr = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
depth_dam2 = dam2.infer_image(image_bgr)
depth = torch.tensor(depth_dam2).unsqueeze(0).unsqueeze(0).float().cuda()
depth_image = cv2.applyColorMap((normalize_disp(depth_dam2) * 255).astype(np.uint8), cv2.COLORMAP_JET)
return image, depth_image, image, depth
def cb_generate(image, depth: Tensor, scale_factor):
norm_disp = normalize_disp(depth)
disp = norm_disp * scale_factor / 100 * IMAGE_SIZE
renders = genstereo(
src_image=image,
src_disparity=disp,
ratio=None,
)
warped = (renders['warped'] + 1) / 2
synthesized = renders['synthesized']
mask = renders['mask']
fusion_image = fusion_model(synthesized.float(), warped.float(), mask.float())
warped_pil = to_pil_image(warped[0])
fusion_pil = to_pil_image(fusion_image[0])
return warped_pil, fusion_pil
# Events
file.change(
fn=cb_mde,
inputs=file,
outputs=[image_widget, depth_widget, src_image, src_depth]
)
button.click(
fn=cb_generate,
inputs=[src_image, src_depth, scale_slider],
outputs=[warped_widget, gen_widget]
)
demo.launch(share=True) |