Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,852 Bytes
0297809 4be2365 0297809 4be2365 0297809 4be2365 37e13fe 4be2365 37e13fe 4be2365 37e13fe 0297809 4be2365 0297809 4be2365 0297809 4be2365 0297809 4be2365 0297809 |
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 |
import gc
import os
import spaces
import gradio as gr
import random
import tempfile
import time
from easydict import EasyDict
import numpy as np
import torch
from dav.pipelines import DAVPipeline
from dav.models import UNetSpatioTemporalRopeConditionModel
from diffusers import AutoencoderKLTemporalDecoder, FlowMatchEulerDiscreteScheduler
from dav.utils import img_utils
def seed_all(seed: int = 0):
"""
Set random seeds for reproducibility.
"""
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
examples = [
["demos/wooly_mammoth.mp4", 3, 32, 16, 16, 6, 960],
]
def load_models(model_base, device):
vae = AutoencoderKLTemporalDecoder.from_pretrained(model_base, subfolder="vae")
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
model_base, subfolder="scheduler"
)
unet = UNetSpatioTemporalRopeConditionModel.from_pretrained(
model_base, subfolder="unet"
)
unet_interp = UNetSpatioTemporalRopeConditionModel.from_pretrained(
model_base, subfolder="unet_interp"
)
pipe = DAVPipeline(
vae=vae,
unet=unet,
unet_interp=unet_interp,
scheduler=scheduler,
)
pipe = pipe.to(device)
return pipe
model_base = "hhyangcs/depth-any-video"
device_type = "cuda"
device = torch.device(device_type)
pipe = load_models(model_base, device)
@spaces.GPU(duration=140)
def infer_depth(
file: str,
denoise_steps: int = 3,
num_frames: int = 32,
decode_chunk_size: int = 16,
num_interp_frames: int = 16,
num_overlap_frames: int = 6,
max_resolution: int = 1024,
seed: int = 66,
output_dir: str = "./outputs",
):
seed_all(seed)
max_frames = (num_interp_frames + 2 - num_overlap_frames) * (num_frames // 2)
image, fps = img_utils.read_video(file, max_frames=max_frames)
image = img_utils.imresize_max(image, max_resolution)
image = img_utils.imcrop_multi(image)
image_tensor = np.ascontiguousarray(
[_img.transpose(2, 0, 1) / 255.0 for _img in image]
)
image_tensor = torch.from_numpy(image_tensor).to(device)
print(f"==> video name: {file}, frames shape: {image_tensor.shape}")
with torch.no_grad(), torch.autocast(device_type=device_type, dtype=torch.float16):
pipe_out = pipe(
image_tensor,
num_frames=num_frames,
num_overlap_frames=num_overlap_frames,
num_interp_frames=num_interp_frames,
decode_chunk_size=decode_chunk_size,
num_inference_steps=denoise_steps,
)
disparity = pipe_out.disparity
disparity_colored = pipe_out.disparity_colored
image = pipe_out.image
# (N, H, 2 * W, 3)
merged = np.concatenate(
[
image,
disparity_colored,
],
axis=2,
)
file_name = os.path.splitext(os.path.basename(file))[0]
os.makedirs(output_dir, exist_ok=True)
output_path = os.path.join(output_dir, f"{file_name}_depth.mp4")
img_utils.write_video(
output_path,
merged,
fps,
)
# clear the cache for the next video
gc.collect()
torch.cuda.empty_cache()
return output_path
def construct_demo():
with gr.Blocks(analytics_enabled=False) as depthanyvideo_iface:
with gr.Row(equal_height=True):
with gr.Column(scale=1):
input_video = gr.Video(label="Input Video")
with gr.Column(scale=1):
with gr.Row(equal_height=True):
output_video = gr.Video(
label="Ouput Video Depth",
interactive=False,
autoplay=True,
loop=True,
show_share_button=True,
scale=1,
)
with gr.Row(equal_height=True):
with gr.Column(scale=1):
with gr.Row(equal_height=False):
with gr.Accordion("Advanced Settings", open=False):
denoise_steps = gr.Slider(
label="Denoise Steps",
minimum=1,
maximum=10,
value=3,
step=1,
)
num_frames = gr.Slider(
label="Number of Key Frames",
minimum=16,
maximum=32,
value=24,
step=2,
)
decode_chunk_size = gr.Slider(
label="Decode Chunk Size",
minimum=8,
maximum=32,
value=16,
step=1,
)
num_interp_frames = gr.Slider(
label="Number of Interpolation Frames",
minimum=8,
maximum=32,
value=16,
step=1,
)
num_overlap_frames = gr.Slider(
label="Number of Overlap Frames",
minimum=2,
maximum=10,
value=6,
step=1,
)
max_resolution = gr.Slider(
label="Maximum Resolution",
minimum=512,
maximum=2048,
value=1024,
step=32,
)
generate_btn = gr.Button("Generate")
with gr.Column(scale=2):
pass
gr.Examples(
examples=examples,
inputs=[
input_video,
denoise_steps,
num_frames,
decode_chunk_size,
num_interp_frames,
num_overlap_frames,
max_resolution,
],
outputs=output_video,
fn=infer_depth,
cache_examples="lazy",
)
generate_btn.click(
fn=infer_depth,
inputs=[
input_video,
denoise_steps,
num_frames,
decode_chunk_size,
num_interp_frames,
num_overlap_frames,
max_resolution,
],
outputs=output_video,
)
return depthanyvideo_iface
demo = construct_demo()
if __name__ == "__main__":
demo.queue()
demo.launch(share=True)
|