Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,633 Bytes
fc13e66 |
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 |
import gc
import os
import numpy as np
import torch
from diffusers.training_utils import set_seed
from diffusers import AutoencoderKLTemporalDecoder
from fire import Fire
from normalcrafter.normal_crafter_ppl import NormalCrafterPipeline
from normalcrafter.unet import DiffusersUNetSpatioTemporalConditionModelNormalCrafter
from normalcrafter.utils import vis_sequence_normal, save_video, read_video_frames
class DepthCrafterDemo:
def __init__(
self,
unet_path: str,
pre_train_path: str,
cpu_offload: str = "model",
):
unet = DiffusersUNetSpatioTemporalConditionModelNormalCrafter.from_pretrained(
unet_path,
subfolder="unet",
low_cpu_mem_usage=True,
)
vae = AutoencoderKLTemporalDecoder.from_pretrained(
unet_path, subfolder="vae"
)
weight_dtype = torch.float16
vae.to(dtype=weight_dtype)
unet.to(dtype=weight_dtype)
# load weights of other components from the provided checkpoint
self.pipe = NormalCrafterPipeline.from_pretrained(
pre_train_path,
unet=unet,
vae=vae,
torch_dtype=weight_dtype,
variant="fp16",
)
# for saving memory, we can offload the model to CPU, or even run the model sequentially to save more memory
if cpu_offload is not None:
if cpu_offload == "sequential":
# This will slow, but save more memory
self.pipe.enable_sequential_cpu_offload()
elif cpu_offload == "model":
self.pipe.enable_model_cpu_offload()
else:
raise ValueError(f"Unknown cpu offload option: {cpu_offload}")
else:
self.pipe.to("cuda")
# enable attention slicing and xformers memory efficient attention
try:
self.pipe.enable_xformers_memory_efficient_attention()
except Exception as e:
print(e)
print("Xformers is not enabled")
# self.pipe.enable_attention_slicing()
def infer(
self,
video: str,
save_folder: str = "./demo_output",
window_size: int = 14,
time_step_size: int = 10,
process_length: int = 195,
decode_chunk_size: int = 7,
max_res: int = 1024,
dataset: str = "open",
target_fps: int = 15,
seed: int = 42,
save_npz: bool = False,
):
set_seed(seed)
frames, target_fps = read_video_frames(
video,
process_length,
target_fps,
max_res,
)
# inference the depth map using the DepthCrafter pipeline
with torch.inference_mode():
res = self.pipe(
frames,
decode_chunk_size=decode_chunk_size,
time_step_size=time_step_size,
window_size=window_size,
).frames[0]
# visualize the depth map and save the results
vis = vis_sequence_normal(res)
# save the depth map and visualization with the target FPS
save_path = os.path.join(
save_folder, os.path.splitext(os.path.basename(video))[0]
)
os.makedirs(os.path.dirname(save_path), exist_ok=True)
save_video(vis, save_path + "_vis.mp4", fps=target_fps)
save_video(frames, save_path + "_input.mp4", fps=target_fps)
if save_npz:
np.savez_compressed(save_path + ".npz", depth=res)
return [
save_path + "_input.mp4",
save_path + "_vis.mp4",
]
def run(
self,
input_video,
num_denoising_steps,
guidance_scale,
max_res=1024,
process_length=195,
):
res_path = self.infer(
input_video,
num_denoising_steps,
guidance_scale,
max_res=max_res,
process_length=process_length,
)
# clear the cache for the next video
gc.collect()
torch.cuda.empty_cache()
return res_path[:2]
def main(
video_path: str,
save_folder: str = "./demo_output",
unet_path: str = "Yanrui95/NormalCrafter",
pre_train_path: str = "stabilityai/stable-video-diffusion-img2vid-xt",
process_length: int = -1,
cpu_offload: str = "model",
target_fps: int = -1,
seed: int = 42,
window_size: int = 14,
time_step_size: int = 10,
max_res: int = 1024,
dataset: str = "open",
save_npz: bool = False
):
depthcrafter_demo = DepthCrafterDemo(
unet_path=unet_path,
pre_train_path=pre_train_path,
cpu_offload=cpu_offload,
)
# process the videos, the video paths are separated by comma
video_paths = video_path.split(",")
for video in video_paths:
depthcrafter_demo.infer(
video,
save_folder=save_folder,
window_size=window_size,
process_length=process_length,
time_step_size=time_step_size,
max_res=max_res,
dataset=dataset,
target_fps=target_fps,
seed=seed,
save_npz=save_npz,
)
# clear the cache for the next video
gc.collect()
torch.cuda.empty_cache()
if __name__ == "__main__":
# running configs
# the most important arguments for memory saving are `cpu_offload`, `enable_xformers`, `max_res`, and `window_size`
# the most important arguments for trade-off between quality and speed are
# `num_inference_steps`, `guidance_scale`, and `max_res`
Fire(main)
|