Spaces:
Running
on
Zero
Running
on
Zero
ZeroGPU
#5
by
hysts
HF Staff
- opened
- hf_gradio_app.py +36 -31
hf_gradio_app.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import os, random, time
|
2 |
-
|
3 |
import uuid
|
4 |
import tempfile, shutil
|
5 |
from pydub import AudioSegment
|
@@ -22,22 +22,22 @@ for subfolder in subfolders:
|
|
22 |
|
23 |
snapshot_download(
|
24 |
repo_id = "memoavatar/memo",
|
25 |
-
local_dir = "./checkpoints"
|
26 |
)
|
27 |
|
28 |
snapshot_download(
|
29 |
repo_id = "stabilityai/sd-vae-ft-mse",
|
30 |
-
local_dir = "./checkpoints/vae"
|
31 |
)
|
32 |
|
33 |
snapshot_download(
|
34 |
repo_id = "facebook/wav2vec2-base-960h",
|
35 |
-
local_dir = "./checkpoints/wav2vec2"
|
36 |
)
|
37 |
|
38 |
snapshot_download(
|
39 |
repo_id = "emotion2vec/emotion2vec_plus_large",
|
40 |
-
local_dir = "./checkpoints/emotion2vec_plus_large"
|
41 |
)
|
42 |
|
43 |
import torch
|
@@ -65,51 +65,53 @@ from memo.utils.vision_utils import preprocess_image, tensor_to_video
|
|
65 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
66 |
weight_dtype = torch.bfloat16
|
67 |
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
pipeline.to(device=device, dtype=weight_dtype)
|
84 |
|
85 |
def process_audio(file_path, temp_dir):
|
86 |
# Load the audio file
|
87 |
audio = AudioSegment.from_file(file_path)
|
88 |
-
|
89 |
# Check and cut the audio if longer than 4 seconds
|
90 |
max_duration = 4 * 1000 # 4 seconds in milliseconds
|
91 |
if len(audio) > max_duration:
|
92 |
audio = audio[:max_duration]
|
93 |
-
|
94 |
# Save the processed audio in the temporary directory
|
95 |
output_path = os.path.join(temp_dir, "trimmed_audio.wav")
|
96 |
audio.export(output_path, format="wav")
|
97 |
-
|
98 |
# Return the path to the trimmed file
|
99 |
print(f"Processed audio saved at: {output_path}")
|
100 |
return output_path
|
101 |
|
102 |
-
|
|
|
103 |
@torch.inference_mode()
|
104 |
def generate(input_video, input_audio, seed, progress=gr.Progress(track_tqdm=True)):
|
105 |
-
|
|
|
|
|
106 |
is_shared_ui = True if "fffiloni/MEMO" in os.environ['SPACE_ID'] else False
|
107 |
temp_dir = None
|
108 |
if is_shared_ui:
|
109 |
temp_dir = tempfile.mkdtemp()
|
110 |
input_audio = process_audio(input_audio, temp_dir)
|
111 |
print(f"Processed file was stored temporarily at: {input_audio}")
|
112 |
-
|
113 |
resolution = 512
|
114 |
num_generated_frames_per_clip = 16
|
115 |
fps = 30
|
@@ -125,7 +127,7 @@ def generate(input_video, input_audio, seed, progress=gr.Progress(track_tqdm=Tru
|
|
125 |
generator = torch.manual_seed(seed)
|
126 |
img_size = (resolution, resolution)
|
127 |
pixel_values, face_emb = preprocess_image(face_analysis_model="./checkpoints/misc/face_analysis", image_path=input_video, image_size=resolution)
|
128 |
-
|
129 |
output_dir = "./outputs"
|
130 |
os.makedirs(output_dir, exist_ok=True)
|
131 |
cache_dir = os.path.join(output_dir, "audio_preprocess")
|
@@ -190,6 +192,9 @@ def generate(input_video, input_audio, seed, progress=gr.Progress(track_tqdm=Tru
|
|
190 |
)
|
191 |
video_frames.append(pipeline_output.videos)
|
192 |
|
|
|
|
|
|
|
193 |
video_frames = torch.cat(video_frames, dim=2)
|
194 |
video_frames = video_frames.squeeze(0)
|
195 |
video_frames = video_frames[:, :audio_length]
|
@@ -210,7 +215,7 @@ with gr.Blocks(analytics_enabled=False) as demo:
|
|
210 |
<div style="display:flex;column-gap:4px;">
|
211 |
<a href="https://github.com/memoavatar/memo">
|
212 |
<img src='https://img.shields.io/badge/GitHub-Repo-blue'>
|
213 |
-
</a>
|
214 |
<a href="https://memoavatar.github.io/">
|
215 |
<img src='https://img.shields.io/badge/Project-Page-green'>
|
216 |
</a>
|
@@ -225,7 +230,7 @@ with gr.Blocks(analytics_enabled=False) as demo:
|
|
225 |
</a>
|
226 |
</div>
|
227 |
""")
|
228 |
-
|
229 |
with gr.Row():
|
230 |
with gr.Column():
|
231 |
input_video = gr.Image(label="Upload Input Image", type="filepath")
|
@@ -241,4 +246,4 @@ with gr.Blocks(analytics_enabled=False) as demo:
|
|
241 |
outputs=[video_output],
|
242 |
)
|
243 |
|
244 |
-
demo.queue().launch(share=False, show_api=False, show_error=True)
|
|
|
1 |
import os, random, time
|
2 |
+
import spaces
|
3 |
import uuid
|
4 |
import tempfile, shutil
|
5 |
from pydub import AudioSegment
|
|
|
22 |
|
23 |
snapshot_download(
|
24 |
repo_id = "memoavatar/memo",
|
25 |
+
local_dir = "./checkpoints"
|
26 |
)
|
27 |
|
28 |
snapshot_download(
|
29 |
repo_id = "stabilityai/sd-vae-ft-mse",
|
30 |
+
local_dir = "./checkpoints/vae"
|
31 |
)
|
32 |
|
33 |
snapshot_download(
|
34 |
repo_id = "facebook/wav2vec2-base-960h",
|
35 |
+
local_dir = "./checkpoints/wav2vec2"
|
36 |
)
|
37 |
|
38 |
snapshot_download(
|
39 |
repo_id = "emotion2vec/emotion2vec_plus_large",
|
40 |
+
local_dir = "./checkpoints/emotion2vec_plus_large"
|
41 |
)
|
42 |
|
43 |
import torch
|
|
|
65 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
66 |
weight_dtype = torch.bfloat16
|
67 |
|
68 |
+
|
69 |
+
vae = AutoencoderKL.from_pretrained("./checkpoints/vae").to(device=device, dtype=weight_dtype)
|
70 |
+
reference_net = UNet2DConditionModel.from_pretrained("./checkpoints", subfolder="reference_net", use_safetensors=True)
|
71 |
+
diffusion_net = UNet3DConditionModel.from_pretrained("./checkpoints", subfolder="diffusion_net", use_safetensors=True)
|
72 |
+
image_proj = ImageProjModel.from_pretrained("./checkpoints", subfolder="image_proj", use_safetensors=True)
|
73 |
+
audio_proj = AudioProjModel.from_pretrained("./checkpoints", subfolder="audio_proj", use_safetensors=True)
|
74 |
+
vae.requires_grad_(False).eval()
|
75 |
+
reference_net.requires_grad_(False).eval()
|
76 |
+
diffusion_net.requires_grad_(False).eval()
|
77 |
+
image_proj.requires_grad_(False).eval()
|
78 |
+
audio_proj.requires_grad_(False).eval()
|
79 |
+
noise_scheduler = FlowMatchEulerDiscreteScheduler()
|
80 |
+
pipeline = VideoPipeline(vae=vae, reference_net=reference_net, diffusion_net=diffusion_net, scheduler=noise_scheduler, image_proj=image_proj)
|
81 |
+
pipeline.to(device=device, dtype=weight_dtype)
|
82 |
+
|
|
|
83 |
|
84 |
def process_audio(file_path, temp_dir):
|
85 |
# Load the audio file
|
86 |
audio = AudioSegment.from_file(file_path)
|
87 |
+
|
88 |
# Check and cut the audio if longer than 4 seconds
|
89 |
max_duration = 4 * 1000 # 4 seconds in milliseconds
|
90 |
if len(audio) > max_duration:
|
91 |
audio = audio[:max_duration]
|
92 |
+
|
93 |
# Save the processed audio in the temporary directory
|
94 |
output_path = os.path.join(temp_dir, "trimmed_audio.wav")
|
95 |
audio.export(output_path, format="wav")
|
96 |
+
|
97 |
# Return the path to the trimmed file
|
98 |
print(f"Processed audio saved at: {output_path}")
|
99 |
return output_path
|
100 |
|
101 |
+
|
102 |
+
@spaces.GPU(duration=240)
|
103 |
@torch.inference_mode()
|
104 |
def generate(input_video, input_audio, seed, progress=gr.Progress(track_tqdm=True)):
|
105 |
+
pipeline.reference_net.enable_xformers_memory_efficient_attention()
|
106 |
+
pipeline.diffusion_net.enable_xformers_memory_efficient_attention()
|
107 |
+
|
108 |
is_shared_ui = True if "fffiloni/MEMO" in os.environ['SPACE_ID'] else False
|
109 |
temp_dir = None
|
110 |
if is_shared_ui:
|
111 |
temp_dir = tempfile.mkdtemp()
|
112 |
input_audio = process_audio(input_audio, temp_dir)
|
113 |
print(f"Processed file was stored temporarily at: {input_audio}")
|
114 |
+
|
115 |
resolution = 512
|
116 |
num_generated_frames_per_clip = 16
|
117 |
fps = 30
|
|
|
127 |
generator = torch.manual_seed(seed)
|
128 |
img_size = (resolution, resolution)
|
129 |
pixel_values, face_emb = preprocess_image(face_analysis_model="./checkpoints/misc/face_analysis", image_path=input_video, image_size=resolution)
|
130 |
+
|
131 |
output_dir = "./outputs"
|
132 |
os.makedirs(output_dir, exist_ok=True)
|
133 |
cache_dir = os.path.join(output_dir, "audio_preprocess")
|
|
|
192 |
)
|
193 |
video_frames.append(pipeline_output.videos)
|
194 |
|
195 |
+
pipeline.reference_net.disable_xformers_memory_efficient_attention()
|
196 |
+
pipeline.diffusion_net.disable_xformers_memory_efficient_attention()
|
197 |
+
|
198 |
video_frames = torch.cat(video_frames, dim=2)
|
199 |
video_frames = video_frames.squeeze(0)
|
200 |
video_frames = video_frames[:, :audio_length]
|
|
|
215 |
<div style="display:flex;column-gap:4px;">
|
216 |
<a href="https://github.com/memoavatar/memo">
|
217 |
<img src='https://img.shields.io/badge/GitHub-Repo-blue'>
|
218 |
+
</a>
|
219 |
<a href="https://memoavatar.github.io/">
|
220 |
<img src='https://img.shields.io/badge/Project-Page-green'>
|
221 |
</a>
|
|
|
230 |
</a>
|
231 |
</div>
|
232 |
""")
|
233 |
+
|
234 |
with gr.Row():
|
235 |
with gr.Column():
|
236 |
input_video = gr.Image(label="Upload Input Image", type="filepath")
|
|
|
246 |
outputs=[video_output],
|
247 |
)
|
248 |
|
249 |
+
demo.queue().launch(share=False, show_api=False, show_error=True)
|