Spaces:
Running
on
Zero
Running
on
Zero
Update sonic.py
Browse files
sonic.py
CHANGED
@@ -33,9 +33,11 @@ def test(
|
|
33 |
height,
|
34 |
batch
|
35 |
):
|
|
|
36 |
for k, v in batch.items():
|
37 |
if isinstance(v, torch.Tensor):
|
38 |
batch[k] = v.unsqueeze(0).to(pipe.device).float()
|
|
|
39 |
ref_img = batch['ref_img']
|
40 |
clip_img = batch['clip_images']
|
41 |
face_mask = batch['face_mask']
|
@@ -45,11 +47,11 @@ def test(
|
|
45 |
audio_len = batch['audio_len']
|
46 |
step = int(config.step)
|
47 |
|
48 |
-
#
|
49 |
-
|
50 |
-
window = 16000 # (1초 단위로 chunk 처리)
|
51 |
audio_prompts = []
|
52 |
last_audio_prompts = []
|
|
|
53 |
for i in range(0, audio_feature.shape[-1], window):
|
54 |
audio_clip_chunk = audio_feature[:, :, i:i+window]
|
55 |
# Whisper encoder
|
@@ -61,30 +63,38 @@ def test(
|
|
61 |
audio_prompts.append(audio_prompt)
|
62 |
last_audio_prompts.append(last_audio_prompt)
|
63 |
|
64 |
-
#
|
65 |
if len(audio_prompts) == 0:
|
66 |
raise ValueError(
|
67 |
"[ERROR] No speech recognized from the audio. "
|
68 |
"Please provide a valid speech audio (with clear voice)."
|
69 |
)
|
70 |
-
# -------------------------------------------------------------
|
71 |
|
72 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
73 |
-
# audio_len*2 부분은 모델 내부 로직에 따라 필요한 padding 처리
|
74 |
audio_prompts = audio_prompts[:, :audio_len*2]
|
75 |
-
audio_prompts = torch.cat([
|
|
|
|
|
|
|
|
|
76 |
|
77 |
last_audio_prompts = torch.cat(last_audio_prompts, dim=1)
|
78 |
last_audio_prompts = last_audio_prompts[:, :audio_len*2]
|
79 |
-
last_audio_prompts = torch.cat([
|
|
|
|
|
|
|
|
|
80 |
|
81 |
ref_tensor_list = []
|
82 |
audio_tensor_list = []
|
83 |
uncond_audio_tensor_list = []
|
84 |
motion_buckets = []
|
|
|
85 |
for i in tqdm(range(audio_len // step)):
|
86 |
audio_clip = audio_prompts[:, i*2*step : i*2*step + 10].unsqueeze(0)
|
87 |
audio_clip_for_bucket = last_audio_prompts[:, i*2*step : i*2*step + 50].unsqueeze(0)
|
|
|
88 |
motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds)
|
89 |
motion_bucket = motion_bucket * 16 + 16
|
90 |
motion_buckets.append(motion_bucket[0])
|
@@ -138,29 +148,33 @@ class Sonic():
|
|
138 |
config = self.config
|
139 |
config.use_interframe = enable_interpolate_frame
|
140 |
|
141 |
-
device = 'cuda:{}'
|
142 |
-
|
143 |
config.pretrained_model_name_or_path = os.path.join(BASE_DIR, config.pretrained_model_name_or_path)
|
144 |
|
|
|
145 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
146 |
config.pretrained_model_name_or_path,
|
147 |
subfolder="vae",
|
148 |
variant="fp16")
|
149 |
|
|
|
150 |
val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
|
151 |
config.pretrained_model_name_or_path,
|
152 |
subfolder="scheduler")
|
153 |
|
|
|
154 |
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
155 |
config.pretrained_model_name_or_path,
|
156 |
subfolder="image_encoder",
|
157 |
variant="fp16")
|
158 |
|
|
|
159 |
unet = UNetSpatioTemporalConditionModel.from_pretrained(
|
160 |
config.pretrained_model_name_or_path,
|
161 |
subfolder="unet",
|
162 |
variant="fp16")
|
163 |
|
|
|
164 |
add_ip_adapters(unet, [32], [config.ip_audio_scale])
|
165 |
|
166 |
audio2token = AudioProjModel(
|
@@ -174,6 +188,7 @@ class Sonic():
|
|
174 |
context_tokens=2
|
175 |
).to(device)
|
176 |
|
|
|
177 |
unet_checkpoint_path = os.path.join(BASE_DIR, config.unet_checkpoint_path)
|
178 |
audio2token_checkpoint_path = os.path.join(BASE_DIR, config.audio2token_checkpoint_path)
|
179 |
audio2bucket_checkpoint_path = os.path.join(BASE_DIR, config.audio2bucket_checkpoint_path)
|
@@ -193,6 +208,7 @@ class Sonic():
|
|
193 |
strict=True,
|
194 |
)
|
195 |
|
|
|
196 |
if config.weight_dtype == "fp16":
|
197 |
weight_dtype = torch.float16
|
198 |
elif config.weight_dtype == "fp32":
|
@@ -200,26 +216,34 @@ class Sonic():
|
|
200 |
elif config.weight_dtype == "bf16":
|
201 |
weight_dtype = torch.bfloat16
|
202 |
else:
|
203 |
-
raise ValueError(
|
204 |
-
f"Do not support weight dtype: {config.weight_dtype}"
|
205 |
-
)
|
206 |
|
207 |
-
|
|
|
|
|
|
|
208 |
whisper.requires_grad_(False)
|
209 |
|
210 |
-
self.feature_extractor = AutoFeatureExtractor.from_pretrained(
|
|
|
|
|
211 |
|
|
|
212 |
det_path = os.path.join(BASE_DIR, 'checkpoints/yoloface_v5m.pt')
|
213 |
self.face_det = AlignImage(device, det_path=det_path)
|
|
|
|
|
214 |
if config.use_interframe:
|
215 |
rife = RIFEModel(device=device)
|
216 |
rife.load_model(os.path.join(BASE_DIR, 'checkpoints', 'RIFE/'))
|
217 |
self.rife = rife
|
218 |
|
|
|
219 |
image_encoder.to(weight_dtype)
|
220 |
vae.to(weight_dtype)
|
221 |
unet.to(weight_dtype)
|
222 |
|
|
|
223 |
pipe = SonicPipeline(
|
224 |
unet=unet,
|
225 |
image_encoder=image_encoder,
|
@@ -237,13 +261,13 @@ class Sonic():
|
|
237 |
|
238 |
print('Sonic init done')
|
239 |
|
240 |
-
|
241 |
def preprocess(self, image_path, expand_ratio=1.0):
|
242 |
face_image = cv2.imread(image_path)
|
243 |
h, w = face_image.shape[:2]
|
244 |
_, _, bboxes = self.face_det(face_image, maxface=True)
|
245 |
face_num = len(bboxes)
|
246 |
bbox_s = None
|
|
|
247 |
if face_num > 0:
|
248 |
x1, y1, ww, hh = bboxes[0]
|
249 |
x2, y2 = x1 + ww, y1 + hh
|
@@ -270,7 +294,7 @@ class Sonic():
|
|
270 |
dynamic_scale=1.0,
|
271 |
keep_resolution=False,
|
272 |
seed=None):
|
273 |
-
|
274 |
config = self.config
|
275 |
device = self.device
|
276 |
pipe = self.pipe
|
@@ -279,6 +303,7 @@ class Sonic():
|
|
279 |
audio2bucket = self.audio2bucket
|
280 |
image_encoder = self.image_encoder
|
281 |
|
|
|
282 |
if seed:
|
283 |
config.seed = seed
|
284 |
config.num_inference_steps = inference_steps
|
@@ -288,17 +313,16 @@ class Sonic():
|
|
288 |
video_path = output_path.replace('.mp4', '_noaudio.mp4')
|
289 |
audio_video_path = output_path
|
290 |
|
291 |
-
#
|
292 |
test_data = image_audio_to_tensor(
|
293 |
self.face_det,
|
294 |
self.feature_extractor,
|
295 |
image_path,
|
296 |
audio_path,
|
297 |
-
limit=-1, #
|
298 |
image_size=min_resolution,
|
299 |
area=config.area
|
300 |
)
|
301 |
-
|
302 |
if test_data is None:
|
303 |
return -1
|
304 |
|
@@ -310,6 +334,7 @@ class Sonic():
|
|
310 |
else:
|
311 |
resolution = f'{width}x{height}'
|
312 |
|
|
|
313 |
video = test(
|
314 |
pipe,
|
315 |
config,
|
@@ -322,7 +347,7 @@ class Sonic():
|
|
322 |
batch=test_data,
|
323 |
)
|
324 |
|
325 |
-
# 중간프레임 보간
|
326 |
if config.use_interframe:
|
327 |
rife = self.rife
|
328 |
out = video.to(device)
|
@@ -337,6 +362,12 @@ class Sonic():
|
|
337 |
results.append(out[:, :, video_len - 1])
|
338 |
video = torch.stack(results, 2).cpu()
|
339 |
|
|
|
340 |
save_videos_grid(video, video_path, n_rows=video.shape[0], fps=config.fps * (2 if config.use_interframe else 1))
|
341 |
-
|
|
|
|
|
|
|
|
|
|
|
342 |
return 0
|
|
|
33 |
height,
|
34 |
batch
|
35 |
):
|
36 |
+
# 배치 텐서를 (1,B,C,H,W) 형태로
|
37 |
for k, v in batch.items():
|
38 |
if isinstance(v, torch.Tensor):
|
39 |
batch[k] = v.unsqueeze(0).to(pipe.device).float()
|
40 |
+
|
41 |
ref_img = batch['ref_img']
|
42 |
clip_img = batch['clip_images']
|
43 |
face_mask = batch['face_mask']
|
|
|
47 |
audio_len = batch['audio_len']
|
48 |
step = int(config.step)
|
49 |
|
50 |
+
# window=3000 -> 16000으로 변경(1초 간격)
|
51 |
+
window = 16000
|
|
|
52 |
audio_prompts = []
|
53 |
last_audio_prompts = []
|
54 |
+
|
55 |
for i in range(0, audio_feature.shape[-1], window):
|
56 |
audio_clip_chunk = audio_feature[:, :, i:i+window]
|
57 |
# Whisper encoder
|
|
|
63 |
audio_prompts.append(audio_prompt)
|
64 |
last_audio_prompts.append(last_audio_prompt)
|
65 |
|
66 |
+
# ★ 여기서 비었으면 예외
|
67 |
if len(audio_prompts) == 0:
|
68 |
raise ValueError(
|
69 |
"[ERROR] No speech recognized from the audio. "
|
70 |
"Please provide a valid speech audio (with clear voice)."
|
71 |
)
|
|
|
72 |
|
73 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
|
|
74 |
audio_prompts = audio_prompts[:, :audio_len*2]
|
75 |
+
audio_prompts = torch.cat([
|
76 |
+
torch.zeros_like(audio_prompts[:, :4]),
|
77 |
+
audio_prompts,
|
78 |
+
torch.zeros_like(audio_prompts[:, :6])
|
79 |
+
], dim=1)
|
80 |
|
81 |
last_audio_prompts = torch.cat(last_audio_prompts, dim=1)
|
82 |
last_audio_prompts = last_audio_prompts[:, :audio_len*2]
|
83 |
+
last_audio_prompts = torch.cat([
|
84 |
+
torch.zeros_like(last_audio_prompts[:, :24]),
|
85 |
+
last_audio_prompts,
|
86 |
+
torch.zeros_like(last_audio_prompts[:, :26])
|
87 |
+
], dim=1)
|
88 |
|
89 |
ref_tensor_list = []
|
90 |
audio_tensor_list = []
|
91 |
uncond_audio_tensor_list = []
|
92 |
motion_buckets = []
|
93 |
+
|
94 |
for i in tqdm(range(audio_len // step)):
|
95 |
audio_clip = audio_prompts[:, i*2*step : i*2*step + 10].unsqueeze(0)
|
96 |
audio_clip_for_bucket = last_audio_prompts[:, i*2*step : i*2*step + 50].unsqueeze(0)
|
97 |
+
|
98 |
motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds)
|
99 |
motion_bucket = motion_bucket * 16 + 16
|
100 |
motion_buckets.append(motion_bucket[0])
|
|
|
148 |
config = self.config
|
149 |
config.use_interframe = enable_interpolate_frame
|
150 |
|
151 |
+
device = f'cuda:{device_id}' if device_id > -1 else 'cpu'
|
|
|
152 |
config.pretrained_model_name_or_path = os.path.join(BASE_DIR, config.pretrained_model_name_or_path)
|
153 |
|
154 |
+
# VAE
|
155 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
156 |
config.pretrained_model_name_or_path,
|
157 |
subfolder="vae",
|
158 |
variant="fp16")
|
159 |
|
160 |
+
# 스케줄러
|
161 |
val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
|
162 |
config.pretrained_model_name_or_path,
|
163 |
subfolder="scheduler")
|
164 |
|
165 |
+
# CLIP Vision
|
166 |
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
167 |
config.pretrained_model_name_or_path,
|
168 |
subfolder="image_encoder",
|
169 |
variant="fp16")
|
170 |
|
171 |
+
# UNet
|
172 |
unet = UNetSpatioTemporalConditionModel.from_pretrained(
|
173 |
config.pretrained_model_name_or_path,
|
174 |
subfolder="unet",
|
175 |
variant="fp16")
|
176 |
|
177 |
+
# Adapter
|
178 |
add_ip_adapters(unet, [32], [config.ip_audio_scale])
|
179 |
|
180 |
audio2token = AudioProjModel(
|
|
|
188 |
context_tokens=2
|
189 |
).to(device)
|
190 |
|
191 |
+
# 로컬 체크포인트 로드
|
192 |
unet_checkpoint_path = os.path.join(BASE_DIR, config.unet_checkpoint_path)
|
193 |
audio2token_checkpoint_path = os.path.join(BASE_DIR, config.audio2token_checkpoint_path)
|
194 |
audio2bucket_checkpoint_path = os.path.join(BASE_DIR, config.audio2bucket_checkpoint_path)
|
|
|
208 |
strict=True,
|
209 |
)
|
210 |
|
211 |
+
# weight_dtype 설정
|
212 |
if config.weight_dtype == "fp16":
|
213 |
weight_dtype = torch.float16
|
214 |
elif config.weight_dtype == "fp32":
|
|
|
216 |
elif config.weight_dtype == "bf16":
|
217 |
weight_dtype = torch.bfloat16
|
218 |
else:
|
219 |
+
raise ValueError(f"Do not support weight dtype: {config.weight_dtype}")
|
|
|
|
|
220 |
|
221 |
+
# Whisper
|
222 |
+
whisper = WhisperModel.from_pretrained(
|
223 |
+
os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/')
|
224 |
+
).to(device).eval()
|
225 |
whisper.requires_grad_(False)
|
226 |
|
227 |
+
self.feature_extractor = AutoFeatureExtractor.from_pretrained(
|
228 |
+
os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/')
|
229 |
+
)
|
230 |
|
231 |
+
# Face detect
|
232 |
det_path = os.path.join(BASE_DIR, 'checkpoints/yoloface_v5m.pt')
|
233 |
self.face_det = AlignImage(device, det_path=det_path)
|
234 |
+
|
235 |
+
# RIFE 중간프레임 보간
|
236 |
if config.use_interframe:
|
237 |
rife = RIFEModel(device=device)
|
238 |
rife.load_model(os.path.join(BASE_DIR, 'checkpoints', 'RIFE/'))
|
239 |
self.rife = rife
|
240 |
|
241 |
+
# dtype 변경
|
242 |
image_encoder.to(weight_dtype)
|
243 |
vae.to(weight_dtype)
|
244 |
unet.to(weight_dtype)
|
245 |
|
246 |
+
# SonicPipeline 초기화
|
247 |
pipe = SonicPipeline(
|
248 |
unet=unet,
|
249 |
image_encoder=image_encoder,
|
|
|
261 |
|
262 |
print('Sonic init done')
|
263 |
|
|
|
264 |
def preprocess(self, image_path, expand_ratio=1.0):
|
265 |
face_image = cv2.imread(image_path)
|
266 |
h, w = face_image.shape[:2]
|
267 |
_, _, bboxes = self.face_det(face_image, maxface=True)
|
268 |
face_num = len(bboxes)
|
269 |
bbox_s = None
|
270 |
+
|
271 |
if face_num > 0:
|
272 |
x1, y1, ww, hh = bboxes[0]
|
273 |
x2, y2 = x1 + ww, y1 + hh
|
|
|
294 |
dynamic_scale=1.0,
|
295 |
keep_resolution=False,
|
296 |
seed=None):
|
297 |
+
|
298 |
config = self.config
|
299 |
device = self.device
|
300 |
pipe = self.pipe
|
|
|
303 |
audio2bucket = self.audio2bucket
|
304 |
image_encoder = self.image_encoder
|
305 |
|
306 |
+
# 시드 설정
|
307 |
if seed:
|
308 |
config.seed = seed
|
309 |
config.num_inference_steps = inference_steps
|
|
|
313 |
video_path = output_path.replace('.mp4', '_noaudio.mp4')
|
314 |
audio_video_path = output_path
|
315 |
|
316 |
+
# 오디오+이미지 -> tensor
|
317 |
test_data = image_audio_to_tensor(
|
318 |
self.face_det,
|
319 |
self.feature_extractor,
|
320 |
image_path,
|
321 |
audio_path,
|
322 |
+
limit=-1, # 전체 사용
|
323 |
image_size=min_resolution,
|
324 |
area=config.area
|
325 |
)
|
|
|
326 |
if test_data is None:
|
327 |
return -1
|
328 |
|
|
|
334 |
else:
|
335 |
resolution = f'{width}x{height}'
|
336 |
|
337 |
+
# 여기서 test(...) 호출
|
338 |
video = test(
|
339 |
pipe,
|
340 |
config,
|
|
|
347 |
batch=test_data,
|
348 |
)
|
349 |
|
350 |
+
# 중간프레임 보간
|
351 |
if config.use_interframe:
|
352 |
rife = self.rife
|
353 |
out = video.to(device)
|
|
|
362 |
results.append(out[:, :, video_len - 1])
|
363 |
video = torch.stack(results, 2).cpu()
|
364 |
|
365 |
+
# 비디오 저장
|
366 |
save_videos_grid(video, video_path, n_rows=video.shape[0], fps=config.fps * (2 if config.use_interframe else 1))
|
367 |
+
|
368 |
+
# 오디오 합성 후 최종 mp4
|
369 |
+
os.system(
|
370 |
+
f"ffmpeg -i '{video_path}' -i '{audio_path}' -s {resolution} "
|
371 |
+
f"-vcodec libx264 -acodec aac -crf 18 -shortest '{audio_video_path}' -y; rm '{video_path}'"
|
372 |
+
)
|
373 |
return 0
|