Spaces:
Running
on
Zero
Running
on
Zero
Update sonic.py
Browse files
sonic.py
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
import os
|
|
|
2 |
import torch
|
3 |
import torch.utils.checkpoint
|
4 |
from PIL import Image
|
5 |
-
import numpy as np
|
6 |
from omegaconf import OmegaConf
|
7 |
from tqdm import tqdm
|
8 |
import cv2
|
@@ -13,7 +13,9 @@ from transformers import WhisperModel, CLIPVisionModelWithProjection, AutoFeatur
|
|
13 |
|
14 |
from src.utils.util import save_videos_grid, seed_everything
|
15 |
from src.dataset.test_preprocess import process_bbox, image_audio_to_tensor
|
16 |
-
from src.models.base.unet_spatio_temporal_condition import
|
|
|
|
|
17 |
from src.pipelines.pipeline_sonic import SonicPipeline
|
18 |
from src.models.audio_adapter.audio_proj import AudioProjModel
|
19 |
from src.models.audio_adapter.audio_to_bucket import Audio2bucketModel
|
@@ -22,6 +24,10 @@ from src.dataset.face_align.align import AlignImage
|
|
22 |
|
23 |
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
|
24 |
|
|
|
|
|
|
|
|
|
25 |
def test(
|
26 |
pipe,
|
27 |
config,
|
@@ -31,81 +37,88 @@ def test(
|
|
31 |
image_encoder,
|
32 |
width,
|
33 |
height,
|
34 |
-
batch
|
35 |
):
|
36 |
-
#
|
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
|
42 |
-
clip_img
|
43 |
-
face_mask = batch[
|
44 |
image_embeds = image_encoder(clip_img).image_embeds
|
45 |
|
46 |
-
audio_feature = batch[
|
47 |
-
audio_len
|
48 |
-
step
|
49 |
|
50 |
-
#
|
|
|
|
|
|
|
51 |
window = 16000
|
52 |
-
|
53 |
-
|
54 |
|
|
|
|
|
55 |
for i in range(0, audio_feature.shape[-1], window):
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
|
62 |
-
|
63 |
-
|
64 |
-
last_audio_prompts.append(last_audio_prompt)
|
65 |
|
66 |
-
#
|
67 |
if len(audio_prompts) == 0:
|
68 |
raise ValueError(
|
69 |
"[ERROR] No speech recognized from the audio. "
|
70 |
-
"Please provide a valid speech
|
71 |
)
|
72 |
|
73 |
-
|
74 |
-
audio_prompts
|
75 |
-
audio_prompts
|
76 |
-
torch.zeros_like(audio_prompts[:, :4]),
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
last_audio_prompts = torch.cat(
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
uncond_audio_tensor_list = []
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
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])
|
101 |
|
102 |
-
|
103 |
-
|
104 |
|
105 |
ref_tensor_list.append(ref_img[0])
|
106 |
-
audio_tensor_list.append(
|
107 |
-
uncond_audio_tensor_list.append(
|
|
|
|
|
|
|
|
|
108 |
|
|
|
109 |
video = pipe(
|
110 |
ref_img,
|
111 |
clip_img,
|
@@ -128,246 +141,172 @@ def test(
|
|
128 |
shift_offset=config.shift_offset,
|
129 |
frames_per_batch=config.n_sample_frames,
|
130 |
num_inference_steps=config.num_inference_steps,
|
131 |
-
i2i_noise_strength=config.i2i_noise_strength
|
132 |
).frames
|
|
|
133 |
|
134 |
video = (video * 0.5 + 0.5).clamp(0, 1)
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
156 |
-
|
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(
|
181 |
-
seq_len=10, blocks=5, channels=384,
|
182 |
-
intermediate_dim=1024, output_dim=1024, context_tokens=32
|
183 |
-
).to(device)
|
184 |
-
|
185 |
-
audio2bucket = Audio2bucketModel(
|
186 |
-
seq_len=50, blocks=1, channels=384,
|
187 |
-
clip_channels=1024, intermediate_dim=1024, output_dim=1,
|
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)
|
195 |
-
|
196 |
-
unet.load_state_dict(
|
197 |
-
torch.load(unet_checkpoint_path, map_location="cpu"),
|
198 |
-
strict=True,
|
199 |
-
)
|
200 |
-
|
201 |
-
audio2token.load_state_dict(
|
202 |
-
torch.load(audio2token_checkpoint_path, map_location="cpu"),
|
203 |
-
strict=True,
|
204 |
)
|
205 |
-
|
206 |
-
|
207 |
-
torch.load(audio2bucket_checkpoint_path, map_location="cpu"),
|
208 |
-
strict=True,
|
209 |
)
|
210 |
-
|
211 |
-
|
212 |
-
if config.weight_dtype == "fp16":
|
213 |
-
weight_dtype = torch.float16
|
214 |
-
elif config.weight_dtype == "fp32":
|
215 |
-
weight_dtype = torch.float32
|
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 |
-
|
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,
|
250 |
-
vae=vae,
|
251 |
-
scheduler=val_noise_scheduler,
|
252 |
)
|
253 |
-
|
254 |
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
self.audio2bucket = audio2bucket
|
259 |
-
self.image_encoder = image_encoder
|
260 |
-
self.device = device
|
261 |
|
262 |
-
|
|
|
|
|
|
|
263 |
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
_, _, bboxes = self.face_det(face_image, maxface=True)
|
268 |
-
face_num = len(bboxes)
|
269 |
-
bbox_s = None
|
270 |
|
271 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
272 |
x1, y1, ww, hh = bboxes[0]
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
return {
|
278 |
-
'face_num': face_num,
|
279 |
-
'crop_bbox': bbox_s,
|
280 |
-
}
|
281 |
-
|
282 |
-
def crop_image(self, input_image_path, output_image_path, crop_bbox):
|
283 |
-
face_image = cv2.imread(input_image_path)
|
284 |
-
crop_image = face_image[crop_bbox[1]:crop_bbox[3], crop_bbox[0]:crop_bbox[2]]
|
285 |
-
cv2.imwrite(output_image_path, crop_image)
|
286 |
|
|
|
287 |
@torch.no_grad()
|
288 |
-
def process(
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
#
|
307 |
-
|
308 |
-
|
309 |
-
config.num_inference_steps = inference_steps
|
310 |
-
config.motion_bucket_scale = dynamic_scale
|
311 |
-
seed_everything(config.seed)
|
312 |
-
|
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,
|
323 |
-
image_size=min_resolution,
|
324 |
-
area=
|
325 |
)
|
326 |
if test_data is None:
|
327 |
return -1
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
#
|
|
|
338 |
video = test(
|
339 |
-
pipe,
|
340 |
-
|
341 |
-
wav_enc=whisper,
|
342 |
-
audio_pe=audio2token,
|
343 |
-
audio2bucket=audio2bucket,
|
344 |
-
image_encoder=image_encoder,
|
345 |
-
width=
|
346 |
-
height=
|
347 |
batch=test_data,
|
348 |
)
|
349 |
|
350 |
-
#
|
351 |
-
if
|
352 |
-
|
353 |
-
out
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
I2 = out[:, :, idx + 1]
|
359 |
-
middle = rife.inference(I1, I2).clamp(0, 1).detach()
|
360 |
-
results.append(out[:, :, idx])
|
361 |
-
results.append(middle)
|
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 |
-
#
|
|
|
|
|
|
|
|
|
369 |
os.system(
|
370 |
-
f"ffmpeg -i '{
|
371 |
-
f"-vcodec libx264 -acodec aac -crf 18 -shortest '{
|
372 |
)
|
|
|
373 |
return 0
|
|
|
1 |
import os
|
2 |
+
import math # [★ 수정] ceil 계산용
|
3 |
import torch
|
4 |
import torch.utils.checkpoint
|
5 |
from PIL import Image
|
|
|
6 |
from omegaconf import OmegaConf
|
7 |
from tqdm import tqdm
|
8 |
import cv2
|
|
|
13 |
|
14 |
from src.utils.util import save_videos_grid, seed_everything
|
15 |
from src.dataset.test_preprocess import process_bbox, image_audio_to_tensor
|
16 |
+
from src.models.base.unet_spatio_temporal_condition import (
|
17 |
+
UNetSpatioTemporalConditionModel, add_ip_adapters,
|
18 |
+
)
|
19 |
from src.pipelines.pipeline_sonic import SonicPipeline
|
20 |
from src.models.audio_adapter.audio_proj import AudioProjModel
|
21 |
from src.models.audio_adapter.audio_to_bucket import Audio2bucketModel
|
|
|
24 |
|
25 |
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
|
26 |
|
27 |
+
|
28 |
+
# ------------------------------------------------------------------
|
29 |
+
# test() : 한 장의 얼굴 + 오디오 → 프레임 텐서 시퀀스
|
30 |
+
# ------------------------------------------------------------------
|
31 |
def test(
|
32 |
pipe,
|
33 |
config,
|
|
|
37 |
image_encoder,
|
38 |
width,
|
39 |
height,
|
40 |
+
batch,
|
41 |
):
|
42 |
+
# (B,C,H,W) → (1,B,C,H,W)
|
43 |
for k, v in batch.items():
|
44 |
if isinstance(v, torch.Tensor):
|
45 |
batch[k] = v.unsqueeze(0).to(pipe.device).float()
|
46 |
|
47 |
+
ref_img = batch["ref_img"]
|
48 |
+
clip_img = batch["clip_images"]
|
49 |
+
face_mask = batch["face_mask"]
|
50 |
image_embeds = image_encoder(clip_img).image_embeds
|
51 |
|
52 |
+
audio_feature = batch["audio_feature"] # (C,T)
|
53 |
+
audio_len = batch["audio_len"] # # of whisper tokens
|
54 |
+
step = int(config.step)
|
55 |
|
56 |
+
# ----------------------------- [★ 수정] -----------------------------
|
57 |
+
# ① 1 초 구간 단위를 위해 window 16000 → whisper‐tiny 기준 1 초
|
58 |
+
# ② audio_len < step 이면 step 을 줄여 빈 리스트 방지
|
59 |
+
# --------------------------------------------------------------------
|
60 |
window = 16000
|
61 |
+
if audio_len < step:
|
62 |
+
step = max(1, audio_len)
|
63 |
|
64 |
+
# ── 오디오를 1 초 단위로 자르면서 Whisper 인코딩
|
65 |
+
audio_prompts, last_audio_prompts = [], []
|
66 |
for i in range(0, audio_feature.shape[-1], window):
|
67 |
+
chunk = audio_feature[:, :, i : i + window] # (B,C,window)
|
68 |
+
|
69 |
+
# whisper encoder
|
70 |
+
prompt_layers = wav_enc.encoder(chunk, output_hidden_states=True).hidden_states
|
71 |
+
last_hidden = wav_enc.encoder(chunk).last_hidden_state.unsqueeze(-2)
|
72 |
|
73 |
+
audio_prompts.append(torch.stack(prompt_layers, dim=2))
|
74 |
+
last_audio_prompts.append(last_hidden)
|
|
|
75 |
|
76 |
+
# ── 예외: 아무 내용도 없으면 종료
|
77 |
if len(audio_prompts) == 0:
|
78 |
raise ValueError(
|
79 |
"[ERROR] No speech recognized from the audio. "
|
80 |
+
"Please provide a valid speech recording."
|
81 |
)
|
82 |
|
83 |
+
# Whisper token 시퀀스 재구성 (+ 모델 padding 규칙)
|
84 |
+
audio_prompts = torch.cat(audio_prompts, dim=1)[:, : audio_len * 2]
|
85 |
+
audio_prompts = torch.cat(
|
86 |
+
[torch.zeros_like(audio_prompts[:, :4]), audio_prompts, torch.zeros_like(audio_prompts[:, :6])],
|
87 |
+
dim=1,
|
88 |
+
)
|
89 |
+
|
90 |
+
last_audio_prompts = torch.cat(last_audio_prompts, dim=1)[:, : audio_len * 2]
|
91 |
+
last_audio_prompts = torch.cat(
|
92 |
+
[torch.zeros_like(last_audio_prompts[:, :24]), last_audio_prompts, torch.zeros_like(last_audio_prompts[:, :26])],
|
93 |
+
dim=1,
|
94 |
+
)
|
95 |
+
|
96 |
+
# --------------------------------------------------------------------
|
97 |
+
# step 조정 결과를 반영해 총 chunk 횟수 계산 (ceil)
|
98 |
+
# --------------------------------------------------------------------
|
99 |
+
num_chunks = math.ceil(audio_len / step)
|
100 |
+
|
101 |
+
ref_tensor_list, audio_tensor_list, uncond_audio_tensor_list, motion_buckets = [], [], [], []
|
102 |
+
for i in tqdm(range(num_chunks)):
|
103 |
+
start = i * 2 * step
|
104 |
+
audio_clip = audio_prompts[:, start : start + 10].unsqueeze(0)
|
105 |
+
audio_clip_for_bucket = last_audio_prompts[:, start : start + 50].unsqueeze(0)
|
106 |
+
|
107 |
+
motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds) * 16 + 16
|
|
|
|
|
108 |
motion_buckets.append(motion_bucket[0])
|
109 |
|
110 |
+
cond_audio = audio_pe(audio_clip).squeeze(0)
|
111 |
+
uncond_audio = audio_pe(torch.zeros_like(audio_clip)).squeeze(0)
|
112 |
|
113 |
ref_tensor_list.append(ref_img[0])
|
114 |
+
audio_tensor_list.append(cond_audio[0])
|
115 |
+
uncond_audio_tensor_list.append(uncond_audio[0])
|
116 |
+
|
117 |
+
# 빈 리스트 방지
|
118 |
+
if len(audio_tensor_list) == 0:
|
119 |
+
raise ValueError("[ERROR] Audio too short for the configured 'step' size; no frames produced.")
|
120 |
|
121 |
+
# --------------------------------------------------------------------
|
122 |
video = pipe(
|
123 |
ref_img,
|
124 |
clip_img,
|
|
|
141 |
shift_offset=config.shift_offset,
|
142 |
frames_per_batch=config.n_sample_frames,
|
143 |
num_inference_steps=config.num_inference_steps,
|
144 |
+
i2i_noise_strength=config.i2i_noise_strength,
|
145 |
).frames
|
146 |
+
# --------------------------------------------------------------------
|
147 |
|
148 |
video = (video * 0.5 + 0.5).clamp(0, 1)
|
149 |
+
return video.to(pipe.device).unsqueeze(0).cpu()
|
150 |
+
|
151 |
+
|
152 |
+
# ------------------------------------------------------------------
|
153 |
+
# Sonic 클래스
|
154 |
+
# ------------------------------------------------------------------
|
155 |
+
class Sonic:
|
156 |
+
config_file = os.path.join(BASE_DIR, "config/inference/sonic.yaml")
|
157 |
+
config = OmegaConf.load(config_file)
|
158 |
+
|
159 |
+
def __init__(self, device_id: int = 0, enable_interpolate_frame: bool = True):
|
160 |
+
cfg = self.config
|
161 |
+
cfg.use_interframe = enable_interpolate_frame
|
162 |
+
self.device = f"cuda:{device_id}" if device_id >= 0 and torch.cuda.is_available() else "cpu"
|
163 |
+
cfg.pretrained_model_name_or_path = os.path.join(BASE_DIR, cfg.pretrained_model_name_or_path)
|
164 |
+
|
165 |
+
# ───────────── 모델 로드
|
166 |
+
self._load_models(cfg)
|
167 |
+
print("Sonic init done")
|
168 |
+
|
169 |
+
# --------------------------------------------------------------
|
170 |
+
# model / pipeline loader
|
171 |
+
# --------------------------------------------------------------
|
172 |
+
def _load_models(self, cfg):
|
173 |
+
dtype_map = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16}
|
174 |
+
weight_dtype = dtype_map.get(cfg.weight_dtype, torch.float32)
|
175 |
+
|
176 |
+
# backbone
|
177 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
178 |
+
cfg.pretrained_model_name_or_path, subfolder="vae", variant="fp16"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
179 |
)
|
180 |
+
scheduler = EulerDiscreteScheduler.from_pretrained(
|
181 |
+
cfg.pretrained_model_name_or_path, subfolder="scheduler"
|
|
|
|
|
182 |
)
|
183 |
+
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
184 |
+
cfg.pretrained_model_name_or_path, subfolder="image_encoder", variant="fp16"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
185 |
)
|
186 |
+
unet = UNetSpatioTemporalConditionModel.from_pretrained(
|
187 |
+
cfg.pretrained_model_name_or_path, subfolder="unet", variant="fp16"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
)
|
189 |
+
add_ip_adapters(unet, [32], [cfg.ip_audio_scale])
|
190 |
|
191 |
+
# audio adapters
|
192 |
+
audio2token = AudioProjModel(10, 5, 384, 1024, 1024, 32).to(self.device)
|
193 |
+
audio2bucket = Audio2bucketModel(50, 1, 384, 1024, 1024, 1, 2).to(self.device)
|
|
|
|
|
|
|
194 |
|
195 |
+
# checkpoints
|
196 |
+
unet.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.unet_checkpoint_path), map_location="cpu"))
|
197 |
+
audio2token.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2token_checkpoint_path), map_location="cpu"))
|
198 |
+
audio2bucket.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2bucket_checkpoint_path), map_location="cpu"))
|
199 |
|
200 |
+
# whisper
|
201 |
+
whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny")).to(self.device).eval()
|
202 |
+
whisper.requires_grad_(False)
|
|
|
|
|
|
|
203 |
|
204 |
+
# extras
|
205 |
+
self.feature_extractor = AutoFeatureExtractor.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny"))
|
206 |
+
self.face_det = AlignImage(self.device, det_path=os.path.join(BASE_DIR, "checkpoints/yoloface_v5m.pt"))
|
207 |
+
if cfg.use_interframe:
|
208 |
+
self.rife = RIFEModel(device=self.device)
|
209 |
+
self.rife.load_model(os.path.join(BASE_DIR, "checkpoints/RIFE/"))
|
210 |
+
|
211 |
+
# dtype
|
212 |
+
for m in (image_encoder, vae, unet):
|
213 |
+
m.to(weight_dtype)
|
214 |
+
|
215 |
+
# pipeline
|
216 |
+
pipe = SonicPipeline(unet=unet, image_encoder=image_encoder, vae=vae, scheduler=scheduler)
|
217 |
+
self.pipe = pipe.to(device=self.device, dtype=weight_dtype)
|
218 |
+
self.audio2token = audio2token
|
219 |
+
self.audio2bucket = audio2bucket
|
220 |
+
self.image_encoder = image_encoder
|
221 |
+
self.whisper = whisper
|
222 |
+
|
223 |
+
# --------------------------------------------------------------
|
224 |
+
def preprocess(self, image_path: str, expand_ratio: float = 1.0):
|
225 |
+
img = cv2.imread(image_path)
|
226 |
+
h, w = img.shape[:2]
|
227 |
+
_, _, bboxes = self.face_det(img, maxface=True)
|
228 |
+
if bboxes:
|
229 |
x1, y1, ww, hh = bboxes[0]
|
230 |
+
bbox = (x1, y1, x1 + ww, y1 + hh)
|
231 |
+
crop_bbox = process_bbox(bbox, expand_radio=expand_ratio, height=h, width=w)
|
232 |
+
return {"face_num": len(bboxes), "crop_bbox": crop_bbox}
|
233 |
+
return {"face_num": 0, "crop_bbox": None}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
234 |
|
235 |
+
# --------------------------------------------------------------
|
236 |
@torch.no_grad()
|
237 |
+
def process(
|
238 |
+
self,
|
239 |
+
image_path: str,
|
240 |
+
audio_path: str,
|
241 |
+
output_path: str,
|
242 |
+
min_resolution: int = 512,
|
243 |
+
inference_steps: int = 25,
|
244 |
+
dynamic_scale: float = 1.0,
|
245 |
+
keep_resolution: bool = False,
|
246 |
+
seed: int | None = None,
|
247 |
+
):
|
248 |
+
cfg = self.config
|
249 |
+
if seed is not None:
|
250 |
+
cfg.seed = seed
|
251 |
+
cfg.num_inference_steps = inference_steps
|
252 |
+
cfg.motion_bucket_scale = dynamic_scale
|
253 |
+
seed_everything(cfg.seed)
|
254 |
+
|
255 |
+
# ----------------------------------------------------------
|
256 |
+
# 이미지·오디오 → 텐서
|
257 |
+
# ----------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
test_data = image_audio_to_tensor(
|
259 |
+
self.face_det,
|
260 |
+
self.feature_extractor,
|
261 |
+
image_path,
|
262 |
+
audio_path,
|
263 |
+
limit=-1, # 전체 오디오 사용
|
264 |
+
image_size=min_resolution,
|
265 |
+
area=cfg.area,
|
266 |
)
|
267 |
if test_data is None:
|
268 |
return -1
|
269 |
+
|
270 |
+
h, w = test_data["ref_img"].shape[-2:]
|
271 |
+
resolution = (
|
272 |
+
f"{(Image.open(image_path).width // 2)*2}x{(Image.open(image_path).height // 2)*2}"
|
273 |
+
if keep_resolution
|
274 |
+
else f"{w}x{h}"
|
275 |
+
)
|
276 |
+
|
277 |
+
# ----------------------------------------------------------
|
278 |
+
# 프레임 생성
|
279 |
+
# ----------------------------------------------------------
|
280 |
video = test(
|
281 |
+
self.pipe,
|
282 |
+
cfg,
|
283 |
+
wav_enc=self.whisper,
|
284 |
+
audio_pe=self.audio2token,
|
285 |
+
audio2bucket=self.audio2bucket,
|
286 |
+
image_encoder=self.image_encoder,
|
287 |
+
width=w,
|
288 |
+
height=h,
|
289 |
batch=test_data,
|
290 |
)
|
291 |
|
292 |
+
# 중간 프레임 보간
|
293 |
+
if cfg.use_interframe:
|
294 |
+
out, results = video.to(self.device), []
|
295 |
+
for i in tqdm(range(out.shape[2] - 1), ncols=0):
|
296 |
+
I1, I2 = out[:, :, i], out[:, :, i + 1]
|
297 |
+
middle = self.rife.inference(I1, I2).clamp(0, 1).detach()
|
298 |
+
results.extend([out[:, :, i], middle])
|
299 |
+
results.append(out[:, :, -1])
|
|
|
|
|
|
|
|
|
|
|
300 |
video = torch.stack(results, 2).cpu()
|
|
|
|
|
|
|
301 |
|
302 |
+
# ----------------------------------------------------------
|
303 |
+
# 파일 저장
|
304 |
+
# ----------------------------------------------------------
|
305 |
+
tmp_video = output_path.replace(".mp4", "_noaudio.mp4")
|
306 |
+
save_videos_grid(video, tmp_video, n_rows=video.shape[0], fps=cfg.fps * (2 if cfg.use_interframe else 1))
|
307 |
os.system(
|
308 |
+
f"ffmpeg -i '{tmp_video}' -i '{audio_path}' -s {resolution} "
|
309 |
+
f"-vcodec libx264 -acodec aac -crf 18 -shortest '{output_path}' -y -loglevel error"
|
310 |
)
|
311 |
+
os.remove(tmp_video)
|
312 |
return 0
|