Spaces:
Running
on
Zero
Running
on
Zero
Update sonic.py
Browse files
sonic.py
CHANGED
@@ -32,9 +32,9 @@ def test(
|
|
32 |
image_encoder,
|
33 |
width,
|
34 |
height,
|
35 |
-
batch
|
36 |
):
|
37 |
-
"""
|
38 |
for k, v in batch.items():
|
39 |
if isinstance(v, torch.Tensor):
|
40 |
batch[k] = v.unsqueeze(0).to(pipe.device).float()
|
@@ -52,30 +52,36 @@ def test(
|
|
52 |
audio_prompts = []
|
53 |
last_audio_prompts = []
|
54 |
for i in range(0, audio_feature.shape[-1], window):
|
55 |
-
audio_prompt = wav_enc.encoder(audio_feature[:, :, i:i
|
56 |
-
last_audio_prompt = wav_enc.encoder(audio_feature[:, :, i:i
|
57 |
last_audio_prompt = last_audio_prompt.unsqueeze(-2)
|
58 |
audio_prompt = torch.stack(audio_prompt, dim=2)
|
59 |
audio_prompts.append(audio_prompt)
|
60 |
last_audio_prompts.append(last_audio_prompt)
|
61 |
|
62 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
63 |
-
audio_prompts = audio_prompts[:, :audio_len
|
64 |
-
audio_prompts = torch.cat([
|
65 |
-
|
|
|
|
|
|
|
66 |
|
67 |
last_audio_prompts = torch.cat(last_audio_prompts, dim=1)
|
68 |
-
last_audio_prompts = last_audio_prompts[:, :audio_len
|
69 |
-
last_audio_prompts = torch.cat([
|
70 |
-
|
|
|
|
|
|
|
71 |
|
72 |
ref_tensor_list = []
|
73 |
audio_tensor_list = []
|
74 |
uncond_audio_tensor_list = []
|
75 |
motion_buckets = []
|
76 |
-
for i in tqdm(range(audio_len
|
77 |
-
audio_clip = audio_prompts[:, i
|
78 |
-
audio_clip_for_bucket = last_audio_prompts[:, i
|
79 |
motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds)
|
80 |
motion_bucket = motion_bucket * 16 + 16
|
81 |
motion_buckets.append(motion_bucket[0])
|
@@ -114,100 +120,67 @@ def test(
|
|
114 |
|
115 |
video = (video * 0.5 + 0.5).clamp(0, 1)
|
116 |
video = torch.cat([video.to(pipe.device)], dim=0).cpu()
|
117 |
-
|
118 |
return video
|
119 |
|
120 |
|
121 |
class Sonic:
|
122 |
-
"""
|
123 |
|
124 |
config_file = os.path.join(BASE_DIR, 'config/inference/sonic.yaml')
|
125 |
config = OmegaConf.load(config_file)
|
126 |
|
127 |
def __init__(self, device_id: int = 0, enable_interpolate_frame: bool = True):
|
128 |
-
# --------- load config & device ---------
|
129 |
config = self.config
|
130 |
config.use_interframe = enable_interpolate_frame
|
131 |
|
132 |
device = f'cuda:{device_id}' if device_id > -1 else 'cpu'
|
133 |
self.device = device
|
134 |
|
135 |
-
# --------- Model paths ---------
|
136 |
config.pretrained_model_name_or_path = os.path.join(BASE_DIR, config.pretrained_model_name_or_path)
|
137 |
|
138 |
-
# --------- Load sub‑modules ---------
|
139 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
140 |
-
config.pretrained_model_name_or_path,
|
141 |
-
subfolder="vae",
|
142 |
-
variant="fp16"
|
143 |
-
)
|
144 |
-
|
145 |
val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
|
146 |
-
config.pretrained_model_name_or_path,
|
147 |
-
subfolder="scheduler"
|
148 |
-
)
|
149 |
-
|
150 |
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
151 |
-
config.pretrained_model_name_or_path,
|
152 |
-
subfolder="image_encoder",
|
153 |
-
variant="fp16"
|
154 |
-
)
|
155 |
-
|
156 |
unet = UNetSpatioTemporalConditionModel.from_pretrained(
|
157 |
-
config.pretrained_model_name_or_path,
|
158 |
-
subfolder="unet",
|
159 |
-
variant="fp16"
|
160 |
-
)
|
161 |
add_ip_adapters(unet, [32], [config.ip_audio_scale])
|
162 |
|
163 |
-
audio2token = AudioProjModel(seq_len=10, blocks=5, channels=384, intermediate_dim=1024,
|
164 |
-
context_tokens=32).to(device)
|
165 |
-
audio2bucket = Audio2bucketModel(seq_len=50, blocks=1, channels=384, clip_channels=1024,
|
166 |
-
output_dim=1, context_tokens=2).to(device)
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
if config.weight_dtype == "fp16":
|
179 |
-
weight_dtype = torch.float16
|
180 |
-
elif config.weight_dtype == "fp32":
|
181 |
-
weight_dtype = torch.float32
|
182 |
-
elif config.weight_dtype == "bf16":
|
183 |
-
weight_dtype = torch.bfloat16
|
184 |
-
else:
|
185 |
raise ValueError(f"Unsupported weight dtype: {config.weight_dtype}")
|
186 |
|
187 |
-
# --------- Whisper encoder for audio ---------
|
188 |
whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/')).to(device).eval()
|
189 |
whisper.requires_grad_(False)
|
190 |
-
self.feature_extractor = AutoFeatureExtractor.from_pretrained(
|
|
|
191 |
|
192 |
-
|
193 |
-
det_path = os.path.join(BASE_DIR, 'checkpoints/yoloface_v5m.pt')
|
194 |
-
self.face_det = AlignImage(device, det_path=det_path)
|
195 |
if config.use_interframe:
|
196 |
self.rife = RIFEModel(device=device)
|
197 |
self.rife.load_model(os.path.join(BASE_DIR, 'checkpoints', 'RIFE/'))
|
198 |
|
199 |
-
# --------- Move modules to device & dtype ---------
|
200 |
image_encoder.to(weight_dtype)
|
201 |
vae.to(weight_dtype)
|
202 |
unet.to(weight_dtype)
|
203 |
|
204 |
-
# --------- Compose pipeline ---------
|
205 |
pipe = SonicPipeline(
|
206 |
-
unet=unet,
|
207 |
-
image_encoder=image_encoder,
|
208 |
-
vae=vae,
|
209 |
-
scheduler=val_noise_scheduler,
|
210 |
-
)
|
211 |
self.pipe = pipe.to(device=device, dtype=weight_dtype)
|
212 |
self.whisper = whisper
|
213 |
self.audio2token = audio2token
|
@@ -216,9 +189,7 @@ class Sonic:
|
|
216 |
|
217 |
print('Sonic initialization complete.')
|
218 |
|
219 |
-
# -------------------------- Public helpers --------------------------
|
220 |
def preprocess(self, image_path: str, expand_ratio: float = 1.0):
|
221 |
-
"""Detect face and compute crop bbox (optional)."""
|
222 |
face_image = cv2.imread(image_path)
|
223 |
h, w = face_image.shape[:2]
|
224 |
_, _, bboxes = self.face_det(face_image, maxface=True)
|
@@ -227,15 +198,63 @@ class Sonic:
|
|
227 |
if face_num > 0:
|
228 |
x1, y1, ww, hh = bboxes[0]
|
229 |
x2, y2 = x1 + ww, y1 + hh
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
return {
|
234 |
-
'face_num': face_num,
|
235 |
-
'crop_bbox': bbox_s,
|
236 |
-
}
|
237 |
|
238 |
def crop_image(self, input_image_path: str, output_image_path: str, crop_bbox):
|
239 |
face_image = cv2.imread(input_image_path)
|
240 |
-
|
241 |
-
cv2.imwrite(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
image_encoder,
|
33 |
width,
|
34 |
height,
|
35 |
+
batch,
|
36 |
):
|
37 |
+
"""Generate a video tensor for the given batch."""
|
38 |
for k, v in batch.items():
|
39 |
if isinstance(v, torch.Tensor):
|
40 |
batch[k] = v.unsqueeze(0).to(pipe.device).float()
|
|
|
52 |
audio_prompts = []
|
53 |
last_audio_prompts = []
|
54 |
for i in range(0, audio_feature.shape[-1], window):
|
55 |
+
audio_prompt = wav_enc.encoder(audio_feature[:, :, i:i+window], output_hidden_states=True).hidden_states
|
56 |
+
last_audio_prompt = wav_enc.encoder(audio_feature[:, :, i:i+window]).last_hidden_state
|
57 |
last_audio_prompt = last_audio_prompt.unsqueeze(-2)
|
58 |
audio_prompt = torch.stack(audio_prompt, dim=2)
|
59 |
audio_prompts.append(audio_prompt)
|
60 |
last_audio_prompts.append(last_audio_prompt)
|
61 |
|
62 |
audio_prompts = torch.cat(audio_prompts, dim=1)
|
63 |
+
audio_prompts = audio_prompts[:, :audio_len*2]
|
64 |
+
audio_prompts = torch.cat([
|
65 |
+
torch.zeros_like(audio_prompts[:, :4]),
|
66 |
+
audio_prompts,
|
67 |
+
torch.zeros_like(audio_prompts[:, :6])
|
68 |
+
], 1)
|
69 |
|
70 |
last_audio_prompts = torch.cat(last_audio_prompts, dim=1)
|
71 |
+
last_audio_prompts = last_audio_prompts[:, :audio_len*2]
|
72 |
+
last_audio_prompts = torch.cat([
|
73 |
+
torch.zeros_like(last_audio_prompts[:, :24]),
|
74 |
+
last_audio_prompts,
|
75 |
+
torch.zeros_like(last_audio_prompts[:, :26])
|
76 |
+
], 1)
|
77 |
|
78 |
ref_tensor_list = []
|
79 |
audio_tensor_list = []
|
80 |
uncond_audio_tensor_list = []
|
81 |
motion_buckets = []
|
82 |
+
for i in tqdm(range(audio_len//step), ncols=0):
|
83 |
+
audio_clip = audio_prompts[:, i*2*step:i*2*step+10].unsqueeze(0)
|
84 |
+
audio_clip_for_bucket = last_audio_prompts[:, i*2*step:i*2*step+50].unsqueeze(0)
|
85 |
motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds)
|
86 |
motion_bucket = motion_bucket * 16 + 16
|
87 |
motion_buckets.append(motion_bucket[0])
|
|
|
120 |
|
121 |
video = (video * 0.5 + 0.5).clamp(0, 1)
|
122 |
video = torch.cat([video.to(pipe.device)], dim=0).cpu()
|
|
|
123 |
return video
|
124 |
|
125 |
|
126 |
class Sonic:
|
127 |
+
"""High-level interface for the Sonic portrait animation pipeline."""
|
128 |
|
129 |
config_file = os.path.join(BASE_DIR, 'config/inference/sonic.yaml')
|
130 |
config = OmegaConf.load(config_file)
|
131 |
|
132 |
def __init__(self, device_id: int = 0, enable_interpolate_frame: bool = True):
|
|
|
133 |
config = self.config
|
134 |
config.use_interframe = enable_interpolate_frame
|
135 |
|
136 |
device = f'cuda:{device_id}' if device_id > -1 else 'cpu'
|
137 |
self.device = device
|
138 |
|
|
|
139 |
config.pretrained_model_name_or_path = os.path.join(BASE_DIR, config.pretrained_model_name_or_path)
|
140 |
|
|
|
141 |
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
142 |
+
config.pretrained_model_name_or_path, subfolder='vae', variant='fp16')
|
|
|
|
|
|
|
|
|
143 |
val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
|
144 |
+
config.pretrained_model_name_or_path, subfolder='scheduler')
|
|
|
|
|
|
|
145 |
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
146 |
+
config.pretrained_model_name_or_path, subfolder='image_encoder', variant='fp16')
|
|
|
|
|
|
|
|
|
147 |
unet = UNetSpatioTemporalConditionModel.from_pretrained(
|
148 |
+
config.pretrained_model_name_or_path, subfolder='unet', variant='fp16')
|
|
|
|
|
|
|
149 |
add_ip_adapters(unet, [32], [config.ip_audio_scale])
|
150 |
|
151 |
+
audio2token = AudioProjModel(seq_len=10, blocks=5, channels=384, intermediate_dim=1024,
|
152 |
+
output_dim=1024, context_tokens=32).to(device)
|
153 |
+
audio2bucket = Audio2bucketModel(seq_len=50, blocks=1, channels=384, clip_channels=1024,
|
154 |
+
intermediate_dim=1024, output_dim=1, context_tokens=2).to(device)
|
155 |
+
|
156 |
+
unet.load_state_dict(
|
157 |
+
torch.load(os.path.join(BASE_DIR, config.unet_checkpoint_path), map_location='cpu'), strict=True)
|
158 |
+
audio2token.load_state_dict(
|
159 |
+
torch.load(os.path.join(BASE_DIR, config.audio2token_checkpoint_path), map_location='cpu'), strict=True)
|
160 |
+
audio2bucket.load_state_dict(
|
161 |
+
torch.load(os.path.join(BASE_DIR, config.audio2bucket_checkpoint_path), map_location='cpu'), strict=True)
|
162 |
+
|
163 |
+
dtype_map = {'fp16': torch.float16, 'fp32': torch.float32, 'bf16': torch.bfloat16}
|
164 |
+
weight_dtype = dtype_map.get(config.weight_dtype)
|
165 |
+
if weight_dtype is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
raise ValueError(f"Unsupported weight dtype: {config.weight_dtype}")
|
167 |
|
|
|
168 |
whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/')).to(device).eval()
|
169 |
whisper.requires_grad_(False)
|
170 |
+
self.feature_extractor = AutoFeatureExtractor.from_pretrained(
|
171 |
+
os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/'))
|
172 |
|
173 |
+
self.face_det = AlignImage(device, det_path=os.path.join(BASE_DIR, 'checkpoints/yoloface_v5m.pt'))
|
|
|
|
|
174 |
if config.use_interframe:
|
175 |
self.rife = RIFEModel(device=device)
|
176 |
self.rife.load_model(os.path.join(BASE_DIR, 'checkpoints', 'RIFE/'))
|
177 |
|
|
|
178 |
image_encoder.to(weight_dtype)
|
179 |
vae.to(weight_dtype)
|
180 |
unet.to(weight_dtype)
|
181 |
|
|
|
182 |
pipe = SonicPipeline(
|
183 |
+
unet=unet, image_encoder=image_encoder, vae=vae, scheduler=val_noise_scheduler)
|
|
|
|
|
|
|
|
|
184 |
self.pipe = pipe.to(device=device, dtype=weight_dtype)
|
185 |
self.whisper = whisper
|
186 |
self.audio2token = audio2token
|
|
|
189 |
|
190 |
print('Sonic initialization complete.')
|
191 |
|
|
|
192 |
def preprocess(self, image_path: str, expand_ratio: float = 1.0):
|
|
|
193 |
face_image = cv2.imread(image_path)
|
194 |
h, w = face_image.shape[:2]
|
195 |
_, _, bboxes = self.face_det(face_image, maxface=True)
|
|
|
198 |
if face_num > 0:
|
199 |
x1, y1, ww, hh = bboxes[0]
|
200 |
x2, y2 = x1 + ww, y1 + hh
|
201 |
+
bbox_s = process_bbox((x1, y1, x2, y2), expand_radio=expand_ratio, height=h, width=w)
|
202 |
+
return {'face_num': face_num, 'crop_bbox': bbox_s}
|
|
|
|
|
|
|
|
|
|
|
203 |
|
204 |
def crop_image(self, input_image_path: str, output_image_path: str, crop_bbox):
|
205 |
face_image = cv2.imread(input_image_path)
|
206 |
+
crop_img = face_image[crop_bbox[1]:crop_bbox[3], crop_bbox[0]:crop_bbox[2]]
|
207 |
+
cv2.imwrite(output_image_path, crop_img)
|
208 |
+
|
209 |
+
@torch.no_grad()
|
210 |
+
def process(self, image_path, audio_path, output_path, min_resolution=512,
|
211 |
+
inference_steps=25, dynamic_scale=1.0, keep_resolution=False, seed=None):
|
212 |
+
config = self.config
|
213 |
+
device = self.device
|
214 |
+
|
215 |
+
pipe = self.pipe
|
216 |
+
whisper = self.whisper
|
217 |
+
audio2token = self.audio2token
|
218 |
+
audio2bucket = self.audio2bucket
|
219 |
+
image_encoder = self.image_encoder
|
220 |
+
|
221 |
+
if seed is not None:
|
222 |
+
config.seed = seed
|
223 |
+
seed_everything(config.seed)
|
224 |
+
|
225 |
+
config.num_inference_steps = inference_steps
|
226 |
+
config.frame_num = config.fps * 60
|
227 |
+
config.motion_bucket_scale = dynamic_scale
|
228 |
+
|
229 |
+
video_path = output_path.replace('.mp4', '_noaudio.mp4')
|
230 |
+
audio_video_path = output_path
|
231 |
+
|
232 |
+
imSrc_ = Image.open(image_path).convert('RGB')
|
233 |
+
raw_w, raw_h = imSrc_.size
|
234 |
+
|
235 |
+
test_data = image_audio_to_tensor(
|
236 |
+
self.face_det, self.feature_extractor, image_path, audio_path,
|
237 |
+
limit=config.frame_num, image_size=min_resolution, area=config.area)
|
238 |
+
if test_data is None:
|
239 |
+
return -1
|
240 |
+
height, width = test_data['ref_img'].shape[-2:]
|
241 |
+
resolution = f"{width}x{height}" if not keep_resolution else f"{raw_w//2*2}x{raw_h//2*2}"
|
242 |
+
|
243 |
+
video = test(pipe, config, wav_enc=whisper, audio_pe=audio2token,
|
244 |
+
audio2bucket=audio2bucket, image_encoder=image_encoder,
|
245 |
+
width=width, height=height, batch=test_data)
|
246 |
+
|
247 |
+
if config.use_interframe:
|
248 |
+
out = video.to(device)
|
249 |
+
results = []
|
250 |
+
for idx in tqdm(range(out.shape[2]-1), ncols=0):
|
251 |
+
I1 = out[:, :, idx]
|
252 |
+
I2 = out[:, :, idx+1]
|
253 |
+
mid = self.rife.inference(I1, I2).clamp(0,1).detach()
|
254 |
+
results.extend([out[:, :, idx], mid])
|
255 |
+
results.append(out[:, :, -1])
|
256 |
+
video = torch.stack(results, 2).cpu()
|
257 |
+
|
258 |
+
save_videos_grid(video, video_path, n_rows=video.shape[0], fps=config.fps * (2 if config.use_interframe else 1))
|
259 |
+
os.system(f"ffmpeg -i '{video_path}' -i '{audio_path}' -s {resolution} -vcodec libx264 -acodec aac -crf 18 -shortest '{audio_video_path}' -y; rm '{video_path}'")
|
260 |
+
return 0
|