import os, math, torch, cv2 from PIL import Image from omegaconf import OmegaConf from tqdm import tqdm from diffusers import AutoencoderKLTemporalDecoder from diffusers.schedulers import EulerDiscreteScheduler from transformers import WhisperModel, CLIPVisionModelWithProjection, AutoFeatureExtractor from src.utils.util import save_videos_grid, seed_everything from src.dataset.test_preprocess import process_bbox, image_audio_to_tensor from src.models.base.unet_spatio_temporal_condition import ( UNetSpatioTemporalConditionModel, add_ip_adapters, ) from src.pipelines.pipeline_sonic import SonicPipeline from src.models.audio_adapter.audio_proj import AudioProjModel from src.models.audio_adapter.audio_to_bucket import Audio2bucketModel from src.utils.RIFE.RIFE_HDv3 import RIFEModel from src.dataset.face_align.align import AlignImage BASE_DIR = os.path.dirname(os.path.abspath(__file__)) # ------------------------------------------------------------------ # single image + speech → video-tensor generator # ------------------------------------------------------------------ def test( pipe, config, wav_enc, audio_pe, audio2bucket, image_encoder, width, height, batch, ): # --- 배치 차원 맞추기 -------------------------------------------------- for k, v in batch.items(): if isinstance(v, torch.Tensor): batch[k] = v.unsqueeze(0).to(pipe.device).float() ref_img = batch["ref_img"] # (1,C,H,W) clip_img = batch["clip_images"] face_mask = batch["face_mask"] image_embeds = image_encoder(clip_img).image_embeds audio_feature = batch["audio_feature"] # (1,80,T) audio_len = int(batch["audio_len"]) # Python int step = int(config.step) # --- step 보정 (최소 1) ----------------------------------------------- if audio_len < step: step = max(1, audio_len) window = 16000 # 1초 chunk audio_prompts, last_prompts = [], [] # --- window 단위 Whisper 인코딩 -------------------------------------- for i in range(0, audio_feature.shape[-1], window): chunk = audio_feature[:, :, i : i + window] prompt_layers = wav_enc.encoder(chunk, output_hidden_states=True).hidden_states last_hidden = wav_enc.encoder(chunk).last_hidden_state.unsqueeze(-2) # (1,t,1,384) audio_prompts.append(torch.stack(prompt_layers, dim=2)) # (1,L,12,80) last_prompts.append(last_hidden) # (1,L,1,384) if len(audio_prompts) == 0: raise ValueError("[ERROR] No speech recognised in the provided audio.") audio_prompts = torch.cat(audio_prompts, dim=1) last_prompts = torch.cat(last_prompts, dim=1) # padding 규칙 audio_prompts = torch.cat( [torch.zeros_like(audio_prompts[:, :4]), audio_prompts, torch.zeros_like(audio_prompts[:, :6])], dim=1) last_prompts = torch.cat( [torch.zeros_like(last_prompts[:, :24]), last_prompts, torch.zeros_like(last_prompts[:, :26])], dim=1) # --- 반드시 ≥1 chunk -------------------------------------------------- total_tokens = audio_prompts.shape[1] num_chunks = max(1, math.ceil(total_tokens / (2 * step))) ref_list, audio_list, uncond_list, motion_buckets = [], [], [], [] for i in tqdm(range(num_chunks)): start = i * 2 * step cond_clip = audio_prompts[:, start : start + 10] # (1,10,12,80) if cond_clip.shape[1] < 10: # 짧으면 패딩 pad = torch.zeros_like(cond_clip[:, : 10 - cond_clip.shape[1]]) cond_clip = torch.cat([cond_clip, pad], dim=1) # ------------------ (★) bucket_clip 차원 맞춤 ------------------- bucket_clip = last_prompts[:, start : start + 50] # (1,50,1,384) if bucket_clip.shape[1] < 50: # 짧으면 패딩 pad = torch.zeros_like(bucket_clip[:, : 50 - bucket_clip.shape[1]]) bucket_clip = torch.cat([bucket_clip, pad], dim=1) bucket_clip = bucket_clip.unsqueeze(1) # → (1,1,50,1,384) ✔ 5-D # ----------------------------------------------------------------- motion = audio2bucket(bucket_clip, image_embeds) * 16 + 16 ref_list.append(ref_img[0]) audio_list.append(audio_pe(cond_clip.unsqueeze(1)).squeeze(0)[0]) # (10,···)→ unsqueeze 후 4-D uncond_list.append(audio_pe(torch.zeros_like(cond_clip).unsqueeze(1)).squeeze(0)[0]) motion_buckets.append(motion[0]) # ---------------------------------------------------------------------- video = pipe( ref_img, clip_img, face_mask, audio_list, uncond_list, motion_buckets, height=height, width=width, num_frames=len(audio_list), decode_chunk_size=config.decode_chunk_size, motion_bucket_scale=config.motion_bucket_scale, fps=config.fps, noise_aug_strength=config.noise_aug_strength, min_guidance_scale1=config.min_appearance_guidance_scale, max_guidance_scale1=config.max_appearance_guidance_scale, min_guidance_scale2=config.audio_guidance_scale, max_guidance_scale2=config.audio_guidance_scale, overlap=config.overlap, shift_offset=config.shift_offset, frames_per_batch=config.n_sample_frames, num_inference_steps=config.num_inference_steps, i2i_noise_strength=config.i2i_noise_strength, ).frames video = (video * 0.5 + 0.5).clamp(0, 1) return video.to(pipe.device).unsqueeze(0).cpu() # ------------------------------------------------------------------ # Sonic 클래스 # ------------------------------------------------------------------ class Sonic: config_file = os.path.join(BASE_DIR, "config/inference/sonic.yaml") config = OmegaConf.load(config_file) def __init__(self, device_id: int = 0, enable_interpolate_frame: bool = True): cfg = self.config cfg.use_interframe = enable_interpolate_frame self.device = f"cuda:{device_id}" if device_id >= 0 and torch.cuda.is_available() else "cpu" cfg.pretrained_model_name_or_path = os.path.join(BASE_DIR, cfg.pretrained_model_name_or_path) self._load_models(cfg) print("Sonic init done") # -------------------------------------------------------------- def _load_models(self, cfg): dtype = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16}[cfg.weight_dtype] vae = AutoencoderKLTemporalDecoder.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="vae", variant="fp16") sched = EulerDiscreteScheduler.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="scheduler") imgE = CLIPVisionModelWithProjection.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="image_encoder", variant="fp16") unet = UNetSpatioTemporalConditionModel.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="unet", variant="fp16") add_ip_adapters(unet, [32], [cfg.ip_audio_scale]) a2t = AudioProjModel(10, 5, 384, 1024, 1024, 32).to(self.device) a2b = Audio2bucketModel(50, 1, 384, 1024, 1024, 1, 2).to(self.device) unet.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.unet_checkpoint_path), map_location="cpu")) a2t.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2token_checkpoint_path), map_location="cpu")) a2b.load_state_dict(torch.load(os.path.join(BASE_DIR, cfg.audio2bucket_checkpoint_path), map_location="cpu")) whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny")).to(self.device).eval() whisper.requires_grad_(False) self.feature_extractor = AutoFeatureExtractor.from_pretrained(os.path.join(BASE_DIR, "checkpoints/whisper-tiny")) self.face_det = AlignImage(self.device, det_path=os.path.join(BASE_DIR, "checkpoints/yoloface_v5m.pt")) if cfg.use_interframe: self.rife = RIFEModel(device=self.device) self.rife.load_model(os.path.join(BASE_DIR, "checkpoints/RIFE/")) for m in (imgE, vae, unet): m.to(dtype) self.pipe = SonicPipeline(unet=unet, image_encoder=imgE, vae=vae, scheduler=sched).to(device=self.device, dtype=dtype) self.image_encoder = imgE self.audio2token = a2t self.audio2bucket = a2b self.whisper = whisper # -------------------------------------------------------------- def preprocess(self, image_path: str, expand_ratio: float = 1.0): img = cv2.imread(image_path) h, w = img.shape[:2] _, _, bboxes = self.face_det(img, maxface=True) if bboxes: x1, y1, ww, hh = bboxes[0] return {"face_num": 1, "crop_bbox": process_bbox((x1, y1, x1 + ww, y1 + hh), expand_ratio, h, w)} return {"face_num": 0, "crop_bbox": None} # -------------------------------------------------------------- @torch.no_grad() def process( self, image_path: str, audio_path: str, output_path: str, min_resolution: int = 512, inference_steps: int = 25, dynamic_scale: float = 1.0, keep_resolution: bool = False, seed: int | None = None, ): cfg = self.config if seed is not None: cfg.seed = seed cfg.num_inference_steps = inference_steps cfg.motion_bucket_scale = dynamic_scale seed_everything(cfg.seed) # 이미지·오디오 → tensor test_data = image_audio_to_tensor( self.face_det, self.feature_extractor, image_path, audio_path, limit=-1, image_size=min_resolution, area=cfg.area, ) if test_data is None: return -1 h, w = test_data["ref_img"].shape[-2:] resolution = ( f"{(Image.open(image_path).width // 2) * 2}x{(Image.open(image_path).height // 2) * 2}" if keep_resolution else f"{w}x{h}" ) # 비디오 프레임 생성 video = test( self.pipe, cfg, wav_enc=self.whisper, audio_pe=self.audio2token, audio2bucket=self.audio2bucket, image_encoder=self.image_encoder, width=w, height=h, batch=test_data, ) # 중간 프레임 보간 if cfg.use_interframe: out = video.to(self.device) frames = [] for i in tqdm(range(out.shape[2] - 1), ncols=0): mid = self.rife.inference(out[:, :, i], out[:, :, i + 1]).clamp(0, 1).detach() frames.extend([out[:, :, i], mid]) frames.append(out[:, :, -1]) video = torch.stack(frames, 2).cpu() # 저장 tmp = output_path.replace(".mp4", "_noaudio.mp4") save_videos_grid(video, tmp, n_rows=video.shape[0], fps=cfg.fps * (2 if cfg.use_interframe else 1)) os.system( f"ffmpeg -i '{tmp}' -i '{audio_path}' -s {resolution} " f"-vcodec libx264 -acodec aac -crf 18 -shortest '{output_path}' -y -loglevel error" ) os.remove(tmp) return 0