Kerry
4b731ff98b
Float project applicable to ComfyUI.Generates speaking portrait video frames from an image and audio.适用于ComfyUI的强大[FLOAT]项目，实现由肖像生成音频驱动的说话视频。
74 lines
2.6 KiB
Python
74 lines
2.6 KiB
Python
# ComfyUI_Float_Animator/options/__init__.py
|
|
|
|
from types import SimpleNamespace
|
|
|
|
# BaseOptionsJson will act as a SimpleNamespace to hold model options,
|
|
# replacing the argparse.Namespace used in the original generate.py.
|
|
class BaseOptionsJson(SimpleNamespace):
|
|
def __init__(self, **kwargs):
|
|
super().__init__()
|
|
# Initialize with default values from official base_options.py
|
|
# These are the default values if no specific argument is passed
|
|
self.pretrained_dir = './checkpoints'
|
|
self.seed = 15
|
|
self.fix_noise_seed = False # Original default is False, not an action_store_true
|
|
|
|
# video
|
|
self.input_size = 512
|
|
self.input_nc = 3
|
|
self.fps = 25.0
|
|
|
|
# audio
|
|
self.sampling_rate = 16000
|
|
self.audio_marcing = 2
|
|
self.wav2vec_sec = 2.0
|
|
self.wav2vec_model_path = './checkpoints/wav2vec2-base-960h' # These will be overwritten by Node.py
|
|
self.audio2emotion_path = './checkpoints/wav2vec-english-speech-emotion-recognition' # These will be overwritten by Node.py
|
|
self.attention_window = 2
|
|
|
|
self.only_last_features = False # Original default is False
|
|
self.average_emotion = False # Original default is False
|
|
|
|
# dropout
|
|
self.audio_dropout_prob = 0.1
|
|
self.ref_dropout_prob = 0.1
|
|
self.emotion_dropout_prob = 0.1
|
|
|
|
# model Hyper Parameters
|
|
self.style_dim = 512
|
|
self.dim_a = 512
|
|
self.dim_w = 512
|
|
self.dim_h = 1024
|
|
self.dim_m = 20
|
|
self.dim_e = 7
|
|
|
|
# option for FMT
|
|
self.fmt_depth = 8
|
|
self.num_heads = 8
|
|
self.mlp_ratio = 4.0
|
|
self.no_learned_pe = False # Original default is False
|
|
self.num_prev_frames = 10
|
|
self.max_grad_norm = 1.0
|
|
|
|
self.ode_atol = 1e-5
|
|
self.ode_rtol = 1e-5
|
|
self.nfe = 10
|
|
self.torchdiffeq_ode_method = 'euler'
|
|
self.a_cfg_scale = 2.0
|
|
self.e_cfg_scale = 1.0
|
|
self.r_cfg_scale = 1.0
|
|
|
|
# option for Diffusion (ablation)
|
|
self.n_diff_steps = 500
|
|
self.diff_schedule = 'cosine'
|
|
self.diffusion_mode = 'sample'
|
|
|
|
# Apply any kwargs to override defaults
|
|
for k, v in kwargs.items():
|
|
setattr(self, k, v)
|
|
|
|
# Additional ComfyUI specific parameters (will be set by Node.py)
|
|
self.rank = None # Will be set by ComfyUI's model management
|
|
self.ngpus = 1 # Will be set by ComfyUI's model management
|
|
self.ckpt_path = None # Will be set by Node.py
|
|
self.res_dir = "./results" # Default output dir, though Node.py won't use it directly for saving |