facefusion/processors/modules/lip_syncer.py

from argparse import ArgumentParser
from functools import lru_cache
from typing import List

import cv2
import numpy

import facefusion.jobs.job_manager
import facefusion.jobs.job_store
import facefusion.processors.core as processors
from facefusion import config, content_analyser, face_classifier, face_detector, face_landmarker, face_masker, face_recognizer, inference_manager, logger, process_manager, state_manager, voice_extractor, wording
from facefusion.audio import create_empty_audio_frame, get_voice_frame, read_static_voice
from facefusion.common_helper import get_first
from facefusion.download import conditional_download_hashes, conditional_download_sources, resolve_download_url
from facefusion.face_analyser import get_many_faces, get_one_face
from facefusion.face_helper import create_bounding_box, paste_back, warp_face_by_bounding_box, warp_face_by_face_landmark_5
from facefusion.face_masker import create_mouth_mask, create_occlusion_mask, create_static_box_mask
from facefusion.face_selector import find_similar_faces, sort_and_filter_faces
from facefusion.face_store import get_reference_faces
from facefusion.filesystem import filter_audio_paths, has_audio, in_directory, is_image, is_video, resolve_relative_path, same_file_extension
from facefusion.processors import choices as processors_choices
from facefusion.processors.types import LipSyncerInputs
from facefusion.program_helper import find_argument_group
from facefusion.thread_helper import conditional_thread_semaphore
from facefusion.types import ApplyStateItem, Args, AudioFrame, DownloadScope, Face, InferencePool, ModelOptions, ModelSet, ProcessMode, QueuePayload, UpdateProgress, VisionFrame
from facefusion.vision import read_image, read_static_image, restrict_video_fps, write_image


@lru_cache(maxsize = None)
def create_static_model_set(download_scope : DownloadScope) -> ModelSet:
	return\
	{
		'wav2lip_96':
		{
			'hashes':
			{
				'lip_syncer':
				{
					'url': resolve_download_url('models-3.0.0', 'wav2lip_96.hash'),
					'path': resolve_relative_path('../.assets/models/wav2lip_96.hash')
				}
			},
			'sources':
			{
				'lip_syncer':
				{
					'url': resolve_download_url('models-3.0.0', 'wav2lip_96.onnx'),
					'path': resolve_relative_path('../.assets/models/wav2lip_96.onnx')
				}
			},
			'size': (96, 96)
		},
		'wav2lip_gan_96':
		{
			'hashes':
			{
				'lip_syncer':
				{
					'url': resolve_download_url('models-3.0.0', 'wav2lip_gan_96.hash'),
					'path': resolve_relative_path('../.assets/models/wav2lip_gan_96.hash')
				}
			},
			'sources':
			{
				'lip_syncer':
				{
					'url': resolve_download_url('models-3.0.0', 'wav2lip_gan_96.onnx'),
					'path': resolve_relative_path('../.assets/models/wav2lip_gan_96.onnx')
				}
			},
			'size': (96, 96)
		}
	}


def get_inference_pool() -> InferencePool:
	model_names = [ state_manager.get_item('lip_syncer_model') ]
	model_source_set = get_model_options().get('sources')

	return inference_manager.get_inference_pool(__name__, model_names, model_source_set)


def clear_inference_pool() -> None:
	model_names = [ state_manager.get_item('lip_syncer_model') ]
	inference_manager.clear_inference_pool(__name__, model_names)


def get_model_options() -> ModelOptions:
	lip_syncer_model = state_manager.get_item('lip_syncer_model')
	return create_static_model_set('full').get(lip_syncer_model)


def register_args(program : ArgumentParser) -> None:
	group_processors = find_argument_group(program, 'processors')
	if group_processors:
		group_processors.add_argument('--lip-syncer-model', help = wording.get('help.lip_syncer_model'), default = config.get_str_value('processors', 'lip_syncer_model', 'wav2lip_gan_96'), choices = processors_choices.lip_syncer_models)
		facefusion.jobs.job_store.register_step_keys([ 'lip_syncer_model' ])


def apply_args(args : Args, apply_state_item : ApplyStateItem) -> None:
	apply_state_item('lip_syncer_model', args.get('lip_syncer_model'))


def pre_check() -> bool:
	model_hash_set = get_model_options().get('hashes')
	model_source_set = get_model_options().get('sources')

	return conditional_download_hashes(model_hash_set) and conditional_download_sources(model_source_set)


def pre_process(mode : ProcessMode) -> bool:
	if not has_audio(state_manager.get_item('source_paths')):
		logger.error(wording.get('choose_audio_source') + wording.get('exclamation_mark'), __name__)
		return False
	if mode in [ 'output', 'preview' ] and not is_image(state_manager.get_item('target_path')) and not is_video(state_manager.get_item('target_path')):
		logger.error(wording.get('choose_image_or_video_target') + wording.get('exclamation_mark'), __name__)
		return False
	if mode == 'output' and not in_directory(state_manager.get_item('output_path')):
		logger.error(wording.get('specify_image_or_video_output') + wording.get('exclamation_mark'), __name__)
		return False
	if mode == 'output' and not same_file_extension(state_manager.get_item('target_path'), state_manager.get_item('output_path')):
		logger.error(wording.get('match_target_and_output_extension') + wording.get('exclamation_mark'), __name__)
		return False
	return True


def post_process() -> None:
	read_static_image.cache_clear()
	read_static_voice.cache_clear()
	if state_manager.get_item('video_memory_strategy') in [ 'strict', 'moderate' ]:
		clear_inference_pool()
	if state_manager.get_item('video_memory_strategy') == 'strict':
		content_analyser.clear_inference_pool()
		face_classifier.clear_inference_pool()
		face_detector.clear_inference_pool()
		face_landmarker.clear_inference_pool()
		face_masker.clear_inference_pool()
		face_recognizer.clear_inference_pool()
		voice_extractor.clear_inference_pool()


def sync_lip(target_face : Face, temp_audio_frame : AudioFrame, temp_vision_frame : VisionFrame) -> VisionFrame:
	model_size = get_model_options().get('size')
	temp_audio_frame = prepare_audio_frame(temp_audio_frame)
	crop_vision_frame, affine_matrix = warp_face_by_face_landmark_5(temp_vision_frame, target_face.landmark_set.get('5/68'), 'ffhq_512', (512, 512))
	face_landmark_68 = cv2.transform(target_face.landmark_set.get('68').reshape(1, -1, 2), affine_matrix).reshape(-1, 2)
	bounding_box = create_bounding_box(face_landmark_68)
	bounding_box[1] -= numpy.abs(bounding_box[3] - bounding_box[1]) * 0.125
	mouth_mask = create_mouth_mask(face_landmark_68)
	box_mask = create_static_box_mask(crop_vision_frame.shape[:2][::-1], state_manager.get_item('face_mask_blur'), state_manager.get_item('face_mask_padding'))
	crop_masks =\
	[
		mouth_mask,
		box_mask
	]

	if 'occlusion' in state_manager.get_item('face_mask_types'):
		occlusion_mask = create_occlusion_mask(crop_vision_frame)
		crop_masks.append(occlusion_mask)

	close_vision_frame, close_matrix = warp_face_by_bounding_box(crop_vision_frame, bounding_box, model_size)
	close_vision_frame = prepare_crop_frame(close_vision_frame)
	close_vision_frame = forward(temp_audio_frame, close_vision_frame)
	close_vision_frame = normalize_close_frame(close_vision_frame)
	crop_vision_frame = cv2.warpAffine(close_vision_frame, cv2.invertAffineTransform(close_matrix), (512, 512), borderMode = cv2.BORDER_REPLICATE)
	crop_mask = numpy.minimum.reduce(crop_masks)
	paste_vision_frame = paste_back(temp_vision_frame, crop_vision_frame, crop_mask, affine_matrix)
	return paste_vision_frame


def forward(temp_audio_frame : AudioFrame, close_vision_frame : VisionFrame) -> VisionFrame:
	lip_syncer = get_inference_pool().get('lip_syncer')

	with conditional_thread_semaphore():
		close_vision_frame = lip_syncer.run(None,
		{
			'source': temp_audio_frame,
			'target': close_vision_frame
		})[0]

	return close_vision_frame


def prepare_audio_frame(temp_audio_frame : AudioFrame) -> AudioFrame:
	temp_audio_frame = numpy.maximum(numpy.exp(-5 * numpy.log(10)), temp_audio_frame)
	temp_audio_frame = numpy.log10(temp_audio_frame) * 1.6 + 3.2
	temp_audio_frame = temp_audio_frame.clip(-4, 4).astype(numpy.float32)
	temp_audio_frame = numpy.expand_dims(temp_audio_frame, axis = (0, 1))
	return temp_audio_frame


def prepare_crop_frame(crop_vision_frame : VisionFrame) -> VisionFrame:
	crop_vision_frame = numpy.expand_dims(crop_vision_frame, axis = 0)
	prepare_vision_frame = crop_vision_frame.copy()
	prepare_vision_frame[:, 48:] = 0
	crop_vision_frame = numpy.concatenate((prepare_vision_frame, crop_vision_frame), axis = 3)
	crop_vision_frame = crop_vision_frame.transpose(0, 3, 1, 2).astype('float32') / 255.0
	return crop_vision_frame


def normalize_close_frame(crop_vision_frame : VisionFrame) -> VisionFrame:
	crop_vision_frame = crop_vision_frame[0].transpose(1, 2, 0)
	crop_vision_frame = crop_vision_frame.clip(0, 1) * 255
	crop_vision_frame = crop_vision_frame.astype(numpy.uint8)
	return crop_vision_frame


def get_reference_frame(source_face : Face, target_face : Face, temp_vision_frame : VisionFrame) -> VisionFrame:
	pass


def process_frame(inputs : LipSyncerInputs) -> VisionFrame:
	reference_faces = inputs.get('reference_faces')
	source_audio_frame = inputs.get('source_audio_frame')
	target_vision_frame = inputs.get('target_vision_frame')
	many_faces = sort_and_filter_faces(get_many_faces([ target_vision_frame ]))

	if state_manager.get_item('face_selector_mode') == 'many':
		if many_faces:
			for target_face in many_faces:
				target_vision_frame = sync_lip(target_face, source_audio_frame, target_vision_frame)
	if state_manager.get_item('face_selector_mode') == 'one':
		target_face = get_one_face(many_faces)
		if target_face:
			target_vision_frame = sync_lip(target_face, source_audio_frame, target_vision_frame)
	if state_manager.get_item('face_selector_mode') == 'reference':
		similar_faces = find_similar_faces(many_faces, reference_faces, state_manager.get_item('reference_face_distance'))
		if similar_faces:
			for similar_face in similar_faces:
				target_vision_frame = sync_lip(similar_face, source_audio_frame, target_vision_frame)
	return target_vision_frame


def process_frames(source_paths : List[str], queue_payloads : List[QueuePayload], update_progress : UpdateProgress) -> None:
	reference_faces = get_reference_faces() if 'reference' in state_manager.get_item('face_selector_mode') else None
	source_audio_path = get_first(filter_audio_paths(source_paths))
	temp_video_fps = restrict_video_fps(state_manager.get_item('target_path'), state_manager.get_item('output_video_fps'))

	for queue_payload in process_manager.manage(queue_payloads):
		frame_number = queue_payload.get('frame_number')
		target_vision_path = queue_payload.get('frame_path')
		source_audio_frame = get_voice_frame(source_audio_path, temp_video_fps, frame_number)
		if not numpy.any(source_audio_frame):
			source_audio_frame = create_empty_audio_frame()
		target_vision_frame = read_image(target_vision_path)
		output_vision_frame = process_frame(
		{
			'reference_faces': reference_faces,
			'source_audio_frame': source_audio_frame,
			'target_vision_frame': target_vision_frame
		})
		write_image(target_vision_path, output_vision_frame)
		update_progress(1)


def process_image(source_paths : List[str], target_path : str, output_path : str) -> None:
	reference_faces = get_reference_faces() if 'reference' in state_manager.get_item('face_selector_mode') else None
	source_audio_frame = create_empty_audio_frame()
	target_vision_frame = read_static_image(target_path)
	output_vision_frame = process_frame(
	{
		'reference_faces': reference_faces,
		'source_audio_frame': source_audio_frame,
		'target_vision_frame': target_vision_frame
	})
	write_image(output_path, output_vision_frame)


def process_video(source_paths : List[str], temp_frame_paths : List[str]) -> None:
	source_audio_paths = filter_audio_paths(state_manager.get_item('source_paths'))
	temp_video_fps = restrict_video_fps(state_manager.get_item('target_path'), state_manager.get_item('output_video_fps'))
	for source_audio_path in source_audio_paths:
		read_static_voice(source_audio_path, temp_video_fps)
	processors.multi_process_frames(source_paths, temp_frame_paths, process_frames)