init space

LICENSE.md

@@ -0,0 +1,3 @@
+MIT license
+
+Copyright (c) 2023 Henry Ruhs

app.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+
+from facefusion import core
+
+if __name__ == '__main__':
+    core.cli()

facefusion/choices.py

@@ -0,0 +1,26 @@
+from typing import List
+
+import numpy
+
+from facefusion.typing import FaceSelectorMode, FaceAnalyserOrder, FaceAnalyserAge, FaceAnalyserGender, TempFrameFormat, OutputVideoEncoder
+
+
+face_analyser_orders : List[FaceAnalyserOrder] = [ 'left-right', 'right-left', 'top-bottom', 'bottom-top', 'small-large', 'large-small', 'best-worst', 'worst-best' ]
+face_analyser_ages : List[FaceAnalyserAge] = [ 'child', 'teen', 'adult', 'senior' ]
+face_analyser_genders : List[FaceAnalyserGender] = [ 'male', 'female' ]
+face_detector_models : List[str] = [ 'retinaface', 'yunet' ]
+face_detector_sizes : List[str] = [ '160x160', '320x320', '480x480', '512x512', '640x640', '768x768', '960x960', '1024x1024' ]
+face_selector_modes : List[FaceSelectorMode] = [ 'reference', 'one', 'many' ]
+temp_frame_formats : List[TempFrameFormat] = [ 'jpg', 'png' ]
+output_video_encoders : List[OutputVideoEncoder] = [ 'libx264', 'libx265', 'libvpx-vp9', 'h264_nvenc', 'hevc_nvenc' ]
+
+execution_thread_count_range : List[int] = numpy.arange(1, 129, 1).tolist()
+execution_queue_count_range : List[int] = numpy.arange(1, 33, 1).tolist()
+max_memory_range : List[int] = numpy.arange(0, 129, 1).tolist()
+face_detector_score_range : List[float] = numpy.arange(0.0, 1.05, 0.05).tolist()
+face_mask_blur_range : List[float] = numpy.arange(0.0, 1.05, 0.05).tolist()
+face_mask_padding_range : List[float] = numpy.arange(0, 101, 1).tolist()
+reference_face_distance_range : List[float] = numpy.arange(0.0, 1.55, 0.05).tolist()
+temp_frame_quality_range : List[int] = numpy.arange(0, 101, 1).tolist()
+output_image_quality_range : List[int] = numpy.arange(0, 101, 1).tolist()
+output_video_quality_range : List[int] = numpy.arange(0, 101, 1).tolist()

facefusion/content_analyser.py

@@ -0,0 +1,102 @@
+from typing import Any, Dict
+from functools import lru_cache
+import threading
+import cv2
+import numpy
+import onnxruntime
+from tqdm import tqdm
+
+import facefusion.globals
+from facefusion import wording
+from facefusion.typing import Frame, ModelValue
+from facefusion.vision import get_video_frame, count_video_frame_total, read_image, detect_fps
+from facefusion.utilities import resolve_relative_path, conditional_download
+
+CONTENT_ANALYSER = None
+THREAD_LOCK : threading.Lock = threading.Lock()
+MODELS : Dict[str, ModelValue] =\
+{
+	'open_nsfw':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/open_nsfw.onnx',
+		'path': resolve_relative_path('../.assets/models/open_nsfw.onnx')
+	}
+}
+MAX_PROBABILITY = 0.80
+MAX_RATE = 5
+STREAM_COUNTER = 0
+
+
+def get_content_analyser() -> Any:
+	global CONTENT_ANALYSER
+
+	with THREAD_LOCK:
+		if CONTENT_ANALYSER is None:
+			model_path = MODELS.get('open_nsfw').get('path')
+			CONTENT_ANALYSER = onnxruntime.InferenceSession(model_path, providers = facefusion.globals.execution_providers)
+	return CONTENT_ANALYSER
+
+
+def clear_content_analyser() -> None:
+	global CONTENT_ANALYSER
+
+	CONTENT_ANALYSER = None
+
+
+def pre_check() -> bool:
+	if not facefusion.globals.skip_download:
+		download_directory_path = resolve_relative_path('../.assets/models')
+		model_url = MODELS.get('open_nsfw').get('url')
+		conditional_download(download_directory_path, [ model_url ])
+	return True
+
+
+def analyse_stream(frame : Frame, fps : float) -> bool:
+	global STREAM_COUNTER
+
+	STREAM_COUNTER = STREAM_COUNTER + 1
+	if STREAM_COUNTER % int(fps) == 0:
+		return analyse_frame(frame)
+	return False
+
+
+def prepare_frame(frame : Frame) -> Frame:
+	frame = cv2.resize(frame, (224, 224)).astype(numpy.float32)
+	frame -= numpy.array([ 104, 117, 123 ]).astype(numpy.float32)
+	frame = numpy.expand_dims(frame, axis = 0)
+	return frame
+
+
+def analyse_frame(frame : Frame) -> bool:
+	content_analyser = get_content_analyser()
+	frame = prepare_frame(frame)
+	probability = content_analyser.run(None,
+	{
+		'input:0': frame
+	})[0][0][1]
+	return probability > MAX_PROBABILITY
+
+
+@lru_cache(maxsize = None)
+def analyse_image(image_path : str) -> bool:
+	frame = read_image(image_path)
+	return analyse_frame(frame)
+
+
+@lru_cache(maxsize = None)
+def analyse_video(video_path : str, start_frame : int, end_frame : int) -> bool:
+	video_frame_total = count_video_frame_total(video_path)
+	fps = detect_fps(video_path)
+	frame_range = range(start_frame or 0, end_frame or video_frame_total)
+	rate = 0.0
+	counter = 0
+	with tqdm(total = len(frame_range), desc = wording.get('analysing'), unit = 'frame', ascii = ' =') as progress:
+		for frame_number in frame_range:
+			if frame_number % int(fps) == 0:
+				frame = get_video_frame(video_path, frame_number)
+				if analyse_frame(frame):
+					counter += 1
+			rate = counter * int(fps) / len(frame_range) * 100
+			progress.update()
+			progress.set_postfix(rate = rate)
+	return rate > MAX_RATE

facefusion/core.py

@@ -0,0 +1,274 @@
+import os
+
+os.environ['OMP_NUM_THREADS'] = '1'
+
+import signal
+import sys
+import warnings
+import platform
+import shutil
+import onnxruntime
+from argparse import ArgumentParser, HelpFormatter
+
+import facefusion.choices
+import facefusion.globals
+from facefusion.face_analyser import get_one_face
+from facefusion.face_reference import get_face_reference, set_face_reference
+from facefusion.vision import get_video_frame, read_image
+from facefusion import face_analyser, content_analyser, metadata, wording
+from facefusion.content_analyser import analyse_image, analyse_video
+from facefusion.processors.frame.core import get_frame_processors_modules, load_frame_processor_module
+from facefusion.utilities import is_image, is_video, detect_fps, compress_image, merge_video, extract_frames, get_temp_frame_paths, restore_audio, create_temp, move_temp, clear_temp, list_module_names, encode_execution_providers, decode_execution_providers, normalize_output_path, normalize_padding, create_metavar, update_status
+
+onnxruntime.set_default_logger_severity(3)
+warnings.filterwarnings('ignore', category = UserWarning, module = 'gradio')
+warnings.filterwarnings('ignore', category = UserWarning, module = 'torchvision')
+
+
+def cli() -> None:
+	signal.signal(signal.SIGINT, lambda signal_number, frame: destroy())
+	program = ArgumentParser(formatter_class = lambda prog: HelpFormatter(prog, max_help_position = 120), add_help = False)
+	# general
+	program.add_argument('-s', '--source', help = wording.get('source_help'), dest = 'source_path')
+	program.add_argument('-t', '--target', help = wording.get('target_help'), dest = 'target_path')
+	program.add_argument('-o', '--output', help = wording.get('output_help'), dest = 'output_path')
+	program.add_argument('-v', '--version', version = metadata.get('name') + ' ' + metadata.get('version'), action = 'version')
+	# misc
+	group_misc = program.add_argument_group('misc')
+	group_misc.add_argument('--skip-download', help = wording.get('skip_download_help'), dest = 'skip_download', action = 'store_true')
+	group_misc.add_argument('--headless', help = wording.get('headless_help'), dest = 'headless', action = 'store_true')
+	# execution
+	group_execution = program.add_argument_group('execution')
+	group_execution.add_argument('--execution-providers', help = wording.get('execution_providers_help'), dest = 'execution_providers', default = [ 'cpu' ], choices = encode_execution_providers(onnxruntime.get_available_providers()), nargs = '+')
+	group_execution.add_argument('--execution-thread-count', help = wording.get('execution_thread_count_help'), dest = 'execution_thread_count', type = int, default = 4, choices = facefusion.choices.execution_thread_count_range, metavar = create_metavar(facefusion.choices.execution_thread_count_range))
+	group_execution.add_argument('--execution-queue-count', help = wording.get('execution_queue_count_help'), dest = 'execution_queue_count', type = int, default = 1, choices = facefusion.choices.execution_queue_count_range, metavar = create_metavar(facefusion.choices.execution_queue_count_range))
+	group_execution.add_argument('--max-memory', help = wording.get('max_memory_help'), dest = 'max_memory', type = int, choices = facefusion.choices.max_memory_range, metavar = create_metavar(facefusion.choices.max_memory_range))
+	# face analyser
+	group_face_analyser = program.add_argument_group('face analyser')
+	group_face_analyser.add_argument('--face-analyser-order', help = wording.get('face_analyser_order_help'), dest = 'face_analyser_order', default = 'left-right', choices = facefusion.choices.face_analyser_orders)
+	group_face_analyser.add_argument('--face-analyser-age', help = wording.get('face_analyser_age_help'), dest = 'face_analyser_age', choices = facefusion.choices.face_analyser_ages)
+	group_face_analyser.add_argument('--face-analyser-gender', help = wording.get('face_analyser_gender_help'), dest = 'face_analyser_gender', choices = facefusion.choices.face_analyser_genders)
+	group_face_analyser.add_argument('--face-detector-model', help = wording.get('face_detector_model_help'), dest = 'face_detector_model', default = 'retinaface', choices = facefusion.choices.face_detector_models)
+	group_face_analyser.add_argument('--face-detector-size', help = wording.get('face_detector_size_help'), dest = 'face_detector_size', default = '640x640', choices = facefusion.choices.face_detector_sizes)
+	group_face_analyser.add_argument('--face-detector-score', help = wording.get('face_detector_score_help'), dest = 'face_detector_score', type = float, default = 0.5, choices = facefusion.choices.face_detector_score_range, metavar = create_metavar(facefusion.choices.face_detector_score_range))
+	# face selector
+	group_face_selector = program.add_argument_group('face selector')
+	group_face_selector.add_argument('--face-selector-mode', help = wording.get('face_selector_mode_help'), dest = 'face_selector_mode', default = 'reference', choices = facefusion.choices.face_selector_modes)
+	group_face_selector.add_argument('--reference-face-position', help = wording.get('reference_face_position_help'), dest = 'reference_face_position', type = int, default = 0)
+	group_face_selector.add_argument('--reference-face-distance', help = wording.get('reference_face_distance_help'), dest = 'reference_face_distance', type = float, default = 0.6, choices = facefusion.choices.reference_face_distance_range, metavar = create_metavar(facefusion.choices.reference_face_distance_range))
+	group_face_selector.add_argument('--reference-frame-number', help = wording.get('reference_frame_number_help'), dest = 'reference_frame_number', type = int, default = 0)
+	# face mask
+	group_face_mask = program.add_argument_group('face mask')
+	group_face_mask.add_argument('--face-mask-blur', help = wording.get('face_mask_blur_help'), dest = 'face_mask_blur', type = float, default = 0.3, choices = facefusion.choices.face_mask_blur_range, metavar = create_metavar(facefusion.choices.face_mask_blur_range))
+	group_face_mask.add_argument('--face-mask-padding', help = wording.get('face_mask_padding_help'), dest = 'face_mask_padding', type = int, default = [ 0, 0, 0, 0 ], nargs = '+')
+	# frame extraction
+	group_frame_extraction = program.add_argument_group('frame extraction')
+	group_frame_extraction.add_argument('--trim-frame-start', help = wording.get('trim_frame_start_help'), dest = 'trim_frame_start', type = int)
+	group_frame_extraction.add_argument('--trim-frame-end', help = wording.get('trim_frame_end_help'), dest = 'trim_frame_end', type = int)
+	group_frame_extraction.add_argument('--temp-frame-format', help = wording.get('temp_frame_format_help'), dest = 'temp_frame_format', default = 'jpg', choices = facefusion.choices.temp_frame_formats)
+	group_frame_extraction.add_argument('--temp-frame-quality', help = wording.get('temp_frame_quality_help'), dest = 'temp_frame_quality', type = int, default = 100, choices = facefusion.choices.temp_frame_quality_range, metavar = create_metavar(facefusion.choices.temp_frame_quality_range))
+	group_frame_extraction.add_argument('--keep-temp', help = wording.get('keep_temp_help'), dest = 'keep_temp', action = 'store_true')
+	# output creation
+	group_output_creation = program.add_argument_group('output creation')
+	group_output_creation.add_argument('--output-image-quality', help = wording.get('output_image_quality_help'), dest = 'output_image_quality', type = int, default = 80, choices = facefusion.choices.output_image_quality_range, metavar = create_metavar(facefusion.choices.output_image_quality_range))
+	group_output_creation.add_argument('--output-video-encoder', help = wording.get('output_video_encoder_help'), dest = 'output_video_encoder', default = 'libx264', choices = facefusion.choices.output_video_encoders)
+	group_output_creation.add_argument('--output-video-quality', help = wording.get('output_video_quality_help'), dest = 'output_video_quality', type = int, default = 80, choices = facefusion.choices.output_video_quality_range, metavar = create_metavar(facefusion.choices.output_video_quality_range))
+	group_output_creation.add_argument('--keep-fps', help = wording.get('keep_fps_help'), dest = 'keep_fps', action = 'store_true')
+	group_output_creation.add_argument('--skip-audio', help = wording.get('skip_audio_help'), dest = 'skip_audio', action = 'store_true')
+	# frame processors
+	available_frame_processors = list_module_names('facefusion/processors/frame/modules')
+	program = ArgumentParser(parents = [ program ], formatter_class = program.formatter_class, add_help = True)
+	group_frame_processors = program.add_argument_group('frame processors')
+	group_frame_processors.add_argument('--frame-processors', help = wording.get('frame_processors_help').format(choices = ', '.join(available_frame_processors)), dest = 'frame_processors', default = [ 'face_swapper' ], nargs = '+')
+	for frame_processor in available_frame_processors:
+		frame_processor_module = load_frame_processor_module(frame_processor)
+		frame_processor_module.register_args(group_frame_processors)
+	# uis
+	group_uis = program.add_argument_group('uis')
+	group_uis.add_argument('--ui-layouts', help = wording.get('ui_layouts_help').format(choices = ', '.join(list_module_names('facefusion/uis/layouts'))), dest = 'ui_layouts', default = [ 'default' ], nargs = '+')
+	run(program)
+
+
+def apply_args(program : ArgumentParser) -> None:
+	args = program.parse_args()
+	# general
+	facefusion.globals.source_path = args.source_path
+	facefusion.globals.target_path = args.target_path
+	facefusion.globals.output_path = normalize_output_path(facefusion.globals.source_path, facefusion.globals.target_path, args.output_path)
+	# misc
+	facefusion.globals.skip_download = args.skip_download
+	facefusion.globals.headless = args.headless
+	# execution
+	facefusion.globals.execution_providers = decode_execution_providers(args.execution_providers)
+	facefusion.globals.execution_thread_count = args.execution_thread_count
+	facefusion.globals.execution_queue_count = args.execution_queue_count
+	facefusion.globals.max_memory = args.max_memory
+	# face analyser
+	facefusion.globals.face_analyser_order = args.face_analyser_order
+	facefusion.globals.face_analyser_age = args.face_analyser_age
+	facefusion.globals.face_analyser_gender = args.face_analyser_gender
+	facefusion.globals.face_detector_model = args.face_detector_model
+	facefusion.globals.face_detector_size = args.face_detector_size
+	facefusion.globals.face_detector_score = args.face_detector_score
+	# face selector
+	facefusion.globals.face_selector_mode = args.face_selector_mode
+	facefusion.globals.reference_face_position = args.reference_face_position
+	facefusion.globals.reference_face_distance = args.reference_face_distance
+	facefusion.globals.reference_frame_number = args.reference_frame_number
+	# face mask
+	facefusion.globals.face_mask_blur = args.face_mask_blur
+	facefusion.globals.face_mask_padding = normalize_padding(args.face_mask_padding)
+	# frame extraction
+	facefusion.globals.trim_frame_start = args.trim_frame_start
+	facefusion.globals.trim_frame_end = args.trim_frame_end
+	facefusion.globals.temp_frame_format = args.temp_frame_format
+	facefusion.globals.temp_frame_quality = args.temp_frame_quality
+	facefusion.globals.keep_temp = args.keep_temp
+	# output creation
+	facefusion.globals.output_image_quality = args.output_image_quality
+	facefusion.globals.output_video_encoder = args.output_video_encoder
+	facefusion.globals.output_video_quality = args.output_video_quality
+	facefusion.globals.keep_fps = args.keep_fps
+	facefusion.globals.skip_audio = args.skip_audio
+	# frame processors
+	available_frame_processors = list_module_names('facefusion/processors/frame/modules')
+	facefusion.globals.frame_processors = args.frame_processors
+	for frame_processor in available_frame_processors:
+		frame_processor_module = load_frame_processor_module(frame_processor)
+		frame_processor_module.apply_args(program)
+	# uis
+	facefusion.globals.ui_layouts = args.ui_layouts
+
+
+def run(program : ArgumentParser) -> None:
+	apply_args(program)
+	limit_resources()
+	if not pre_check() or not content_analyser.pre_check() or not face_analyser.pre_check():
+		return
+	for frame_processor_module in get_frame_processors_modules(facefusion.globals.frame_processors):
+		if not frame_processor_module.pre_check():
+			return
+	if facefusion.globals.headless:
+		conditional_process()
+	else:
+		import facefusion.uis.core as ui
+
+		for ui_layout in ui.get_ui_layouts_modules(facefusion.globals.ui_layouts):
+			if not ui_layout.pre_check():
+				return
+		ui.launch()
+
+
+def destroy() -> None:
+	if facefusion.globals.target_path:
+		clear_temp(facefusion.globals.target_path)
+	sys.exit()
+
+
+def limit_resources() -> None:
+	if facefusion.globals.max_memory:
+		memory = facefusion.globals.max_memory * 1024 ** 3
+		if platform.system().lower() == 'darwin':
+			memory = facefusion.globals.max_memory * 1024 ** 6
+		if platform.system().lower() == 'windows':
+			import ctypes
+			kernel32 = ctypes.windll.kernel32 # type: ignore[attr-defined]
+			kernel32.SetProcessWorkingSetSize(-1, ctypes.c_size_t(memory), ctypes.c_size_t(memory))
+		else:
+			import resource
+			resource.setrlimit(resource.RLIMIT_DATA, (memory, memory))
+
+
+def pre_check() -> bool:
+	if sys.version_info < (3, 9):
+		update_status(wording.get('python_not_supported').format(version = '3.9'))
+		return False
+	if not shutil.which('ffmpeg'):
+		update_status(wording.get('ffmpeg_not_installed'))
+		return False
+	return True
+
+
+def conditional_process() -> None:
+	conditional_set_face_reference()
+	for frame_processor_module in get_frame_processors_modules(facefusion.globals.frame_processors):
+		if not frame_processor_module.pre_process('output'):
+			return
+	if is_image(facefusion.globals.target_path):
+		process_image()
+	if is_video(facefusion.globals.target_path):
+		process_video()
+
+
+def conditional_set_face_reference() -> None:
+	if 'reference' in facefusion.globals.face_selector_mode and not get_face_reference():
+		if is_video(facefusion.globals.target_path):
+			reference_frame = get_video_frame(facefusion.globals.target_path, facefusion.globals.reference_frame_number)
+		else:
+			reference_frame = read_image(facefusion.globals.target_path)
+		reference_face = get_one_face(reference_frame, facefusion.globals.reference_face_position)
+		set_face_reference(reference_face)
+
+
+def process_image() -> None:
+	if analyse_image(facefusion.globals.target_path):
+		return
+	shutil.copy2(facefusion.globals.target_path, facefusion.globals.output_path)
+	# process frame
+	for frame_processor_module in get_frame_processors_modules(facefusion.globals.frame_processors):
+		update_status(wording.get('processing'), frame_processor_module.NAME)
+		frame_processor_module.process_image(facefusion.globals.source_path, facefusion.globals.output_path, facefusion.globals.output_path)
+		frame_processor_module.post_process()
+	# compress image
+	update_status(wording.get('compressing_image'))
+	if not compress_image(facefusion.globals.output_path):
+		update_status(wording.get('compressing_image_failed'))
+	# validate image
+	if is_image(facefusion.globals.output_path):
+		update_status(wording.get('processing_image_succeed'))
+	else:
+		update_status(wording.get('processing_image_failed'))
+
+
+def process_video() -> None:
+	if analyse_video(facefusion.globals.target_path, facefusion.globals.trim_frame_start, facefusion.globals.trim_frame_end):
+		return
+	fps = detect_fps(facefusion.globals.target_path) if facefusion.globals.keep_fps else 25.0
+	# create temp
+	update_status(wording.get('creating_temp'))
+	create_temp(facefusion.globals.target_path)
+	# extract frames
+	update_status(wording.get('extracting_frames_fps').format(fps = fps))
+	extract_frames(facefusion.globals.target_path, fps)
+	# process frame
+	temp_frame_paths = get_temp_frame_paths(facefusion.globals.target_path)
+	if temp_frame_paths:
+		for frame_processor_module in get_frame_processors_modules(facefusion.globals.frame_processors):
+			update_status(wording.get('processing'), frame_processor_module.NAME)
+			frame_processor_module.process_video(facefusion.globals.source_path, temp_frame_paths)
+			frame_processor_module.post_process()
+	else:
+		update_status(wording.get('temp_frames_not_found'))
+		return
+	# merge video
+	update_status(wording.get('merging_video_fps').format(fps = fps))
+	if not merge_video(facefusion.globals.target_path, fps):
+		update_status(wording.get('merging_video_failed'))
+		return
+	# handle audio
+	if facefusion.globals.skip_audio:
+		update_status(wording.get('skipping_audio'))
+		move_temp(facefusion.globals.target_path, facefusion.globals.output_path)
+	else:
+		update_status(wording.get('restoring_audio'))
+		if not restore_audio(facefusion.globals.target_path, facefusion.globals.output_path):
+			update_status(wording.get('restoring_audio_failed'))
+			move_temp(facefusion.globals.target_path, facefusion.globals.output_path)
+	# clear temp
+	update_status(wording.get('clearing_temp'))
+	clear_temp(facefusion.globals.target_path)
+	# validate video
+	if is_video(facefusion.globals.output_path):
+		update_status(wording.get('processing_video_succeed'))
+	else:
+		update_status(wording.get('processing_video_failed'))

facefusion/face_analyser.py

@@ -0,0 +1,309 @@
+from typing import Any, Optional, List, Dict, Tuple
+import threading
+import cv2
+import numpy
+import onnxruntime
+
+import facefusion.globals
+from facefusion.face_cache import get_faces_cache, set_faces_cache
+from facefusion.face_helper import warp_face, create_static_anchors, distance_to_kps, distance_to_bbox, apply_nms
+from facefusion.typing import Frame, Face, FaceAnalyserOrder, FaceAnalyserAge, FaceAnalyserGender, ModelValue, Bbox, Kps, Score, Embedding
+from facefusion.utilities import resolve_relative_path, conditional_download
+from facefusion.vision import resize_frame_dimension
+
+FACE_ANALYSER = None
+THREAD_SEMAPHORE : threading.Semaphore = threading.Semaphore()
+THREAD_LOCK : threading.Lock = threading.Lock()
+MODELS : Dict[str, ModelValue] =\
+{
+	'face_detector_retinaface':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/retinaface_10g.onnx',
+		'path': resolve_relative_path('../.assets/models/retinaface_10g.onnx')
+	},
+	'face_detector_yunet':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/yunet_2023mar.onnx',
+		'path': resolve_relative_path('../.assets/models/yunet_2023mar.onnx')
+	},
+	'face_recognizer_arcface_blendface':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/arcface_w600k_r50.onnx',
+		'path': resolve_relative_path('../.assets/models/arcface_w600k_r50.onnx')
+	},
+	'face_recognizer_arcface_inswapper':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/arcface_w600k_r50.onnx',
+		'path': resolve_relative_path('../.assets/models/arcface_w600k_r50.onnx')
+	},
+	'face_recognizer_arcface_simswap':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/arcface_simswap.onnx',
+		'path': resolve_relative_path('../.assets/models/arcface_simswap.onnx')
+	},
+	'gender_age':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gender_age.onnx',
+		'path': resolve_relative_path('../.assets/models/gender_age.onnx')
+	}
+}
+
+
+def get_face_analyser() -> Any:
+	global FACE_ANALYSER
+
+	with THREAD_LOCK:
+		if FACE_ANALYSER is None:
+			if facefusion.globals.face_detector_model == 'retinaface':
+				face_detector = onnxruntime.InferenceSession(MODELS.get('face_detector_retinaface').get('path'), providers = facefusion.globals.execution_providers)
+			if facefusion.globals.face_detector_model == 'yunet':
+				face_detector = cv2.FaceDetectorYN.create(MODELS.get('face_detector_yunet').get('path'), '', (0, 0))
+			if facefusion.globals.face_recognizer_model == 'arcface_blendface':
+				face_recognizer = onnxruntime.InferenceSession(MODELS.get('face_recognizer_arcface_blendface').get('path'), providers = facefusion.globals.execution_providers)
+			if facefusion.globals.face_recognizer_model == 'arcface_inswapper':
+				face_recognizer = onnxruntime.InferenceSession(MODELS.get('face_recognizer_arcface_inswapper').get('path'), providers = facefusion.globals.execution_providers)
+			if facefusion.globals.face_recognizer_model == 'arcface_simswap':
+				face_recognizer = onnxruntime.InferenceSession(MODELS.get('face_recognizer_arcface_simswap').get('path'), providers = facefusion.globals.execution_providers)
+			gender_age = onnxruntime.InferenceSession(MODELS.get('gender_age').get('path'), providers = facefusion.globals.execution_providers)
+			FACE_ANALYSER =\
+			{
+				'face_detector': face_detector,
+				'face_recognizer': face_recognizer,
+				'gender_age': gender_age
+			}
+	return FACE_ANALYSER
+
+
+def clear_face_analyser() -> Any:
+	global FACE_ANALYSER
+
+	FACE_ANALYSER = None
+
+
+def pre_check() -> bool:
+	if not facefusion.globals.skip_download:
+		download_directory_path = resolve_relative_path('../.assets/models')
+		model_urls =\
+		[
+			MODELS.get('face_detector_retinaface').get('url'),
+			MODELS.get('face_detector_yunet').get('url'),
+			MODELS.get('face_recognizer_arcface_inswapper').get('url'),
+			MODELS.get('face_recognizer_arcface_simswap').get('url'),
+			MODELS.get('gender_age').get('url')
+		]
+		conditional_download(download_directory_path, model_urls)
+	return True
+
+
+def extract_faces(frame: Frame) -> List[Face]:
+	face_detector_width, face_detector_height = map(int, facefusion.globals.face_detector_size.split('x'))
+	frame_height, frame_width, _ = frame.shape
+	temp_frame = resize_frame_dimension(frame, face_detector_width, face_detector_height)
+	temp_frame_height, temp_frame_width, _ = temp_frame.shape
+	ratio_height = frame_height / temp_frame_height
+	ratio_width = frame_width / temp_frame_width
+	if facefusion.globals.face_detector_model == 'retinaface':
+		bbox_list, kps_list, score_list = detect_with_retinaface(temp_frame, temp_frame_height, temp_frame_width, face_detector_height, face_detector_width, ratio_height, ratio_width)
+		return create_faces(frame, bbox_list, kps_list, score_list)
+	elif facefusion.globals.face_detector_model == 'yunet':
+		bbox_list, kps_list, score_list = detect_with_yunet(temp_frame, temp_frame_height, temp_frame_width, ratio_height, ratio_width)
+		return create_faces(frame, bbox_list, kps_list, score_list)
+	return []
+
+
+def detect_with_retinaface(temp_frame : Frame, temp_frame_height : int, temp_frame_width : int, face_detector_height : int, face_detector_width : int, ratio_height : float, ratio_width : float) -> Tuple[List[Bbox], List[Kps], List[Score]]:
+	face_detector = get_face_analyser().get('face_detector')
+	bbox_list = []
+	kps_list = []
+	score_list = []
+	feature_strides = [ 8, 16, 32 ]
+	feature_map_channel = 3
+	anchor_total = 2
+	prepare_frame = numpy.zeros((face_detector_height, face_detector_width, 3))
+	prepare_frame[:temp_frame_height, :temp_frame_width, :] = temp_frame
+	temp_frame = (prepare_frame - 127.5) / 128.0
+	temp_frame = numpy.expand_dims(temp_frame.transpose(2, 0, 1), axis = 0).astype(numpy.float32)
+	with THREAD_SEMAPHORE:
+		detections = face_detector.run(None,
+		{
+			face_detector.get_inputs()[0].name: temp_frame
+		})
+	for index, feature_stride in enumerate(feature_strides):
+		keep_indices = numpy.where(detections[index] >= facefusion.globals.face_detector_score)[0]
+		if keep_indices.any():
+			stride_height = face_detector_height // feature_stride
+			stride_width = face_detector_width // feature_stride
+			anchors = create_static_anchors(feature_stride, anchor_total, stride_height, stride_width)
+			bbox_raw = (detections[index + feature_map_channel] * feature_stride)
+			kps_raw = detections[index + feature_map_channel * 2] * feature_stride
+			for bbox in distance_to_bbox(anchors, bbox_raw)[keep_indices]:
+				bbox_list.append(numpy.array(
+				[
+					bbox[0] * ratio_width,
+					bbox[1] * ratio_height,
+					bbox[2] * ratio_width,
+					bbox[3] * ratio_height
+				]))
+			for kps in distance_to_kps(anchors, kps_raw)[keep_indices]:
+				kps_list.append(kps * [ ratio_width, ratio_height ])
+			for score in detections[index][keep_indices]:
+				score_list.append(score[0])
+	return bbox_list, kps_list, score_list
+
+
+def detect_with_yunet(temp_frame : Frame, temp_frame_height : int, temp_frame_width : int, ratio_height : float, ratio_width : float) -> Tuple[List[Bbox], List[Kps], List[Score]]:
+	face_detector = get_face_analyser().get('face_detector')
+	face_detector.setInputSize((temp_frame_width, temp_frame_height))
+	face_detector.setScoreThreshold(facefusion.globals.face_detector_score)
+	bbox_list = []
+	kps_list = []
+	score_list = []
+	with THREAD_SEMAPHORE:
+		_, detections = face_detector.detect(temp_frame)
+	if detections.any():
+		for detection in detections:
+			bbox_list.append(numpy.array(
+			[
+				detection[0] * ratio_width,
+				detection[1] * ratio_height,
+				(detection[0] + detection[2]) * ratio_width,
+				(detection[1] + detection[3]) * ratio_height
+			]))
+			kps_list.append(detection[4:14].reshape((5, 2)) * [ ratio_width, ratio_height])
+			score_list.append(detection[14])
+	return bbox_list, kps_list, score_list
+
+
+def create_faces(frame : Frame, bbox_list : List[Bbox], kps_list : List[Kps], score_list : List[Score]) -> List[Face] :
+	faces : List[Face] = []
+	if facefusion.globals.face_detector_score > 0:
+		keep_indices = apply_nms(bbox_list, 0.4)
+		for index in keep_indices:
+			bbox = bbox_list[index]
+			kps = kps_list[index]
+			score = score_list[index]
+			embedding, normed_embedding = calc_embedding(frame, kps)
+			gender, age = detect_gender_age(frame, kps)
+			faces.append(Face(
+				bbox = bbox,
+				kps = kps,
+				score = score,
+				embedding = embedding,
+				normed_embedding = normed_embedding,
+				gender = gender,
+				age = age
+			))
+	return faces
+
+
+def calc_embedding(temp_frame : Frame, kps : Kps) -> Tuple[Embedding, Embedding]:
+	face_recognizer = get_face_analyser().get('face_recognizer')
+	crop_frame, matrix = warp_face(temp_frame, kps, 'arcface_v2', (112, 112))
+	crop_frame = crop_frame.astype(numpy.float32) / 127.5 - 1
+	crop_frame = crop_frame[:, :, ::-1].transpose(2, 0, 1)
+	crop_frame = numpy.expand_dims(crop_frame, axis = 0)
+	embedding = face_recognizer.run(None,
+	{
+		face_recognizer.get_inputs()[0].name: crop_frame
+	})[0]
+	embedding = embedding.ravel()
+	normed_embedding = embedding / numpy.linalg.norm(embedding)
+	return embedding, normed_embedding
+
+
+def detect_gender_age(frame : Frame, kps : Kps) -> Tuple[int, int]:
+	gender_age = get_face_analyser().get('gender_age')
+	crop_frame, affine_matrix = warp_face(frame, kps, 'arcface_v2', (96, 96))
+	crop_frame = numpy.expand_dims(crop_frame, axis = 0).transpose(0, 3, 1, 2).astype(numpy.float32)
+	prediction = gender_age.run(None,
+	{
+		gender_age.get_inputs()[0].name: crop_frame
+	})[0][0]
+	gender = int(numpy.argmax(prediction[:2]))
+	age = int(numpy.round(prediction[2] * 100))
+	return gender, age
+
+
+def get_one_face(frame : Frame, position : int = 0) -> Optional[Face]:
+	many_faces = get_many_faces(frame)
+	if many_faces:
+		try:
+			return many_faces[position]
+		except IndexError:
+			return many_faces[-1]
+	return None
+
+
+def get_many_faces(frame : Frame) -> List[Face]:
+	try:
+		faces_cache = get_faces_cache(frame)
+		if faces_cache:
+			faces = faces_cache
+		else:
+			faces = extract_faces(frame)
+			set_faces_cache(frame, faces)
+		if facefusion.globals.face_analyser_order:
+			faces = sort_by_order(faces, facefusion.globals.face_analyser_order)
+		if facefusion.globals.face_analyser_age:
+			faces = filter_by_age(faces, facefusion.globals.face_analyser_age)
+		if facefusion.globals.face_analyser_gender:
+			faces = filter_by_gender(faces, facefusion.globals.face_analyser_gender)
+		return faces
+	except (AttributeError, ValueError):
+		return []
+
+
+def find_similar_faces(frame : Frame, reference_face : Face, face_distance : float) -> List[Face]:
+	many_faces = get_many_faces(frame)
+	similar_faces = []
+	if many_faces:
+		for face in many_faces:
+			if hasattr(face, 'normed_embedding') and hasattr(reference_face, 'normed_embedding'):
+				current_face_distance = 1 - numpy.dot(face.normed_embedding, reference_face.normed_embedding)
+				if current_face_distance < face_distance:
+					similar_faces.append(face)
+	return similar_faces
+
+
+def sort_by_order(faces : List[Face], order : FaceAnalyserOrder) -> List[Face]:
+	if order == 'left-right':
+		return sorted(faces, key = lambda face: face.bbox[0])
+	if order == 'right-left':
+		return sorted(faces, key = lambda face: face.bbox[0], reverse = True)
+	if order == 'top-bottom':
+		return sorted(faces, key = lambda face: face.bbox[1])
+	if order == 'bottom-top':
+		return sorted(faces, key = lambda face: face.bbox[1], reverse = True)
+	if order == 'small-large':
+		return sorted(faces, key = lambda face: (face.bbox[2] - face.bbox[0]) * (face.bbox[3] - face.bbox[1]))
+	if order == 'large-small':
+		return sorted(faces, key = lambda face: (face.bbox[2] - face.bbox[0]) * (face.bbox[3] - face.bbox[1]), reverse = True)
+	if order == 'best-worst':
+		return sorted(faces, key = lambda face: face.score, reverse = True)
+	if order == 'worst-best':
+		return sorted(faces, key = lambda face: face.score)
+	return faces
+
+
+def filter_by_age(faces : List[Face], age : FaceAnalyserAge) -> List[Face]:
+	filter_faces = []
+	for face in faces:
+		if face.age < 13 and age == 'child':
+			filter_faces.append(face)
+		elif face.age < 19 and age == 'teen':
+			filter_faces.append(face)
+		elif face.age < 60 and age == 'adult':
+			filter_faces.append(face)
+		elif face.age > 59 and age == 'senior':
+			filter_faces.append(face)
+	return filter_faces
+
+
+def filter_by_gender(faces : List[Face], gender : FaceAnalyserGender) -> List[Face]:
+	filter_faces = []
+	for face in faces:
+		if face.gender == 0 and gender == 'female':
+			filter_faces.append(face)
+		if face.gender == 1 and gender == 'male':
+			filter_faces.append(face)
+	return filter_faces

facefusion/face_cache.py

@@ -0,0 +1,29 @@
+from typing import Optional, List, Dict
+import hashlib
+
+from facefusion.typing import Frame, Face
+
+FACES_CACHE : Dict[str, List[Face]] = {}
+
+
+def get_faces_cache(frame : Frame) -> Optional[List[Face]]:
+	frame_hash = create_frame_hash(frame)
+	if frame_hash in FACES_CACHE:
+		return FACES_CACHE[frame_hash]
+	return None
+
+
+def set_faces_cache(frame : Frame, faces : List[Face]) -> None:
+	frame_hash = create_frame_hash(frame)
+	if frame_hash:
+		FACES_CACHE[frame_hash] = faces
+
+
+def clear_faces_cache() -> None:
+	global FACES_CACHE
+
+	FACES_CACHE = {}
+
+
+def create_frame_hash(frame : Frame) -> Optional[str]:
+	return hashlib.sha1(frame.tobytes()).hexdigest() if frame.any() else None

facefusion/face_helper.py

@@ -0,0 +1,119 @@
+from typing import Any, Dict, Tuple, List
+from functools import lru_cache
+from cv2.typing import Size
+import cv2
+import numpy
+
+from facefusion.typing import Bbox, Kps, Frame, Matrix, Template, Padding
+
+TEMPLATES : Dict[Template, numpy.ndarray[Any, Any]] =\
+{
+	'arcface_v1': numpy.array(
+	[
+		[ 39.7300, 51.1380 ],
+		[ 72.2700, 51.1380 ],
+		[ 56.0000, 68.4930 ],
+		[ 42.4630, 87.0100 ],
+		[ 69.5370, 87.0100 ]
+	]),
+	'arcface_v2': numpy.array(
+	[
+		[ 38.2946, 51.6963 ],
+		[ 73.5318, 51.5014 ],
+		[ 56.0252, 71.7366 ],
+		[ 41.5493, 92.3655 ],
+		[ 70.7299, 92.2041 ]
+	]),
+	'ffhq': numpy.array(
+	[
+		[ 192.98138, 239.94708 ],
+		[ 318.90277, 240.1936 ],
+		[ 256.63416, 314.01935 ],
+		[ 201.26117, 371.41043 ],
+		[ 313.08905, 371.15118 ]
+	])
+}
+
+
+def warp_face(temp_frame : Frame, kps : Kps, template : Template, size : Size) -> Tuple[Frame, Matrix]:
+	normed_template = TEMPLATES.get(template) * size[1] / size[0]
+	affine_matrix = cv2.estimateAffinePartial2D(kps, normed_template, method = cv2.LMEDS)[0]
+	crop_frame = cv2.warpAffine(temp_frame, affine_matrix, (size[1], size[1]), borderMode = cv2.BORDER_REPLICATE)
+	return crop_frame, affine_matrix
+
+
+def paste_back(temp_frame : Frame, crop_frame: Frame, affine_matrix : Matrix, face_mask_blur : float, face_mask_padding : Padding) -> Frame:
+	inverse_matrix = cv2.invertAffineTransform(affine_matrix)
+	temp_frame_size = temp_frame.shape[:2][::-1]
+	mask_size = tuple(crop_frame.shape[:2])
+	mask_frame = create_static_mask_frame(mask_size, face_mask_blur, face_mask_padding)
+	inverse_mask_frame = cv2.warpAffine(mask_frame, inverse_matrix, temp_frame_size).clip(0, 1)
+	inverse_crop_frame = cv2.warpAffine(crop_frame, inverse_matrix, temp_frame_size, borderMode = cv2.BORDER_REPLICATE)
+	paste_frame = temp_frame.copy()
+	paste_frame[:, :, 0] = inverse_mask_frame * inverse_crop_frame[:, :, 0] + (1 - inverse_mask_frame) * temp_frame[:, :, 0]
+	paste_frame[:, :, 1] = inverse_mask_frame * inverse_crop_frame[:, :, 1] + (1 - inverse_mask_frame) * temp_frame[:, :, 1]
+	paste_frame[:, :, 2] = inverse_mask_frame * inverse_crop_frame[:, :, 2] + (1 - inverse_mask_frame) * temp_frame[:, :, 2]
+	return paste_frame
+
+
+@lru_cache(maxsize = None)
+def create_static_mask_frame(mask_size : Size, face_mask_blur : float, face_mask_padding : Padding) -> Frame:
+	mask_frame = numpy.ones(mask_size, numpy.float32)
+	blur_amount = int(mask_size[0] * 0.5 * face_mask_blur)
+	blur_area = max(blur_amount // 2, 1)
+	mask_frame[:max(blur_area, int(mask_size[1] * face_mask_padding[0] / 100)), :] = 0
+	mask_frame[-max(blur_area, int(mask_size[1] * face_mask_padding[2] / 100)):, :] = 0
+	mask_frame[:, :max(blur_area, int(mask_size[0] * face_mask_padding[3] / 100))] = 0
+	mask_frame[:, -max(blur_area, int(mask_size[0] * face_mask_padding[1] / 100)):] = 0
+	if blur_amount > 0:
+		mask_frame = cv2.GaussianBlur(mask_frame, (0, 0), blur_amount * 0.25)
+	return mask_frame
+
+
+@lru_cache(maxsize = None)
+def create_static_anchors(feature_stride : int, anchor_total : int, stride_height : int, stride_width : int) -> numpy.ndarray[Any, Any]:
+	y, x = numpy.mgrid[:stride_height, :stride_width][::-1]
+	anchors = numpy.stack((y, x), axis = -1)
+	anchors = (anchors * feature_stride).reshape((-1, 2))
+	anchors = numpy.stack([ anchors ] * anchor_total, axis = 1).reshape((-1, 2))
+	return anchors
+
+
+def distance_to_bbox(points : numpy.ndarray[Any, Any], distance : numpy.ndarray[Any, Any]) -> Bbox:
+	x1 = points[:, 0] - distance[:, 0]
+	y1 = points[:, 1] - distance[:, 1]
+	x2 = points[:, 0] + distance[:, 2]
+	y2 = points[:, 1] + distance[:, 3]
+	bbox = numpy.column_stack([ x1, y1, x2, y2 ])
+	return bbox
+
+
+def distance_to_kps(points : numpy.ndarray[Any, Any], distance : numpy.ndarray[Any, Any]) -> Kps:
+	x = points[:, 0::2] + distance[:, 0::2]
+	y = points[:, 1::2] + distance[:, 1::2]
+	kps = numpy.stack((x, y), axis = -1)
+	return kps
+
+
+def apply_nms(bbox_list : List[Bbox], iou_threshold : float) -> List[int]:
+	keep_indices = []
+	dimension_list = numpy.reshape(bbox_list, (-1, 4))
+	x1 = dimension_list[:, 0]
+	y1 = dimension_list[:, 1]
+	x2 = dimension_list[:, 2]
+	y2 = dimension_list[:, 3]
+	areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+	indices = numpy.arange(len(bbox_list))
+	while indices.size > 0:
+		index = indices[0]
+		remain_indices = indices[1:]
+		keep_indices.append(index)
+		xx1 = numpy.maximum(x1[index], x1[remain_indices])
+		yy1 = numpy.maximum(y1[index], y1[remain_indices])
+		xx2 = numpy.minimum(x2[index], x2[remain_indices])
+		yy2 = numpy.minimum(y2[index], y2[remain_indices])
+		width = numpy.maximum(0, xx2 - xx1 + 1)
+		height = numpy.maximum(0, yy2 - yy1 + 1)
+		iou = width * height / (areas[index] + areas[remain_indices] - width * height)
+		indices = indices[numpy.where(iou <= iou_threshold)[0] + 1]
+	return keep_indices

facefusion/face_reference.py

@@ -0,0 +1,21 @@
+from typing import Optional
+
+from facefusion.typing import Face
+
+FACE_REFERENCE = None
+
+
+def get_face_reference() -> Optional[Face]:
+	return FACE_REFERENCE
+
+
+def set_face_reference(face : Face) -> None:
+	global FACE_REFERENCE
+
+	FACE_REFERENCE = face
+
+
+def clear_face_reference() -> None:
+	global FACE_REFERENCE
+
+	FACE_REFERENCE = None

facefusion/globals.py

@@ -0,0 +1,48 @@
+from typing import List, Optional
+
+from facefusion.typing import FaceSelectorMode, FaceAnalyserOrder, FaceAnalyserAge, FaceAnalyserGender, OutputVideoEncoder, FaceDetectorModel, FaceRecognizerModel, TempFrameFormat, Padding
+
+# general
+source_path : Optional[str] = None
+target_path : Optional[str] = None
+output_path : Optional[str] = None
+# misc
+skip_download : Optional[bool] = None
+headless : Optional[bool] = None
+# execution
+execution_providers : List[str] = []
+execution_thread_count : Optional[int] = None
+execution_queue_count : Optional[int] = None
+max_memory : Optional[int] = None
+# face analyser
+face_analyser_order : Optional[FaceAnalyserOrder] = None
+face_analyser_age : Optional[FaceAnalyserAge] = None
+face_analyser_gender : Optional[FaceAnalyserGender] = None
+face_detector_model : Optional[FaceDetectorModel] = None
+face_detector_size : Optional[str] = None
+face_detector_score : Optional[float] = None
+face_recognizer_model : Optional[FaceRecognizerModel] = None
+# face selector
+face_selector_mode : Optional[FaceSelectorMode] = None
+reference_face_position : Optional[int] = None
+reference_face_distance : Optional[float] = None
+reference_frame_number : Optional[int] = None
+# face mask
+face_mask_blur : Optional[float] = None
+face_mask_padding : Optional[Padding] = None
+# frame extraction
+trim_frame_start : Optional[int] = None
+trim_frame_end : Optional[int] = None
+temp_frame_format : Optional[TempFrameFormat] = None
+temp_frame_quality : Optional[int] = None
+keep_temp : Optional[bool] = None
+# output creation
+output_image_quality : Optional[int] = None
+output_video_encoder : Optional[OutputVideoEncoder] = None
+output_video_quality : Optional[int] = None
+keep_fps : Optional[bool] = None
+skip_audio : Optional[bool] = None
+# frame processors
+frame_processors : List[str] = []
+# uis
+ui_layouts : List[str] = []

facefusion/installer.py

@@ -0,0 +1,63 @@
+from typing import Dict, Tuple
+import subprocess
+from argparse import ArgumentParser, HelpFormatter
+
+subprocess.call([ 'pip', 'install' , 'inquirer', '-q' ])
+
+import inquirer
+
+from facefusion import metadata, wording
+
+TORCH : Dict[str, str] =\
+{
+	'default': 'default',
+	'cpu': 'cpu',
+	'cuda': 'cu118',
+	'rocm': 'rocm5.6'
+}
+ONNXRUNTIMES : Dict[str, Tuple[str, str]] =\
+{
+	'default': ('onnxruntime', '1.16.3'),
+	'cuda': ('onnxruntime-gpu', '1.16.3'),
+	'coreml-legacy': ('onnxruntime-coreml', '1.13.1'),
+	'coreml-silicon': ('onnxruntime-silicon', '1.16.0'),
+	'directml': ('onnxruntime-directml', '1.16.3'),
+	'openvino': ('onnxruntime-openvino', '1.16.0')
+}
+
+
+def cli() -> None:
+	program = ArgumentParser(formatter_class = lambda prog: HelpFormatter(prog, max_help_position = 120))
+	program.add_argument('--torch', help = wording.get('install_dependency_help').format(dependency = 'torch'), dest = 'torch', choices = TORCH.keys())
+	program.add_argument('--onnxruntime', help = wording.get('install_dependency_help').format(dependency = 'onnxruntime'), dest = 'onnxruntime', choices = ONNXRUNTIMES.keys())
+	program.add_argument('-v', '--version', version = metadata.get('name') + ' ' + metadata.get('version'), action = 'version')
+	run(program)
+
+
+def run(program : ArgumentParser) -> None:
+	args = program.parse_args()
+
+	if args.torch and args.onnxruntime:
+		answers =\
+		{
+			'torch': args.torch,
+			'onnxruntime': args.onnxruntime
+		}
+	else:
+		answers = inquirer.prompt(
+		[
+			inquirer.List('torch', message = wording.get('install_dependency_help').format(dependency = 'torch'), choices = list(TORCH.keys())),
+			inquirer.List('onnxruntime', message = wording.get('install_dependency_help').format(dependency = 'onnxruntime'), choices = list(ONNXRUNTIMES.keys()))
+		])
+	if answers:
+		torch = answers['torch']
+		torch_wheel = TORCH[torch]
+		onnxruntime = answers['onnxruntime']
+		onnxruntime_name, onnxruntime_version = ONNXRUNTIMES[onnxruntime]
+		subprocess.call([ 'pip', 'uninstall', 'torch', '-y' ])
+		if torch_wheel == 'default':
+			subprocess.call([ 'pip', 'install', '-r', 'requirements.txt' ])
+		else:
+			subprocess.call([ 'pip', 'install', '-r', 'requirements.txt', '--extra-index-url', 'https://download.pytorch.org/whl/' + torch_wheel ])
+		subprocess.call([ 'pip', 'uninstall', 'onnxruntime', onnxruntime_name, '-y' ])
+		subprocess.call([ 'pip', 'install', onnxruntime_name + '==' + onnxruntime_version ])

facefusion/metadata.py

@@ -0,0 +1,13 @@
+METADATA =\
+{
+	'name': 'FaceFusion',
+	'description': 'Next generation face swapper and enhancer',
+	'version': '2.0.0',
+	'license': 'MIT',
+	'author': 'Henry Ruhs',
+	'url': 'https://facefusion.io'
+}
+
+
+def get(key : str) -> str:
+	return METADATA[key]

facefusion/processors/frame/choices.py

@@ -0,0 +1,13 @@
+from typing import List
+import numpy
+
+from facefusion.processors.frame.typings import FaceSwapperModel, FaceEnhancerModel, FrameEnhancerModel, FaceDebuggerItem
+
+face_swapper_models : List[FaceSwapperModel] = [ 'blendface_256', 'inswapper_128', 'inswapper_128_fp16', 'simswap_256', 'simswap_512_unofficial' ]
+face_enhancer_models : List[FaceEnhancerModel] = [ 'codeformer', 'gfpgan_1.2', 'gfpgan_1.3', 'gfpgan_1.4', 'gpen_bfr_256', 'gpen_bfr_512', 'restoreformer' ]
+frame_enhancer_models : List[FrameEnhancerModel] = [ 'real_esrgan_x2plus', 'real_esrgan_x4plus', 'real_esrnet_x4plus' ]
+
+face_enhancer_blend_range : List[int] = numpy.arange(0, 101, 1).tolist()
+frame_enhancer_blend_range : List[int] = numpy.arange(0, 101, 1).tolist()
+
+face_debugger_items : List[FaceDebuggerItem] = [ 'bbox', 'kps', 'face-mask', 'score' ]

facefusion/processors/frame/core.py

@@ -0,0 +1,96 @@
+import sys
+import importlib
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from queue import Queue
+from types import ModuleType
+from typing import Any, List
+from tqdm import tqdm
+
+import facefusion.globals
+from facefusion.typing import Process_Frames
+from facefusion import wording
+from facefusion.utilities import encode_execution_providers
+
+FRAME_PROCESSORS_MODULES : List[ModuleType] = []
+FRAME_PROCESSORS_METHODS =\
+[
+	'get_frame_processor',
+	'clear_frame_processor',
+	'get_options',
+	'set_options',
+	'register_args',
+	'apply_args',
+	'pre_check',
+	'pre_process',
+	'process_frame',
+	'process_frames',
+	'process_image',
+	'process_video',
+	'post_process'
+]
+
+
+def load_frame_processor_module(frame_processor : str) -> Any:
+	try:
+		frame_processor_module = importlib.import_module('facefusion.processors.frame.modules.' + frame_processor)
+		for method_name in FRAME_PROCESSORS_METHODS:
+			if not hasattr(frame_processor_module, method_name):
+				raise NotImplementedError
+	except ModuleNotFoundError:
+		sys.exit(wording.get('frame_processor_not_loaded').format(frame_processor = frame_processor))
+	except NotImplementedError:
+		sys.exit(wording.get('frame_processor_not_implemented').format(frame_processor = frame_processor))
+	return frame_processor_module
+
+
+def get_frame_processors_modules(frame_processors : List[str]) -> List[ModuleType]:
+	global FRAME_PROCESSORS_MODULES
+
+	if not FRAME_PROCESSORS_MODULES:
+		for frame_processor in frame_processors:
+			frame_processor_module = load_frame_processor_module(frame_processor)
+			FRAME_PROCESSORS_MODULES.append(frame_processor_module)
+	return FRAME_PROCESSORS_MODULES
+
+
+def clear_frame_processors_modules() -> None:
+	global FRAME_PROCESSORS_MODULES
+
+	for frame_processor_module in get_frame_processors_modules(facefusion.globals.frame_processors):
+		frame_processor_module.clear_frame_processor()
+	FRAME_PROCESSORS_MODULES = []
+
+
+def multi_process_frames(source_path : str, temp_frame_paths : List[str], process_frames : Process_Frames) -> None:
+	with tqdm(total = len(temp_frame_paths), desc = wording.get('processing'), unit = 'frame', ascii = ' =') as progress:
+		progress.set_postfix(
+		{
+			'execution_providers': encode_execution_providers(facefusion.globals.execution_providers),
+			'execution_thread_count': facefusion.globals.execution_thread_count,
+			'execution_queue_count': facefusion.globals.execution_queue_count
+		})
+		with ThreadPoolExecutor(max_workers = facefusion.globals.execution_thread_count) as executor:
+			futures = []
+			queue_temp_frame_paths : Queue[str] = create_queue(temp_frame_paths)
+			queue_per_future = max(len(temp_frame_paths) // facefusion.globals.execution_thread_count * facefusion.globals.execution_queue_count, 1)
+			while not queue_temp_frame_paths.empty():
+				payload_temp_frame_paths = pick_queue(queue_temp_frame_paths, queue_per_future)
+				future = executor.submit(process_frames, source_path, payload_temp_frame_paths, progress.update)
+				futures.append(future)
+			for future_done in as_completed(futures):
+				future_done.result()
+
+
+def create_queue(temp_frame_paths : List[str]) -> Queue[str]:
+	queue : Queue[str] = Queue()
+	for frame_path in temp_frame_paths:
+		queue.put(frame_path)
+	return queue
+
+
+def pick_queue(queue : Queue[str], queue_per_future : int) -> List[str]:
+	queues = []
+	for _ in range(queue_per_future):
+		if not queue.empty():
+			queues.append(queue.get())
+	return queues

facefusion/processors/frame/globals.py

@@ -0,0 +1,10 @@
+from typing import List, Optional
+
+from facefusion.processors.frame.typings import FaceSwapperModel, FaceEnhancerModel, FrameEnhancerModel, FaceDebuggerItem
+
+face_swapper_model : Optional[FaceSwapperModel] = None
+face_enhancer_model : Optional[FaceEnhancerModel] = None
+face_enhancer_blend : Optional[int] = None
+frame_enhancer_model : Optional[FrameEnhancerModel] = None
+frame_enhancer_blend : Optional[int] = None
+face_debugger_items : Optional[List[FaceDebuggerItem]] = None

facefusion/processors/frame/modules/face_debugger.py

@@ -0,0 +1,123 @@
+from typing import Any, List, Literal
+from argparse import ArgumentParser
+import cv2
+import numpy
+
+import facefusion.globals
+import facefusion.processors.frame.core as frame_processors
+from facefusion import wording
+from facefusion.face_analyser import get_one_face, get_many_faces, find_similar_faces, clear_face_analyser
+from facefusion.face_reference import get_face_reference
+from facefusion.content_analyser import clear_content_analyser
+from facefusion.typing import Face, Frame, Update_Process, ProcessMode
+from facefusion.vision import read_image, read_static_image, write_image
+from facefusion.face_helper import warp_face, create_static_mask_frame
+from facefusion.processors.frame import globals as frame_processors_globals, choices as frame_processors_choices
+
+NAME = 'FACEFUSION.FRAME_PROCESSOR.FACE_DEBUGGER'
+
+
+def get_frame_processor() -> None:
+	pass
+
+
+def clear_frame_processor() -> None:
+	pass
+
+
+def get_options(key : Literal['model']) -> None:
+	pass
+
+
+def set_options(key : Literal['model'], value : Any) -> None:
+	pass
+
+
+def register_args(program : ArgumentParser) -> None:
+	program.add_argument('--face-debugger-items', help = wording.get('face_debugger_items_help'), dest = 'face_debugger_items', default = [ 'kps', 'face-mask' ], choices = frame_processors_choices.face_debugger_items, nargs = '+')
+
+
+def apply_args(program : ArgumentParser) -> None:
+	args = program.parse_args()
+	frame_processors_globals.face_debugger_items = args.face_debugger_items
+
+
+def pre_check() -> bool:
+	return True
+
+
+def pre_process(mode : ProcessMode) -> bool:
+	return True
+
+
+def post_process() -> None:
+	clear_frame_processor()
+	clear_face_analyser()
+	clear_content_analyser()
+
+
+def debug_face(source_face : Face, target_face : Face, temp_frame : Frame) -> Frame:
+	primary_color = (0, 0, 255)
+	secondary_color = (0, 255, 0)
+	bounding_box = target_face.bbox.astype(numpy.int32)
+	if 'bbox' in frame_processors_globals.face_debugger_items:
+		cv2.rectangle(temp_frame, (bounding_box[0], bounding_box[1]), (bounding_box[2], bounding_box[3]), secondary_color, 2)
+	if 'face-mask' in frame_processors_globals.face_debugger_items:
+		crop_frame, affine_matrix = warp_face(temp_frame, target_face.kps, 'arcface_v2', (128, 128))
+		inverse_matrix = cv2.invertAffineTransform(affine_matrix)
+		temp_frame_size = temp_frame.shape[:2][::-1]
+		mask_frame = create_static_mask_frame(crop_frame.shape[:2], 0, facefusion.globals.face_mask_padding)
+		mask_frame[mask_frame > 0] = 255
+		inverse_mask_frame = cv2.warpAffine(mask_frame.astype(numpy.uint8), inverse_matrix, temp_frame_size)
+		inverse_mask_contours = cv2.findContours(inverse_mask_frame, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
+		cv2.drawContours(temp_frame, inverse_mask_contours, 0, primary_color, 2)
+	if bounding_box[3] - bounding_box[1] > 60 and bounding_box[2] - bounding_box[0] > 60:
+		if 'kps' in frame_processors_globals.face_debugger_items:
+			kps = target_face.kps.astype(numpy.int32)
+			for index in range(kps.shape[0]):
+				cv2.circle(temp_frame, (kps[index][0], kps[index][1]), 3, primary_color, -1)
+		if 'score' in frame_processors_globals.face_debugger_items:
+			score_text = str(round(target_face.score, 2))
+			score_position = (bounding_box[0] + 10, bounding_box[1] + 20)
+			cv2.putText(temp_frame, score_text, score_position, cv2.FONT_HERSHEY_SIMPLEX, 0.5, secondary_color, 2)
+	return temp_frame
+
+
+def process_frame(source_face : Face, reference_face : Face, temp_frame : Frame) -> Frame:
+	if 'reference' in facefusion.globals.face_selector_mode:
+		similar_faces = find_similar_faces(temp_frame, reference_face, facefusion.globals.reference_face_distance)
+		if similar_faces:
+			for similar_face in similar_faces:
+				temp_frame = debug_face(source_face, similar_face, temp_frame)
+	if 'one' in facefusion.globals.face_selector_mode:
+		target_face = get_one_face(temp_frame)
+		if target_face:
+			temp_frame = debug_face(source_face, target_face, temp_frame)
+	if 'many' in facefusion.globals.face_selector_mode:
+		many_faces = get_many_faces(temp_frame)
+		if many_faces:
+			for target_face in many_faces:
+				temp_frame = debug_face(source_face, target_face, temp_frame)
+	return temp_frame
+
+
+def process_frames(source_path : str, temp_frame_paths : List[str], update_progress : Update_Process) -> None:
+	source_face = get_one_face(read_static_image(source_path))
+	reference_face = get_face_reference() if 'reference' in facefusion.globals.face_selector_mode else None
+	for temp_frame_path in temp_frame_paths:
+		temp_frame = read_image(temp_frame_path)
+		result_frame = process_frame(source_face, reference_face, temp_frame)
+		write_image(temp_frame_path, result_frame)
+		update_progress()
+
+
+def process_image(source_path : str, target_path : str, output_path : str) -> None:
+	source_face = get_one_face(read_static_image(source_path))
+	target_frame = read_static_image(target_path)
+	reference_face = get_one_face(target_frame, facefusion.globals.reference_face_position) if 'reference' in facefusion.globals.face_selector_mode else None
+	result_frame = process_frame(source_face, reference_face, target_frame)
+	write_image(output_path, result_frame)
+
+
+def process_video(source_path : str, temp_frame_paths : List[str]) -> None:
+	frame_processors.multi_process_frames(source_path, temp_frame_paths, process_frames)

facefusion/processors/frame/modules/face_enhancer.py

@@ -0,0 +1,221 @@
+from typing import Any, List, Dict, Literal, Optional
+from argparse import ArgumentParser
+import cv2
+import threading
+import numpy
+import onnxruntime
+
+import facefusion.globals
+import facefusion.processors.frame.core as frame_processors
+from facefusion import wording
+from facefusion.face_analyser import get_many_faces, clear_face_analyser
+from facefusion.face_helper import warp_face, paste_back
+from facefusion.content_analyser import clear_content_analyser
+from facefusion.typing import Face, Frame, Update_Process, ProcessMode, ModelValue, OptionsWithModel
+from facefusion.utilities import conditional_download, resolve_relative_path, is_image, is_video, is_file, is_download_done, create_metavar, update_status
+from facefusion.vision import read_image, read_static_image, write_image
+from facefusion.processors.frame import globals as frame_processors_globals
+from facefusion.processors.frame import choices as frame_processors_choices
+
+FRAME_PROCESSOR = None
+THREAD_SEMAPHORE : threading.Semaphore = threading.Semaphore()
+THREAD_LOCK : threading.Lock = threading.Lock()
+NAME = 'FACEFUSION.FRAME_PROCESSOR.FACE_ENHANCER'
+MODELS : Dict[str, ModelValue] =\
+{
+	'codeformer':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/codeformer.onnx',
+		'path': resolve_relative_path('../.assets/models/codeformer.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	},
+	'gfpgan_1.2':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gfpgan_1.2.onnx',
+		'path': resolve_relative_path('../.assets/models/gfpgan_1.2.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	},
+	'gfpgan_1.3':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gfpgan_1.3.onnx',
+		'path': resolve_relative_path('../.assets/models/gfpgan_1.3.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	},
+	'gfpgan_1.4':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gfpgan_1.4.onnx',
+		'path': resolve_relative_path('../.assets/models/gfpgan_1.4.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	},
+	'gpen_bfr_256':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gpen_bfr_256.onnx',
+		'path': resolve_relative_path('../.assets/models/gpen_bfr_256.onnx'),
+		'template': 'arcface_v2',
+		'size': (128, 256)
+	},
+	'gpen_bfr_512':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/gpen_bfr_512.onnx',
+		'path': resolve_relative_path('../.assets/models/gpen_bfr_512.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	},
+	'restoreformer':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/restoreformer.onnx',
+		'path': resolve_relative_path('../.assets/models/restoreformer.onnx'),
+		'template': 'ffhq',
+		'size': (512, 512)
+	}
+}
+OPTIONS : Optional[OptionsWithModel] = None
+
+
+def get_frame_processor() -> Any:
+	global FRAME_PROCESSOR
+
+	with THREAD_LOCK:
+		if FRAME_PROCESSOR is None:
+			model_path = get_options('model').get('path')
+			FRAME_PROCESSOR = onnxruntime.InferenceSession(model_path, providers = facefusion.globals.execution_providers)
+	return FRAME_PROCESSOR
+
+
+def clear_frame_processor() -> None:
+	global FRAME_PROCESSOR
+
+	FRAME_PROCESSOR = None
+
+
+def get_options(key : Literal['model']) -> Any:
+	global OPTIONS
+
+	if OPTIONS is None:
+		OPTIONS =\
+		{
+			'model': MODELS[frame_processors_globals.face_enhancer_model]
+		}
+	return OPTIONS.get(key)
+
+
+def set_options(key : Literal['model'], value : Any) -> None:
+	global OPTIONS
+
+	OPTIONS[key] = value
+
+
+def register_args(program : ArgumentParser) -> None:
+	program.add_argument('--face-enhancer-model', help = wording.get('frame_processor_model_help'), dest = 'face_enhancer_model', default = 'gfpgan_1.4', choices = frame_processors_choices.face_enhancer_models)
+	program.add_argument('--face-enhancer-blend', help = wording.get('frame_processor_blend_help'), dest = 'face_enhancer_blend', type = int, default = 80, choices = frame_processors_choices.face_enhancer_blend_range, metavar = create_metavar(frame_processors_choices.face_enhancer_blend_range))
+
+
+def apply_args(program : ArgumentParser) -> None:
+	args = program.parse_args()
+	frame_processors_globals.face_enhancer_model = args.face_enhancer_model
+	frame_processors_globals.face_enhancer_blend = args.face_enhancer_blend
+
+
+def pre_check() -> bool:
+	if not facefusion.globals.skip_download:
+		download_directory_path = resolve_relative_path('../.assets/models')
+		model_url = get_options('model').get('url')
+		conditional_download(download_directory_path, [ model_url ])
+	return True
+
+
+def pre_process(mode : ProcessMode) -> bool:
+	model_url = get_options('model').get('url')
+	model_path = get_options('model').get('path')
+	if not facefusion.globals.skip_download and not is_download_done(model_url, model_path):
+		update_status(wording.get('model_download_not_done') + wording.get('exclamation_mark'), NAME)
+		return False
+	elif not is_file(model_path):
+		update_status(wording.get('model_file_not_present') + wording.get('exclamation_mark'), NAME)
+		return False
+	if mode in [ 'output', 'preview' ] and not is_image(facefusion.globals.target_path) and not is_video(facefusion.globals.target_path):
+		update_status(wording.get('select_image_or_video_target') + wording.get('exclamation_mark'), NAME)
+		return False
+	if mode == 'output' and not facefusion.globals.output_path:
+		update_status(wording.get('select_file_or_directory_output') + wording.get('exclamation_mark'), NAME)
+		return False
+	return True
+
+
+def post_process() -> None:
+	clear_frame_processor()
+	clear_face_analyser()
+	clear_content_analyser()
+	read_static_image.cache_clear()
+
+
+def enhance_face(target_face: Face, temp_frame: Frame) -> Frame:
+	frame_processor = get_frame_processor()
+	model_template = get_options('model').get('template')
+	model_size = get_options('model').get('size')
+	crop_frame, affine_matrix = warp_face(temp_frame, target_face.kps, model_template, model_size)
+	crop_frame = prepare_crop_frame(crop_frame)
+	frame_processor_inputs = {}
+	for frame_processor_input in frame_processor.get_inputs():
+		if frame_processor_input.name == 'input':
+			frame_processor_inputs[frame_processor_input.name] = crop_frame
+		if frame_processor_input.name == 'weight':
+			frame_processor_inputs[frame_processor_input.name] = numpy.array([ 1 ], dtype = numpy.double)
+	with THREAD_SEMAPHORE:
+		crop_frame = frame_processor.run(None, frame_processor_inputs)[0][0]
+	crop_frame = normalize_crop_frame(crop_frame)
+	paste_frame = paste_back(temp_frame, crop_frame, affine_matrix, facefusion.globals.face_mask_blur, (0, 0, 0, 0))
+	temp_frame = blend_frame(temp_frame, paste_frame)
+	return temp_frame
+
+
+def prepare_crop_frame(crop_frame : Frame) -> Frame:
+	crop_frame = crop_frame[:, :, ::-1] / 255.0
+	crop_frame = (crop_frame - 0.5) / 0.5
+	crop_frame = numpy.expand_dims(crop_frame.transpose(2, 0, 1), axis = 0).astype(numpy.float32)
+	return crop_frame
+
+
+def normalize_crop_frame(crop_frame : Frame) -> Frame:
+	crop_frame = numpy.clip(crop_frame, -1, 1)
+	crop_frame = (crop_frame + 1) / 2
+	crop_frame = crop_frame.transpose(1, 2, 0)
+	crop_frame = (crop_frame * 255.0).round()
+	crop_frame = crop_frame.astype(numpy.uint8)[:, :, ::-1]
+	return crop_frame
+
+
+def blend_frame(temp_frame : Frame, paste_frame : Frame) -> Frame:
+	face_enhancer_blend = 1 - (frame_processors_globals.face_enhancer_blend / 100)
+	temp_frame = cv2.addWeighted(temp_frame, face_enhancer_blend, paste_frame, 1 - face_enhancer_blend, 0)
+	return temp_frame
+
+
+def process_frame(source_face : Face, reference_face : Face, temp_frame : Frame) -> Frame:
+	many_faces = get_many_faces(temp_frame)
+	if many_faces:
+		for target_face in many_faces:
+			temp_frame = enhance_face(target_face, temp_frame)
+	return temp_frame
+
+
+def process_frames(source_path : str, temp_frame_paths : List[str], update_progress : Update_Process) -> None:
+	for temp_frame_path in temp_frame_paths:
+		temp_frame = read_image(temp_frame_path)
+		result_frame = process_frame(None, None, temp_frame)
+		write_image(temp_frame_path, result_frame)
+		update_progress()
+
+
+def process_image(source_path : str, target_path : str, output_path : str) -> None:
+	target_frame = read_static_image(target_path)
+	result_frame = process_frame(None, None, target_frame)
+	write_image(output_path, result_frame)
+
+
+def process_video(source_path : str, temp_frame_paths : List[str]) -> None:
+	frame_processors.multi_process_frames(None, temp_frame_paths, process_frames)

facefusion/processors/frame/modules/face_swapper.py

@@ -0,0 +1,283 @@
+from typing import Any, List, Dict, Literal, Optional
+from argparse import ArgumentParser
+import threading
+import numpy
+import onnx
+import onnxruntime
+from onnx import numpy_helper
+
+import facefusion.globals
+import facefusion.processors.frame.core as frame_processors
+from facefusion import wording
+from facefusion.face_analyser import get_one_face, get_many_faces, find_similar_faces, clear_face_analyser
+from facefusion.face_helper import warp_face, paste_back
+from facefusion.face_reference import get_face_reference
+from facefusion.content_analyser import clear_content_analyser
+from facefusion.typing import Face, Frame, Update_Process, ProcessMode, ModelValue, OptionsWithModel, Embedding
+from facefusion.utilities import conditional_download, resolve_relative_path, is_image, is_video, is_file, is_download_done, update_status
+from facefusion.vision import read_image, read_static_image, write_image
+from facefusion.processors.frame import globals as frame_processors_globals
+from facefusion.processors.frame import choices as frame_processors_choices
+
+FRAME_PROCESSOR = None
+MODEL_MATRIX = None
+THREAD_LOCK : threading.Lock = threading.Lock()
+NAME = 'FACEFUSION.FRAME_PROCESSOR.FACE_SWAPPER'
+MODELS : Dict[str, ModelValue] =\
+{
+	'blendface_256':
+	{
+		'type': 'blendface',
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/blendface_256.onnx',
+		'path': resolve_relative_path('../.assets/models/blendface_256.onnx'),
+		'template': 'ffhq',
+		'size': (512, 256),
+		'mean': [ 0.0, 0.0, 0.0 ],
+		'standard_deviation': [ 1.0, 1.0, 1.0 ]
+	},
+	'inswapper_128':
+	{
+		'type': 'inswapper',
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/inswapper_128.onnx',
+		'path': resolve_relative_path('../.assets/models/inswapper_128.onnx'),
+		'template': 'arcface_v2',
+		'size': (128, 128),
+		'mean': [ 0.0, 0.0, 0.0 ],
+		'standard_deviation': [ 1.0, 1.0, 1.0 ]
+	},
+	'inswapper_128_fp16':
+	{
+		'type': 'inswapper',
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/inswapper_128_fp16.onnx',
+		'path': resolve_relative_path('../.assets/models/inswapper_128_fp16.onnx'),
+		'template': 'arcface_v2',
+		'size': (128, 128),
+		'mean': [ 0.0, 0.0, 0.0 ],
+		'standard_deviation': [ 1.0, 1.0, 1.0 ]
+	},
+	'simswap_256':
+	{
+		'type': 'simswap',
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/simswap_256.onnx',
+		'path': resolve_relative_path('../.assets/models/simswap_256.onnx'),
+		'template': 'arcface_v1',
+		'size': (112, 256),
+		'mean': [ 0.485, 0.456, 0.406 ],
+		'standard_deviation': [ 0.229, 0.224, 0.225 ]
+	},
+	'simswap_512_unofficial':
+	{
+		'type': 'simswap',
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/simswap_512_unofficial.onnx',
+		'path': resolve_relative_path('../.assets/models/simswap_512_unofficial.onnx'),
+		'template': 'arcface_v1',
+		'size': (112, 512),
+		'mean': [ 0.0, 0.0, 0.0 ],
+		'standard_deviation': [ 1.0, 1.0, 1.0 ]
+	}
+}
+OPTIONS : Optional[OptionsWithModel] = None
+
+
+def get_frame_processor() -> Any:
+	global FRAME_PROCESSOR
+
+	with THREAD_LOCK:
+		if FRAME_PROCESSOR is None:
+			model_path = get_options('model').get('path')
+			FRAME_PROCESSOR = onnxruntime.InferenceSession(model_path, providers = facefusion.globals.execution_providers)
+	return FRAME_PROCESSOR
+
+
+def clear_frame_processor() -> None:
+	global FRAME_PROCESSOR
+
+	FRAME_PROCESSOR = None
+
+
+def get_model_matrix() -> Any:
+	global MODEL_MATRIX
+
+	with THREAD_LOCK:
+		if MODEL_MATRIX is None:
+			model_path = get_options('model').get('path')
+			model = onnx.load(model_path)
+			MODEL_MATRIX = numpy_helper.to_array(model.graph.initializer[-1])
+	return MODEL_MATRIX
+
+
+def clear_model_matrix() -> None:
+	global MODEL_MATRIX
+
+	MODEL_MATRIX = None
+
+
+def get_options(key : Literal['model']) -> Any:
+	global OPTIONS
+
+	if OPTIONS is None:
+		OPTIONS =\
+		{
+			'model': MODELS[frame_processors_globals.face_swapper_model]
+		}
+	return OPTIONS.get(key)
+
+
+def set_options(key : Literal['model'], value : Any) -> None:
+	global OPTIONS
+
+	OPTIONS[key] = value
+
+
+def register_args(program : ArgumentParser) -> None:
+	program.add_argument('--face-swapper-model', help = wording.get('frame_processor_model_help'), dest = 'face_swapper_model', default = 'inswapper_128', choices = frame_processors_choices.face_swapper_models)
+
+
+def apply_args(program : ArgumentParser) -> None:
+	args = program.parse_args()
+	frame_processors_globals.face_swapper_model = args.face_swapper_model
+	if args.face_swapper_model == 'blendface_256':
+		facefusion.globals.face_recognizer_model = 'arcface_blendface'
+	if args.face_swapper_model == 'inswapper_128' or args.face_swapper_model == 'inswapper_128_fp16':
+		facefusion.globals.face_recognizer_model = 'arcface_inswapper'
+	if args.face_swapper_model == 'simswap_256' or args.face_swapper_model == 'simswap_512_unofficial':
+		facefusion.globals.face_recognizer_model = 'arcface_simswap'
+
+
+def pre_check() -> bool:
+	if not facefusion.globals.skip_download:
+		download_directory_path = resolve_relative_path('../.assets/models')
+		model_url = get_options('model').get('url')
+		conditional_download(download_directory_path, [ model_url ])
+	return True
+
+
+def pre_process(mode : ProcessMode) -> bool:
+	model_url = get_options('model').get('url')
+	model_path = get_options('model').get('path')
+	if not facefusion.globals.skip_download and not is_download_done(model_url, model_path):
+		update_status(wording.get('model_download_not_done') + wording.get('exclamation_mark'), NAME)
+		return False
+	elif not is_file(model_path):
+		update_status(wording.get('model_file_not_present') + wording.get('exclamation_mark'), NAME)
+		return False
+	if not is_image(facefusion.globals.source_path):
+		update_status(wording.get('select_image_source') + wording.get('exclamation_mark'), NAME)
+		return False
+	elif not get_one_face(read_static_image(facefusion.globals.source_path)):
+		update_status(wording.get('no_source_face_detected') + wording.get('exclamation_mark'), NAME)
+		return False
+	if mode in [ 'output', 'preview' ] and not is_image(facefusion.globals.target_path) and not is_video(facefusion.globals.target_path):
+		update_status(wording.get('select_image_or_video_target') + wording.get('exclamation_mark'), NAME)
+		return False
+	if mode == 'output' and not facefusion.globals.output_path:
+		update_status(wording.get('select_file_or_directory_output') + wording.get('exclamation_mark'), NAME)
+		return False
+	return True
+
+
+def post_process() -> None:
+	clear_frame_processor()
+	clear_model_matrix()
+	clear_face_analyser()
+	clear_content_analyser()
+	read_static_image.cache_clear()
+
+
+def swap_face(source_face : Face, target_face : Face, temp_frame : Frame) -> Frame:
+	frame_processor = get_frame_processor()
+	model_template = get_options('model').get('template')
+	model_size = get_options('model').get('size')
+	model_type = get_options('model').get('type')
+	crop_frame, affine_matrix = warp_face(temp_frame, target_face.kps, model_template, model_size)
+	crop_frame = prepare_crop_frame(crop_frame)
+	frame_processor_inputs = {}
+	for frame_processor_input in frame_processor.get_inputs():
+		if frame_processor_input.name == 'source':
+			if model_type == 'blendface':
+				frame_processor_inputs[frame_processor_input.name] = prepare_source_frame(source_face)
+			else:
+				frame_processor_inputs[frame_processor_input.name] = prepare_source_embedding(source_face)
+		if frame_processor_input.name == 'target':
+			frame_processor_inputs[frame_processor_input.name] = crop_frame
+	crop_frame = frame_processor.run(None, frame_processor_inputs)[0][0]
+	crop_frame = normalize_crop_frame(crop_frame)
+	temp_frame = paste_back(temp_frame, crop_frame, affine_matrix, facefusion.globals.face_mask_blur, facefusion.globals.face_mask_padding)
+	return temp_frame
+
+
+def prepare_source_frame(source_face : Face) -> numpy.ndarray[Any, Any]:
+	source_frame = read_static_image(facefusion.globals.source_path)
+	source_frame, _ = warp_face(source_frame, source_face.kps, 'arcface_v2', (112, 112))
+	source_frame = source_frame[:, :, ::-1] / 255.0
+	source_frame = source_frame.transpose(2, 0, 1)
+	source_frame = numpy.expand_dims(source_frame, axis = 0).astype(numpy.float32)
+	return source_frame
+
+
+def prepare_source_embedding(source_face : Face) -> Embedding:
+	model_type = get_options('model').get('type')
+	if model_type == 'inswapper':
+		model_matrix = get_model_matrix()
+		source_embedding = source_face.embedding.reshape((1, -1))
+		source_embedding = numpy.dot(source_embedding, model_matrix) / numpy.linalg.norm(source_embedding)
+	else:
+		source_embedding = source_face.normed_embedding.reshape(1, -1)
+	return source_embedding
+
+
+def prepare_crop_frame(crop_frame : Frame) -> Frame:
+	model_mean = get_options('model').get('mean')
+	model_standard_deviation = get_options('model').get('standard_deviation')
+	crop_frame = crop_frame[:, :, ::-1] / 255.0
+	crop_frame = (crop_frame - model_mean) / model_standard_deviation
+	crop_frame = crop_frame.transpose(2, 0, 1)
+	crop_frame = numpy.expand_dims(crop_frame, axis = 0).astype(numpy.float32)
+	return crop_frame
+
+
+def normalize_crop_frame(crop_frame : Frame) -> Frame:
+	crop_frame = crop_frame.transpose(1, 2, 0)
+	crop_frame = (crop_frame * 255.0).round()
+	crop_frame = crop_frame[:, :, ::-1].astype(numpy.uint8)
+	return crop_frame
+
+
+def process_frame(source_face : Face, reference_face : Face, temp_frame : Frame) -> Frame:
+	if 'reference' in facefusion.globals.face_selector_mode:
+		similar_faces = find_similar_faces(temp_frame, reference_face, facefusion.globals.reference_face_distance)
+		if similar_faces:
+			for similar_face in similar_faces:
+				temp_frame = swap_face(source_face, similar_face, temp_frame)
+	if 'one' in facefusion.globals.face_selector_mode:
+		target_face = get_one_face(temp_frame)
+		if target_face:
+			temp_frame = swap_face(source_face, target_face, temp_frame)
+	if 'many' in facefusion.globals.face_selector_mode:
+		many_faces = get_many_faces(temp_frame)
+		if many_faces:
+			for target_face in many_faces:
+				temp_frame = swap_face(source_face, target_face, temp_frame)
+	return temp_frame
+
+
+def process_frames(source_path : str, temp_frame_paths : List[str], update_progress : Update_Process) -> None:
+	source_face = get_one_face(read_static_image(source_path))
+	reference_face = get_face_reference() if 'reference' in facefusion.globals.face_selector_mode else None
+	for temp_frame_path in temp_frame_paths:
+		temp_frame = read_image(temp_frame_path)
+		result_frame = process_frame(source_face, reference_face, temp_frame)
+		write_image(temp_frame_path, result_frame)
+		update_progress()
+
+
+def process_image(source_path : str, target_path : str, output_path : str) -> None:
+	source_face = get_one_face(read_static_image(source_path))
+	target_frame = read_static_image(target_path)
+	reference_face = get_one_face(target_frame, facefusion.globals.reference_face_position) if 'reference' in facefusion.globals.face_selector_mode else None
+	result_frame = process_frame(source_face, reference_face, target_frame)
+	write_image(output_path, result_frame)
+
+
+def process_video(source_path : str, temp_frame_paths : List[str]) -> None:
+	frame_processors.multi_process_frames(source_path, temp_frame_paths, process_frames)

facefusion/processors/frame/modules/frame_enhancer.py

@@ -0,0 +1,165 @@
+from typing import Any, List, Dict, Literal, Optional
+from argparse import ArgumentParser
+import threading
+import cv2
+from basicsr.archs.rrdbnet_arch import RRDBNet
+from realesrgan import RealESRGANer
+
+import facefusion.globals
+import facefusion.processors.frame.core as frame_processors
+from facefusion import wording
+from facefusion.face_analyser import clear_face_analyser
+from facefusion.content_analyser import clear_content_analyser
+from facefusion.typing import Frame, Face, Update_Process, ProcessMode, ModelValue, OptionsWithModel
+from facefusion.utilities import conditional_download, resolve_relative_path, is_file, is_download_done, map_device, create_metavar, update_status
+from facefusion.vision import read_image, read_static_image, write_image
+from facefusion.processors.frame import globals as frame_processors_globals
+from facefusion.processors.frame import choices as frame_processors_choices
+
+FRAME_PROCESSOR = None
+THREAD_SEMAPHORE : threading.Semaphore = threading.Semaphore()
+THREAD_LOCK : threading.Lock = threading.Lock()
+NAME = 'FACEFUSION.FRAME_PROCESSOR.FRAME_ENHANCER'
+MODELS: Dict[str, ModelValue] =\
+{
+	'real_esrgan_x2plus':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/real_esrgan_x2plus.pth',
+		'path': resolve_relative_path('../.assets/models/real_esrgan_x2plus.pth'),
+		'scale': 2
+	},
+	'real_esrgan_x4plus':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/real_esrgan_x4plus.pth',
+		'path': resolve_relative_path('../.assets/models/real_esrgan_x4plus.pth'),
+		'scale': 4
+	},
+	'real_esrnet_x4plus':
+	{
+		'url': 'https://github.com/facefusion/facefusion-assets/releases/download/models/real_esrnet_x4plus.pth',
+		'path': resolve_relative_path('../.assets/models/real_esrnet_x4plus.pth'),
+		'scale': 4
+	}
+}
+OPTIONS : Optional[OptionsWithModel] = None
+
+
+def get_frame_processor() -> Any:
+	global FRAME_PROCESSOR
+
+	with THREAD_LOCK:
+		if FRAME_PROCESSOR is None:
+			model_path = get_options('model').get('path')
+			model_scale = get_options('model').get('scale')
+			FRAME_PROCESSOR = RealESRGANer(
+				model_path = model_path,
+				model = RRDBNet(
+					num_in_ch = 3,
+					num_out_ch = 3,
+					scale = model_scale
+				),
+				device = map_device(facefusion.globals.execution_providers),
+				scale = model_scale
+			)
+	return FRAME_PROCESSOR
+
+
+def clear_frame_processor() -> None:
+	global FRAME_PROCESSOR
+
+	FRAME_PROCESSOR = None
+
+
+def get_options(key : Literal['model']) -> Any:
+	global OPTIONS
+
+	if OPTIONS is None:
+		OPTIONS =\
+		{
+			'model': MODELS[frame_processors_globals.frame_enhancer_model]
+		}
+	return OPTIONS.get(key)
+
+
+def set_options(key : Literal['model'], value : Any) -> None:
+	global OPTIONS
+
+	OPTIONS[key] = value
+
+
+def register_args(program : ArgumentParser) -> None:
+	program.add_argument('--frame-enhancer-model', help = wording.get('frame_processor_model_help'), dest = 'frame_enhancer_model', default = 'real_esrgan_x2plus', choices = frame_processors_choices.frame_enhancer_models)
+	program.add_argument('--frame-enhancer-blend', help = wording.get('frame_processor_blend_help'), dest = 'frame_enhancer_blend', type = int, default = 80, choices = frame_processors_choices.frame_enhancer_blend_range, metavar = create_metavar(frame_processors_choices.frame_enhancer_blend_range))
+
+
+def apply_args(program : ArgumentParser) -> None:
+	args = program.parse_args()
+	frame_processors_globals.frame_enhancer_model = args.frame_enhancer_model
+	frame_processors_globals.frame_enhancer_blend = args.frame_enhancer_blend
+
+
+def pre_check() -> bool:
+	if not facefusion.globals.skip_download:
+		download_directory_path = resolve_relative_path('../.assets/models')
+		model_url = get_options('model').get('url')
+		conditional_download(download_directory_path, [ model_url ])
+	return True
+
+
+def pre_process(mode : ProcessMode) -> bool:
+	model_url = get_options('model').get('url')
+	model_path = get_options('model').get('path')
+	if not facefusion.globals.skip_download and not is_download_done(model_url, model_path):
+		update_status(wording.get('model_download_not_done') + wording.get('exclamation_mark'), NAME)
+		return False
+	elif not is_file(model_path):
+		update_status(wording.get('model_file_not_present') + wording.get('exclamation_mark'), NAME)
+		return False
+	if mode == 'output' and not facefusion.globals.output_path:
+		update_status(wording.get('select_file_or_directory_output') + wording.get('exclamation_mark'), NAME)
+		return False
+	return True
+
+
+def post_process() -> None:
+	clear_frame_processor()
+	clear_face_analyser()
+	clear_content_analyser()
+	read_static_image.cache_clear()
+
+
+def enhance_frame(temp_frame : Frame) -> Frame:
+	with THREAD_SEMAPHORE:
+		paste_frame, _ = get_frame_processor().enhance(temp_frame)
+		temp_frame = blend_frame(temp_frame, paste_frame)
+	return temp_frame
+
+
+def blend_frame(temp_frame : Frame, paste_frame : Frame) -> Frame:
+	frame_enhancer_blend = 1 - (frame_processors_globals.frame_enhancer_blend / 100)
+	paste_frame_height, paste_frame_width = paste_frame.shape[0:2]
+	temp_frame = cv2.resize(temp_frame, (paste_frame_width, paste_frame_height))
+	temp_frame = cv2.addWeighted(temp_frame, frame_enhancer_blend, paste_frame, 1 - frame_enhancer_blend, 0)
+	return temp_frame
+
+
+def process_frame(source_face : Face, reference_face : Face, temp_frame : Frame) -> Frame:
+	return enhance_frame(temp_frame)
+
+
+def process_frames(source_path : str, temp_frame_paths : List[str], update_progress : Update_Process) -> None:
+	for temp_frame_path in temp_frame_paths:
+		temp_frame = read_image(temp_frame_path)
+		result_frame = process_frame(None, None, temp_frame)
+		write_image(temp_frame_path, result_frame)
+		update_progress()
+
+
+def process_image(source_path : str, target_path : str, output_path : str) -> None:
+	target_frame = read_static_image(target_path)
+	result = process_frame(None, None, target_frame)
+	write_image(output_path, result)
+
+
+def process_video(source_path : str, temp_frame_paths : List[str]) -> None:
+	frame_processors.multi_process_frames(None, temp_frame_paths, process_frames)

facefusion/processors/frame/typings.py

@@ -0,0 +1,7 @@
+from typing import Literal
+
+FaceSwapperModel = Literal['blendface_256', 'inswapper_128', 'inswapper_128_fp16', 'simswap_256', 'simswap_512_unofficial']
+FaceEnhancerModel = Literal['codeformer', 'gfpgan_1.2', 'gfpgan_1.3', 'gfpgan_1.4', 'gpen_bfr_256', 'gpen_bfr_512', 'restoreformer']
+FrameEnhancerModel = Literal['real_esrgan_x2plus', 'real_esrgan_x4plus', 'real_esrnet_x4plus']
+
+FaceDebuggerItem = Literal['bbox', 'kps', 'face-mask', 'score']