Next (#384)

* feat/yoloface (#334)

* added yolov8 to face_detector (#323)

* added yolov8 to face_detector

* added yolov8 to face_detector

* Initial cleanup and renaming

* Update README

* refactored detect_with_yoloface (#329)

* refactored detect_with_yoloface

* apply review

* Change order again

* Restore working code

* modified code (#330)

* refactored detect_with_yoloface

* apply review

* use temp_frame in detect_with_yoloface

* reorder

* modified

* reorder models

* Tiny cleanup

---------

Co-authored-by: tamoharu <133945583+tamoharu@users.noreply.github.com>

* include audio file functions (#336)

* Add testing for audio handlers

* Change order

* Fix naming

* Use correct typing in choices

* Update help message for arguments, Notation based wording approach (#347)

* Update help message for arguments, Notation based wording approach

* Fix installer

* Audio functions (#345)

* Update ffmpeg.py

* Create audio.py

* Update ffmpeg.py

* Update audio.py

* Update audio.py

* Update typing.py

* Update ffmpeg.py

* Update audio.py

* Rename Frame to VisionFrame (#346)

* Minor tidy up

* Introduce audio testing

* Add more todo for testing

* Add more todo for testing

* Fix indent

* Enable venv on the fly

* Enable venv on the fly

* Revert venv on the fly

* Revert venv on the fly

* Force Gradio to shut up

* Force Gradio to shut up

* Clear temp before processing

* Reduce terminal output

* include audio file functions

* Enforce output resolution on merge video

* Minor cleanups

* Add age and gender to face debugger items (#353)

* Add age and gender to face debugger items

* Rename like suggested in the code review

* Fix the output framerate vs. time

* Lip Sync (#356)

* Cli implementation of wav2lip

* - create get_first_item()
- remove non gan wav2lip model
- implement video memory strategy
- implement get_reference_frame()
- implement process_image()
- rearrange crop_mask_list
- implement test_cli

* Simplify testing

* Rename to lip syncer

* Fix testing

* Fix testing

* Minor cleanup

* Cuda 12 installer (#362)

* Make cuda nightly (12) the default

* Better keep legacy cuda just in case

* Use CUDA and ROCM versions

* Remove MacOS options from installer (CoreML include in default package)

* Add lip-syncer support to source component

* Add lip-syncer support to source component

* Fix the check in the source component

* Add target image check

* Introduce more helpers to suite the lip-syncer needs

* Downgrade onnxruntime as of buggy 1.17.0 release

* Revert "Downgrade onnxruntime as of buggy 1.17.0 release"

This reverts commit f4a7ae6824fed87f0be50906bbc7e2d61d00617b.

* More testing and add todos

* Fix the frame processor API to at least not throw errors

* Introduce dict based frame processor inputs (#364)

* Introduce dict based frame processor inputs

* Forgot to adjust webcam

* create path payloads (#365)

* create index payload to paths for process_frames

* rename to payload_paths

* This code now is poetry

* Fix the terminal output

* Make lip-syncer work in the preview

* Remove face debugger test for now

* Reoder reference_faces, Fix testing

* Use inswapper_128 on buggy onnxruntime 1.17.0

* Undo inswapper_128_fp16 duo broken onnxruntime 1.17.0

* Undo inswapper_128_fp16 duo broken onnxruntime 1.17.0

* Fix lip_syncer occluder & region mask issue

* Fix preview once in case there was no output video fps

* fix lip_syncer custom fps

* remove unused import

* Add 68 landmark functions (#367)

* Add 68 landmark model

* Add landmark to face object

* Re-arrange and modify typing

* Rename function

* Rearrange

* Rearrange

* ignore type

* ignore type

* change type

* ignore

* name

* Some cleanup

* Some cleanup

* Opps, I broke something

* Feat/face analyser refactoring (#369)

* Restructure face analyser and start TDD

* YoloFace and Yunet testing are passing

* Remove offset from yoloface detection

* Cleanup code

* Tiny fix

* Fix get_many_faces()

* Tiny fix (again)

* Use 320x320 fallback for retinaface

* Fix merging mashup

* Upload wave2lip model

* Upload 2dfan2 model and rename internal to face_predictor

* Downgrade onnxruntime for most cases

* Update for the face debugger to render landmark 68

* Try to make detect_face_landmark_68() and detect_gender_age() more uniform

* Enable retinaface testing for 320x320

* Make detect_face_landmark_68() and detect_gender_age() as uniform as … (#370)

* Make detect_face_landmark_68() and detect_gender_age() as uniform as possible

* Revert landmark scale and translation

* Make box-mask for lip-syncer adjustable

* Add create_bbox_from_landmark()

* Remove currently unused code

* Feat/uniface (#375)

* add uniface (#373)

* Finalize UniFace implementation

---------

Co-authored-by: Harisreedhar <46858047+harisreedhar@users.noreply.github.com>

* My approach how todo it

* edit

* edit

* replace vertical blur with gaussian

* remove region mask

* Rebase against next and restore method

* Minor improvements

* Minor improvements

* rename & add forehead padding

* Adjust and host uniface model

* Use 2dfan4 model

* Rename to face landmarker

* Feat/replace bbox with bounding box (#380)

* Add landmark 68 to 5 convertion

* Add landmark 68 to 5 convertion

* Keep 5, 5/68 and 68 landmarks

* Replace kps with landmark

* Replace bbox with bounding box

* Reshape face_landmark5_list different

* Make yoloface the default

* Move convert_face_landmark_68_to_5 to face_helper

* Minor spacing issue

* Dynamic detector sizes according to model (#382)

* Dynamic detector sizes according to model

* Dynamic detector sizes according to model

* Undo false commited files

* Add lib syncer model to the UI

* fix halo (#383)

* Bump to 2.3.0

* Update README and wording

* Update README and wording

* Fix spacing

* Apply _vision suffix

* Apply _vision suffix

* Apply _vision suffix

* Apply _vision suffix

* Apply _vision suffix

* Apply _vision suffix

* Apply _vision suffix, Move mouth mask to face_masker.py

* Apply _vision suffix

* Apply _vision suffix

* increase forehead padding

---------

Co-authored-by: tamoharu <133945583+tamoharu@users.noreply.github.com>
Co-authored-by: Harisreedhar <46858047+harisreedhar@users.noreply.github.com>

facefusion/face_masker.py

@@ -7,7 +7,7 @@ import numpy
 import onnxruntime
 
 import facefusion.globals
-from facefusion.typing import Frame, Mask, Padding, FaceMaskRegion, ModelSet
+from facefusion.typing import FaceLandmark68, VisionFrame, Mask, Padding, FaceMaskRegion, ModelSet
 from facefusion.execution_helper import apply_execution_provider_options
 from facefusion.filesystem import resolve_relative_path
 from facefusion.download import conditional_download
@@ -91,7 +91,7 @@ def pre_check() -> bool:
 def create_static_box_mask(crop_size : Size, face_mask_blur : float, face_mask_padding : Padding) -> Mask:
 	blur_amount = int(crop_size[0] * 0.5 * face_mask_blur)
 	blur_area = max(blur_amount // 2, 1)
-	box_mask = numpy.ones(crop_size, numpy.float32)
+	box_mask : Mask = numpy.ones(crop_size, numpy.float32)
 	box_mask[:max(blur_area, int(crop_size[1] * face_mask_padding[0] / 100)), :] = 0
 	box_mask[-max(blur_area, int(crop_size[1] * face_mask_padding[2] / 100)):, :] = 0
 	box_mask[:, :max(blur_area, int(crop_size[0] * face_mask_padding[3] / 100))] = 0
@@ -101,29 +101,40 @@ def create_static_box_mask(crop_size : Size, face_mask_blur : float, face_mask_p
 	return box_mask
 
 
-def create_occlusion_mask(crop_frame : Frame) -> Mask:
+def create_occlusion_mask(crop_vision_frame : VisionFrame) -> Mask:
 	face_occluder = get_face_occluder()
-	prepare_frame = cv2.resize(crop_frame, face_occluder.get_inputs()[0].shape[1:3][::-1])
-	prepare_frame = numpy.expand_dims(prepare_frame, axis = 0).astype(numpy.float32) / 255
-	prepare_frame = prepare_frame.transpose(0, 1, 2, 3)
-	occlusion_mask = face_occluder.run(None,
+	prepare_vision_frame = cv2.resize(crop_vision_frame, face_occluder.get_inputs()[0].shape[1:3][::-1])
+	prepare_vision_frame = numpy.expand_dims(prepare_vision_frame, axis = 0).astype(numpy.float32) / 255
+	prepare_vision_frame = prepare_vision_frame.transpose(0, 1, 2, 3)
+	occlusion_mask : Mask = face_occluder.run(None,
 	{
-		face_occluder.get_inputs()[0].name: prepare_frame
+		face_occluder.get_inputs()[0].name: prepare_vision_frame
 	})[0][0]
 	occlusion_mask = occlusion_mask.transpose(0, 1, 2).clip(0, 1).astype(numpy.float32)
-	occlusion_mask = cv2.resize(occlusion_mask, crop_frame.shape[:2][::-1])
+	occlusion_mask = cv2.resize(occlusion_mask, crop_vision_frame.shape[:2][::-1])
+	occlusion_mask = (cv2.GaussianBlur(occlusion_mask.clip(0, 1), (0, 0), 5).clip(0.5, 1) - 0.5) * 2
 	return occlusion_mask
 
 
-def create_region_mask(crop_frame : Frame, face_mask_regions : List[FaceMaskRegion]) -> Mask:
+def create_region_mask(crop_vision_frame : VisionFrame, face_mask_regions : List[FaceMaskRegion]) -> Mask:
 	face_parser = get_face_parser()
-	prepare_frame = cv2.flip(cv2.resize(crop_frame, (512, 512)), 1)
-	prepare_frame = numpy.expand_dims(prepare_frame, axis = 0).astype(numpy.float32)[:, :, ::-1] / 127.5 - 1
-	prepare_frame = prepare_frame.transpose(0, 3, 1, 2)
-	region_mask = face_parser.run(None,
+	prepare_vision_frame = cv2.flip(cv2.resize(crop_vision_frame, (512, 512)), 1)
+	prepare_vision_frame = numpy.expand_dims(prepare_vision_frame, axis = 0).astype(numpy.float32)[:, :, ::-1] / 127.5 - 1
+	prepare_vision_frame = prepare_vision_frame.transpose(0, 3, 1, 2)
+	region_mask : Mask = face_parser.run(None,
 	{
-		face_parser.get_inputs()[0].name: prepare_frame
+		face_parser.get_inputs()[0].name: prepare_vision_frame
 	})[0][0]
 	region_mask = numpy.isin(region_mask.argmax(0), [ FACE_MASK_REGIONS[region] for region in face_mask_regions ])
-	region_mask = cv2.resize(region_mask.astype(numpy.float32), crop_frame.shape[:2][::-1])
+	region_mask = cv2.resize(region_mask.astype(numpy.float32), crop_vision_frame.shape[:2][::-1])
+	region_mask = (cv2.GaussianBlur(region_mask.clip(0, 1), (0, 0), 5).clip(0.5, 1) - 0.5) * 2
 	return region_mask
+
+
+def create_mouth_mask(face_landmark_68 : FaceLandmark68) -> Mask:
+	convex_hull = cv2.convexHull(face_landmark_68[numpy.r_[3:14, 31:36]].astype(numpy.int32))
+	mouth_mask : Mask = numpy.zeros((512, 512), dtype = numpy.float32)
+	mouth_mask = cv2.fillConvexPoly(mouth_mask, convex_hull, 1.0)
+	mouth_mask = cv2.erode(mouth_mask.clip(0, 1), numpy.ones((21, 3)))
+	mouth_mask = cv2.GaussianBlur(mouth_mask, (0, 0), sigmaX = 1, sigmaY = 15)
+	return mouth_mask