Upload folder using huggingface_hub

Image.py

@@ -0,0 +1,92 @@
+
+import cv2
+import numpy as np
+
+emap = np.load("emap.npy")
+input_std = 255.0
+input_mean = 0.0
+
+def postprocess_face(face_tensor):
+    face_tensor = face_tensor.squeeze().cpu().detach()
+    face_np = (face_tensor.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
+    face_np = cv2.cvtColor(face_np, cv2.COLOR_RGB2BGR)
+
+    return face_np
+
+def getBlob(aimg, input_size = (128, 128)):
+    blob = cv2.dnn.blobFromImage(aimg, 1.0 / input_std, input_size,
+                            (input_mean, input_mean, input_mean), swapRB=True)
+    return blob
+
+def getLatent(source_face):
+    latent = source_face.normed_embedding.reshape((1,-1))
+    latent = np.dot(latent, emap)
+    latent /= np.linalg.norm(latent)
+
+    return latent
+
+def blend_swapped_image(swapped_face, target_image, M):
+    # get image size
+    h, w = target_image.shape[:2]
+    
+    # create inverse affine transform
+    M_inv = cv2.invertAffineTransform(M)
+    
+    # warp swapped face back to target space
+    warped_face = cv2.warpAffine(
+        swapped_face,
+        M_inv,
+        (w, h),
+        borderValue=0.0
+    )
+    
+    # create initial white mask
+    img_white = np.full(
+        (swapped_face.shape[0], swapped_face.shape[1]),
+        255,
+        dtype=np.float32
+    )
+    
+    # warp white mask to target space
+    img_mask = cv2.warpAffine(
+        img_white,
+        M_inv,
+        (w, h),
+        borderValue=0.0
+    )
+    
+    # threshold and refine mask
+    img_mask[img_mask > 20] = 255
+    
+    # calculate mask size for kernel scaling
+    mask_h_inds, mask_w_inds = np.where(img_mask == 255)
+    if len(mask_h_inds) > 0 and len(mask_w_inds) > 0:  # safety check
+        mask_h = np.max(mask_h_inds) - np.min(mask_h_inds)
+        mask_w = np.max(mask_w_inds) - np.min(mask_w_inds)
+        mask_size = int(np.sqrt(mask_h * mask_w))
+        
+        # erode mask
+        k = max(mask_size // 10, 10)
+        kernel = np.ones((k, k), np.uint8)
+        img_mask = cv2.erode(img_mask, kernel, iterations=1)
+        
+        # blur mask
+        k = max(mask_size // 20, 5)
+        kernel_size = (k, k)
+        blur_size = tuple(2 * i + 1 for i in kernel_size)
+        img_mask = cv2.GaussianBlur(img_mask, blur_size, 0)
+    
+    # normalize mask
+    img_mask = img_mask / 255.0
+    img_mask = np.reshape(img_mask, [img_mask.shape[0], img_mask.shape[1], 1])
+    
+    # blend images using mask
+    result = img_mask * warped_face + (1 - img_mask) * target_image.astype(np.float32)
+    result = result.astype(np.uint8)
+    
+    return result
+
+def drawKeypoints(image, keypoints, colorBGR, keypointsRadius=2):
+    for kp in keypoints:
+        x, y = int(kp[0]), int(kp[1])
+        cv2.circle(image, (x, y), radius=keypointsRadius, color=colorBGR, thickness=-1) # BGR format, -1 means filled circle

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

ModelFormat.py

@@ -0,0 +1,53 @@
+import numpy as np
+import onnx
+import torch
+
+from StyleTransferModel_128 import StyleTransferModel
+
+def save_as_onnx_model(torch_model_path, save_emap=True, img_size = 128, originalInswapperClassCompatible = True):
+    output_path = torch_model_path.replace(".pth", ".onnx")
+
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    # Initialize model with the pretrained weights
+    torch_model = StyleTransferModel().to(device)
+    torch_model.load_state_dict(torch.load(torch_model_path, map_location=device), strict=False)
+
+    # set the model to inference mode
+    torch_model.eval()
+    
+    if originalInswapperClassCompatible:
+        dynamic_axes = None
+    else:
+        image_axe = {0: 'batch_size', 1: 'channels', 2: 'height', 3: 'width'}
+        dynamic_axes = {'target': image_axe,    # variable length axes
+                        'source': {0: 'batch_size'},
+                        'output' : image_axe}
+
+    torch.onnx.export(torch_model,               # model being run
+                  {
+                      'target' :torch.randn(1, 3, img_size, img_size, requires_grad=True).to(device), 
+                      'source': torch.randn(1, 512, requires_grad=True).to(device),
+                  },                         # model input (or a tuple for multiple inputs)
+                  output_path,   # where to save the model (can be a file or file-like object)
+                  export_params=True,        # store the trained parameter weights inside the model file
+                  opset_version=11,          # the ONNX version to export the model to
+                  do_constant_folding=True,  # whether to execute constant folding for optimization
+                  input_names = ['target', "source"],   # the model's input names
+                  output_names = ['output'], # the model's output names
+                  dynamic_axes=dynamic_axes)
+
+    model = onnx.load(output_path)
+
+    if save_emap :
+        emap = np.load("emap.npy")
+
+        emap_tensor = onnx.helper.make_tensor(
+            name='emap',
+            data_type=onnx.TensorProto.FLOAT,
+            dims=[512, 512],
+            vals=emap
+        )
+        
+        model.graph.initializer.append(emap_tensor)
+        
+        onnx.save(model, output_path)

README.md

@@ -1,12 +1,130 @@
----
-title: Hg
-emoji: 👁
-colorFrom: pink
-colorTo: gray
-sdk: gradio
-sdk_version: 5.15.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ReSwapper
+
+ReSwapper aims to reproduce the implementation of inswapper. This repository provides code for training, inference, and includes pretrained weights.
+
+Here is the comparesion of the output of Inswapper and Reswapper.
+| Target | Source | Inswapper Output | Reswapper Output<br>(256 resolution)<br>(Step 1399500) | Reswapper Output<br>(Step 1019500) | Reswapper Output<br>(Step 429500) | 
+|--------|--------|--------|--------|--------|--------|
+| ![image](example/1/target.jpg) |![image](example/1/source.jpg) | ![image](example/1/inswapperOutput.jpg) | ![image](example/1/reswapperOutput-1399500_256.jpg) |![image](example/1/reswapperOutput-1019500.jpg) | ![image](example/1/reswapperOutput-429500.jpg) |
+| ![image](example/2/target.jpg) |![image](example/2/source.jpg) | ![image](example/2/inswapperOutput.jpg) | ![image](example/2/reswapperOutput-1399500_256.jpg) | ![image](example/2/reswapperOutput-1019500.jpg) | ![image](example/2/reswapperOutput-429500.jpg) |
+| ![image](example/3/target.jpg) |![image](example/3/source.png) | ![image](example/3/inswapperOutput.jpg) | ![image](example/3/reswapperOutput-1399500_256.jpg) | ![image](example/3/reswapperOutput-1019500.jpg) | ![image](example/3/reswapperOutput-429500.jpg) |
+
+## Installation
+
+```bash
+git clone https://github.com/somanchiu/ReSwapper.git
+cd ReSwapper
+python -m venv venv
+
+venv\scripts\activate
+
+pip install -r requirements.txt
+
+pip install torch torchvision --force --index-url https://download.pytorch.org/whl/cu121
+pip install onnxruntime-gpu --force --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/
+```
+
+## The details of inswapper
+
+### Model architecture
+The inswapper model architecture can be visualized in [Netron](https://netron.app). You can compare with ReSwapper implementation to see architectural similarities. Exporting the model with opset_version=10 makes it easier to compare the graph in Netron. However, it will cause issue #8.
+
+We can also use the following Python code to get more details:
+```python
+model = onnx.load('test.onnx')
+printable_graph=onnx.helper.printable_graph(model.graph)
+```
+
+The model architectures of InSwapper and SimSwap are extremely similar and worth paying attention to.
+
+### Model inputs
+- target: [1, 3, 128, 128] shape image in RGB format with face alignment, normalized to [-1, 1] range
+- source (latent): [1, 512] shape vector, the features of the source face
+    - Calculation of latent, "emap" can be extracted from the original inswapper model.
+        ```python
+        latent = source_face.normed_embedding.reshape((1,-1))
+        latent = np.dot(latent, emap)
+        latent /= np.linalg.norm(latent)
+        ```
+    - It can also be used to calculate the similarity between two faces using cosine similarity.
+
+### Model output
+Model inswapper_128 not only changes facial features, but also body shape.
+
+| Target | Source | Inswapper Output | Reswapper Output<br>(Step 429500) |
+|--------|--------|--------|--------|
+| ![image](example/1/target.jpg) |![image](example/1/source.jpg) | ![image](example/1/inswapperOutput.gif) | ![image](example/1/reswapperOutput.gif) |
+
+### Loss Functions
+There is no information released from insightface. It is an important part of the training. However, there are a lot of articles and papers that can be referenced. By reading a substantial number of articles and papers on face swapping, ID fidelity, and style transfer, you'll frequently encounter the following keywords:
+- content loss
+- style loss/id loss
+- perceptual loss
+
+### Face alignment
+Face alignment is handled incorrectly at resolutions other than 128. To resolve this issue, add an offset to "dst" in both x and y directions in the function "face_align.estimate_norm". The offset is approximately given by the formula: Offset = (128/32768) * Resolution - 0.5
+
+## Training
+### 0. Pretrained weights (Optional)
+If you don't want to train the model from scratch, you can download the pretrained weights and pass model_path into the train function in train.py.
+
+### 1. Dataset Preparation
+Download [FFHQ](https://www.kaggle.com/datasets/arnaud58/flickrfaceshq-dataset-ffhq) to use as target and source images. For the swaped face images, we can use the inswapper output.
+
+### 2. Model Training
+
+Optimizer: Adam
+
+Learning rate: 0.0001
+
+Modify the code in train.py if needed. Then, execute:
+```python
+python train.py
+```
+
+The model will be saved as "reswapper-\<total steps\>.pth". You can also save the model as ONNX using the ModelFormat.save_as_onnx_model function. The ONNX model can then be used with the original INSwapper class.
+
+All losses will be logged into TensorBoard.
+
+Using images with different resolutions simultaneously to train the model will enhance its generalization ability. To apply this strategy, you can pass "resolutions" into the train function.
+
+Generalization ability of the model trained with resolutions of 128 and 256:
+
+| Output<br>resolution | 128 | 160 | 256 |
+|--------|--------|--------|--------|
+|Output| ![image](example/GeneralizationAbility/1399500_128.jpg) |![image](example/GeneralizationAbility/1399500_160.jpg) |![image](example/GeneralizationAbility/1399500_256.jpg) |
+
+Enhancing data diversity will improve output quality, you can pass "enableDataAugmentation" into the train function to perform data augmentation.
+
+| Target | Source | Inswapper Output | Reswapper Output<br>(Step 1567500) | Reswapper Output<br>(Step 1399500) |
+|--------|--------|--------|--------|--------|
+|![image](example/DataAugmentation/target.jpg)| ![image](example/DataAugmentation/source.jpg) |![image](example/DataAugmentation/inswapper_output.jpg) |![image](example/DataAugmentation/reswapper_256Output-1567500.jpg) | ![image](example/DataAugmentation/reswapper_256Output-1399500.jpg) |
+
+#### Notes
+- Do not stop the training too early.
+
+- I'm using an RTX3060 12GB for training. It takes around 12 hours for 50,000 steps.
+- The optimizer may need to be changed to SGD for the final training, as many articles show that SGD can result in lower loss.
+- To get inspiration for improving the model, you might want to review the commented code and unused functions in commit [c2a12e10021ecd1342b9ba50570a16b18f9634b9](https://github.com/somanchiu/ReSwapper/commit/c2a12e10021ecd1342b9ba50570a16b18f9634b9).
+
+## Inference
+```python
+python swap.py
+```
+
+## Pretrained Model
+### 256 Resolution
+- [reswapper_256-1567500.pth](https://huggingface.co/somanchiu/reswapper/tree/main)
+- [reswapper_256-1399500.pth](https://huggingface.co/somanchiu/reswapper/tree/main)
+
+### 128 Resolution
+- [reswapper-1019500.pth](https://huggingface.co/somanchiu/reswapper/tree/main)
+- [reswapper-1019500.onnx](https://huggingface.co/somanchiu/reswapper/tree/main)
+- [reswapper-429500.pth](https://huggingface.co/somanchiu/reswapper/tree/main)
+- [reswapper-429500.onnx](https://huggingface.co/somanchiu/reswapper/tree/main)
+
+### Notes
+If you downloaded the ONNX format model before 2024/11/25, please download the model again or export the model with opset_version=11. This is related to issue #8.
+
+## To Do
+- Create a 512-resolution model (alternative to inswapper_512)

StyleTransferLoss.py

@@ -0,0 +1,52 @@
+import cv2
+import torch
+import torch.nn as nn
+import numpy as np
+from insightface.app import FaceAnalysis
+from pytorch_msssim import ssim
+
+import Image
+
+class StyleTransferLoss(nn.Module):
+    def __init__(self, device='cuda', face_analysis = None):
+        super(StyleTransferLoss, self).__init__()
+        if face_analysis is None:
+            self.face_analysis = FaceAnalysis(providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+            self.face_analysis.prepare(ctx_id=0, det_size=(128, 128))
+        else:
+            self.face_analysis = face_analysis
+        self.device = device
+        self.cosine_similarity = nn.CosineSimilarity(dim=0)
+        
+        # Content loss
+        self.content_loss = nn.MSELoss()
+    
+    def extract_face_latent(self, image):
+        # Convert torch tensor to numpy array
+        face_tensor = image.squeeze().cpu().detach()
+        face_np = (face_tensor.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
+        face_np = cv2.cvtColor(face_np, cv2.COLOR_RGB2BGR)
+
+        # Extract face embedding
+        faces = self.face_analysis.get(face_np)
+        if len(faces) == 0:
+            return None
+        return torch.tensor(Image.getLatent(faces[0])[0]).to(self.device)
+    
+    def forward(self, output_image, target_content):
+        # Content loss
+        # content_loss = self.content_loss(output_image, target_content)
+        content_loss = 1 - ssim(output_image, target_content, data_range=1.0)
+ 
+        output_embedding = self.extract_face_latent(output_image)
+        target_embedding = self.extract_face_latent(target_content)
+
+        identity_loss = None
+        
+        if output_embedding is not None and target_embedding is not None:
+            similarity = self.cosine_similarity(output_embedding, target_embedding)
+
+            identity_loss = 1-((similarity + 1) / 2)
+            identity_loss = identity_loss ** 2 * 10
+
+        return content_loss, identity_loss

StyleTransferModel_128.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class StyleTransferModel(nn.Module):
+    def __init__(self):
+        super(StyleTransferModel, self).__init__()
+
+        # self.pad = nn.ReflectionPad2d(3)
+        # Encoder for target face
+        self.target_encoder = nn.Sequential(
+            # self.pad,
+            nn.Conv2d(3, 128, kernel_size=7, stride=1, padding=0),
+            nn.LeakyReLU(0.2),
+            nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1),
+            nn.LeakyReLU(0.2),
+            nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1),
+            nn.LeakyReLU(0.2),
+            nn.Conv2d(512, 1024, kernel_size=3, stride=2, padding=1),
+            nn.LeakyReLU(0.2),
+        )
+
+        # for style_block in self.target_encoder:
+        #     for param in style_block.parameters():
+        #         param.requires_grad = False
+        
+        # Style blocks
+        self.style_blocks = nn.ModuleList([
+            StyleBlock(1024, 1024, blockIndex) for blockIndex in range(6)
+        ])
+        
+        # Decoder (upsampling)
+        self.decoder = nn.Sequential(
+            nn.Conv2d(1024, 512, kernel_size=3, stride=1, padding=1),
+            nn.LeakyReLU(0.2)
+        )
+
+        self.decoderPart1 = nn.Sequential(
+            nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=1),
+            nn.LeakyReLU(0.2),
+            nn.Conv2d(256, 128, kernel_size=3, stride=1, padding=1),
+            nn.LeakyReLU(0.2)
+        )
+
+        self.decoderPart2 = nn.Sequential(
+            # self.pad,
+            nn.Conv2d(128, 3, kernel_size=7, stride=1, padding=0),
+            nn.Tanh()
+        )
+
+    def forward(self, target, source):
+        # Encode target face
+        target = F.pad(target, pad=(3, 3, 3, 3), mode='reflect')
+
+        target_features = self.target_encoder(target)
+        
+        # Apply style blocks
+        x = target_features
+        for style_block in self.style_blocks:
+            x = style_block(x, source)
+
+        
+        # Decode
+        # x = F.interpolate(x, scale_factor=2, mode='linear')
+        x = F.upsample(
+            x,
+            scale_factor=2,  # specify the desired height and width
+            mode='bilinear',  # 'linear' in 2D is called 'bilinear'
+            align_corners=False  # this is typically False for ONNX compatibility
+        )
+        output = self.decoder(x)
+
+        output = F.upsample(
+            output,
+            scale_factor=2,  # specify the desired height and width
+            mode='bilinear',  # 'linear' in 2D is called 'bilinear'
+            align_corners=False  # this is typically False for ONNX compatibility
+        )
+        output = self.decoderPart1(output)
+
+        output = F.pad(output, pad=(3, 3, 3, 3), mode='reflect')
+
+        output = self.decoderPart2(output)
+        
+        return (output + 1) / 2
+
+class StyleBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, blockIndex):
+        super(StyleBlock, self).__init__()
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=0)
+        self.conv2 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=0)
+        self.style1 = nn.Linear(512, 2048)
+        self.style2 = nn.Linear(512, 2048)
+        self.style = [self.style1, self.style2]
+
+        self.blockIndex = blockIndex
+
+    def normalizeConvRMS(self, conv):
+        x = conv - torch.mean(conv, dim=[2, 3], keepdim=True) # centeredConv
+        squareX = x * x
+        meanSquaredX = torch.mean(squareX, dim=[2, 3], keepdim=True)
+        rms = torch.sqrt(meanSquaredX + 0.00000001)
+        return (1 / rms) * x
+
+    def forward(self, residual, style):
+        # print(f'Forward: {self.blockIndex}')
+        style1024 = []
+        for index in range(2):
+            style1 = self.style[index](style)
+            style1 = torch.unsqueeze(style1, 2)
+            style1 = torch.unsqueeze(style1, 3)
+            first_half = style1[:, :1024, :, :]
+            second_half = style1[:, 1024:, :, :]
+
+            style1024.append([first_half, second_half])
+
+        conv1 = self.normalizeConvRMS(self.conv1(F.pad(residual, pad=(1, 1, 1, 1), mode='reflect')))
+
+        out = F.relu(conv1 * style1024[0][0] + style1024[0][1])
+
+        out = F.pad(out, pad=(1, 1, 1, 1), mode='reflect')
+
+        conv2 = self.normalizeConvRMS(self.conv2(out))
+        out = conv2 * style1024[1][0] + style1024[1][1]
+
+        return residual + out

emap.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cee626bc81721d71c5d6cb1f76f830b9ae46f595514b0884dd8ae34785576764
+size 1048704

face_align.py

@@ -0,0 +1,109 @@
+import cv2
+import numpy as np
+from skimage import transform as trans
+
+
+arcface_dst = np.array(
+    [[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366],
+     [41.5493, 92.3655], [70.7299, 92.2041]],
+    dtype=np.float32)
+
+def estimate_norm(lmk, image_size=112,mode='arcface'):
+    # assert lmk.shape == (5, 2)
+    # assert image_size%112==0 or image_size%128==0
+    if image_size%112==0:
+        ratio = float(image_size)/112.0
+        diff_x = 0
+    else:
+        ratio = float(image_size)/128.0
+        diff_x = 8.0*ratio
+    dst = arcface_dst * ratio
+    dst[:,0] += diff_x
+
+    if image_size != 128:
+        offset = (128/32768)*image_size-0.5
+        dst[:,0] += offset
+        dst[:,1] += offset
+
+    tform = trans.SimilarityTransform()
+    tform.estimate(lmk, dst)
+    M = tform.params[0:2, :]
+    return M
+
+def norm_crop(img, landmark, image_size=112, mode='arcface'):
+    M = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped
+
+def norm_crop2(img, landmark, image_size=112, mode='arcface'):
+    M = estimate_norm(landmark, image_size, mode)
+    warped = cv2.warpAffine(img, M, (image_size, image_size), borderValue=0.0)
+    return warped, M
+
+def square_crop(im, S):
+    if im.shape[0] > im.shape[1]:
+        height = S
+        width = int(float(im.shape[1]) / im.shape[0] * S)
+        scale = float(S) / im.shape[0]
+    else:
+        width = S
+        height = int(float(im.shape[0]) / im.shape[1] * S)
+        scale = float(S) / im.shape[1]
+    resized_im = cv2.resize(im, (width, height))
+    det_im = np.zeros((S, S, 3), dtype=np.uint8)
+    det_im[:resized_im.shape[0], :resized_im.shape[1], :] = resized_im
+    return det_im, scale
+
+
+def transform(data, center, output_size, scale, rotation):
+    scale_ratio = scale
+    rot = float(rotation) * np.pi / 180.0
+    #translation = (output_size/2-center[0]*scale_ratio, output_size/2-center[1]*scale_ratio)
+    t1 = trans.SimilarityTransform(scale=scale_ratio)
+    cx = center[0] * scale_ratio
+    cy = center[1] * scale_ratio
+    t2 = trans.SimilarityTransform(translation=(-1 * cx, -1 * cy))
+    t3 = trans.SimilarityTransform(rotation=rot)
+    t4 = trans.SimilarityTransform(translation=(output_size / 2,
+                                                output_size / 2))
+    t = t1 + t2 + t3 + t4
+    M = t.params[0:2]
+    cropped = cv2.warpAffine(data,
+                             M, (output_size, output_size),
+                             borderValue=0.0)
+    return cropped, M
+
+
+def trans_points2d(pts, M):
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i] = new_pt[0:2]
+
+    return new_pts
+
+
+def trans_points3d(pts, M):
+    scale = np.sqrt(M[0][0] * M[0][0] + M[0][1] * M[0][1])
+    #print(scale)
+    new_pts = np.zeros(shape=pts.shape, dtype=np.float32)
+    for i in range(pts.shape[0]):
+        pt = pts[i]
+        new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32)
+        new_pt = np.dot(M, new_pt)
+        #print('new_pt', new_pt.shape, new_pt)
+        new_pts[i][0:2] = new_pt[0:2]
+        new_pts[i][2] = pts[i][2] * scale
+
+    return new_pts
+
+
+def trans_points(pts, M):
+    if pts.shape[1] == 2:
+        return trans_points2d(pts, M)
+    else:
+        return trans_points3d(pts, M)
+

requirements-colab.txt

@@ -0,0 +1,85 @@
+absl-py==2.1.0
+addict==2.4.0
+albucore==0.0.17
+albumentations==1.4.17
+annotated-types==0.7.0
+certifi==2024.8.30
+chardet==3.0.4
+charset-normalizer==3.3.2
+colorama==0.4.6
+coloredlogs==15.0.1
+contourpy==1.3.0
+cycler==0.12.1
+Cython==3.0.11
+easydict==1.13
+einops==0.8.0
+eval_type_backport==0.2.0
+facexlib==0.3.0
+filelock==3.13.1
+filterpy==1.4.5
+flatbuffers==24.3.25
+fonttools==4.54.1
+fsspec==2024.2.0
+ftfy==6.2.3
+future==1.0.0
+grpcio==1.66.1
+huggingface-hub==0.25.0
+humanfriendly==10.0
+idna==3.10
+imageio==2.35.1
+importlib_metadata==8.5.0
+insightface==0.7.3
+Jinja2==3.1.3
+joblib==1.4.2
+kiwisolver==1.4.7
+lazy_loader==0.4
+llvmlite==0.43.0
+lmdb==1.5.1
+Markdown==3.7
+MarkupSafe==2.1.5
+matplotlib==3.9.2
+mpmath==1.3.0
+networkx==3.3
+numba==0.60.0
+numpy==1.26.4
+onnx==1.17.0
+onnxruntime==1.18.1
+onnxruntime-gpu==1.19.2
+opencv-python==4.10.0.84
+opencv-python-headless==4.10.0.84
+packaging==24.1
+pillow==10.4.0
+platformdirs==4.3.6
+prettytable==3.11.0
+protobuf==5.28.2
+pydantic==2.9.2
+pydantic_core==2.23.4
+pyparsing==3.1.4
+pyreadline3==3.4.1
+python-dateutil==2.9.0.post0
+pytorch-msssim==1.0.0
+PyYAML==6.0.2
+regex==2024.9.11
+requests==2.32.3
+safetensors==0.4.5
+scikit-image==0.24.0
+scikit-learn==1.5.2
+scipy==1.14.1
+six==1.16.0
+sympy==1.13.2
+tensorboard==2.17.1
+tensorboard-data-server==0.7.2
+threadpoolctl==3.5.0
+tifffile==2024.9.20
+timm==1.0.9
+tokenizers==0.15.2
+tomli==2.0.1
+torch==2.4.1
+tqdm==4.66.5
+transformers==4.36.2
+typing_extensions==4.12.2
+urllib3==2.2.3
+wcwidth==0.2.13
+Werkzeug==3.0.4
+yapf==0.40.2
+zipp==3.20.2

requirements.txt

@@ -0,0 +1,86 @@
+absl-py==2.1.0
+addict==2.4.0
+albucore==0.0.17
+albumentations==1.4.17
+annotated-types==0.7.0
+certifi==2024.8.30
+chardet==3.0.4
+charset-normalizer==3.3.2
+colorama==0.4.6
+coloredlogs==15.0.1
+contourpy==1.3.0
+cycler==0.12.1
+Cython==3.0.11
+easydict==1.13
+einops==0.8.0
+eval_type_backport==0.2.0
+facexlib==0.3.0
+filelock==3.13.1
+filterpy==1.4.5
+flatbuffers==24.3.25
+fonttools==4.54.1
+fsspec==2024.2.0
+ftfy==6.2.3
+future==1.0.0
+grpcio==1.66.1
+huggingface-hub==0.25.0
+humanfriendly==10.0
+idna==3.10
+imageio==2.35.1
+importlib_metadata==8.5.0
+insightface==0.7.3
+Jinja2==3.1.3
+joblib==1.4.2
+kiwisolver==1.4.7
+lazy_loader==0.4
+llvmlite==0.43.0
+lmdb==1.5.1
+Markdown==3.7
+MarkupSafe==2.1.5
+matplotlib==3.9.2
+mpmath==1.3.0
+networkx==3.3
+numba==0.60.0
+numpy==1.26.4
+onnx==1.17.0
+onnxruntime==1.18.1
+onnxruntime-gpu==1.19.2
+opencv-python==4.10.0.84
+opencv-python-headless==4.10.0.84
+packaging==24.1
+pillow==10.4.0
+platformdirs==4.3.6
+prettytable==3.11.0
+protobuf==5.28.2
+pydantic==2.9.2
+pydantic_core==2.23.4
+pyparsing==3.1.4
+pyreadline3==3.4.1
+python-dateutil==2.9.0.post0
+pytorch-msssim==1.0.0
+PyYAML==6.0.2
+regex==2024.9.11
+requests==2.32.3
+safetensors==0.4.5
+scikit-image==0.24.0
+scikit-learn==1.5.2
+scipy==1.14.1
+six==1.16.0
+sympy==1.13.2
+tensorboard==2.17.1
+tensorboard-data-server==0.7.2
+threadpoolctl==3.5.0
+tifffile==2024.9.20
+timm==1.0.9
+tokenizers==0.15.2
+tomli==2.0.1
+torch==2.4.1+cu121
+torchvision==0.19.1+cu121
+tqdm==4.66.5
+transformers==4.36.2
+typing_extensions==4.12.2
+urllib3==2.2.3
+wcwidth==0.2.13
+Werkzeug==3.0.4
+yapf==0.40.2
+zipp==3.20.2

swap.py

@@ -0,0 +1,93 @@
+import argparse
+import os
+
+import cv2
+import torch
+import Image
+from insightface.app import FaceAnalysis
+import face_align
+
+faceAnalysis = FaceAnalysis(name='buffalo_l')
+faceAnalysis.prepare(ctx_id=0, det_size=(512, 512))
+
+from StyleTransferModel_128 import StyleTransferModel
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description='Process command line arguments')
+    
+    parser.add_argument('--target', required=True, help='Target path')
+    parser.add_argument('--source', required=True, help='Source path')
+    parser.add_argument('--outputPath', required=True, help='Output path')
+    parser.add_argument('--modelPath', required=True, help='Model path')
+    parser.add_argument('--no-paste-back', action='store_true', help='Disable pasting back the swapped face onto the original image')
+    parser.add_argument('--resolution', type=int, default=128, help='Resolution')
+
+    return parser.parse_args()
+
+def get_device():
+    return torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+def load_model(model_path):
+    device = get_device()
+    model = StyleTransferModel().to(device)
+    model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
+    model.eval()
+    return model
+
+def swap_face(model, target_face, source_face_latent):
+    device = get_device()
+
+    target_tensor = torch.from_numpy(target_face).to(device)
+    source_tensor = torch.from_numpy(source_face_latent).to(device)
+
+    with torch.no_grad():
+        swapped_tensor = model(target_tensor, source_tensor)
+    
+    swapped_face = Image.postprocess_face(swapped_tensor)
+    
+    return swapped_face, swapped_tensor
+
+def create_target(target_image, resolution):
+    if isinstance(target_image, str):
+        target_image = cv2.imread(target_image)
+
+    target_face = faceAnalysis.get(target_image)[0]
+    aligned_target_face, M = face_align.norm_crop2(target_image, target_face.kps, resolution)
+    target_face_blob = Image.getBlob(aligned_target_face, (resolution, resolution))
+
+    return target_face_blob, M
+
+def create_source(source_img_path):
+    source_image = cv2.imread(source_img_path)
+
+    source_face = faceAnalysis.get(source_image)[0]
+
+    source_latent = Image.getLatent(source_face)
+
+    return source_latent
+
+def main():
+    args = parse_arguments()
+    
+    # Access the arguments
+    target_image_path = args.target
+    source = args.source
+    output_path = args.outputPath
+    model_path = args.modelPath
+
+    model = load_model(model_path)
+
+    target_img = cv2.imread(target_image_path)
+    target_face_blob, M = create_target(target_img, args.resolution)
+    source_latent = create_source(source)
+    swapped_face, _ = swap_face(model, target_face_blob, source_latent)
+
+    if not args.no_paste_back:
+        swapped_face = Image.blend_swapped_image(swapped_face, target_img, M)
+
+    output_folder = os.path.dirname(output_path)
+    os.makedirs(output_folder, exist_ok=True)
+    cv2.imwrite(output_path, swapped_face)
+
+if __name__ == "__main__":
+    main()

train.py

@@ -0,0 +1,247 @@
+from datetime import datetime
+import os
+import random
+import torch
+import torch.optim as optim
+import torch.nn.functional as F
+
+import Image
+import ModelFormat
+from StyleTransferLoss import StyleTransferLoss
+import onnxruntime as rt
+
+import cv2
+from insightface.data import get_image as ins_get_image
+from insightface.app import FaceAnalysis
+import face_align
+
+from StyleTransferModel_128 import StyleTransferModel
+from torch.utils.tensorboard import SummaryWriter
+
+inswapper_128_path = 'inswapper_128.onnx'
+img_size = 128
+
+providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
+
+inswapperInferenceSession = rt.InferenceSession(inswapper_128_path, providers=providers)
+
+faceAnalysis = FaceAnalysis(name='buffalo_l')
+faceAnalysis.prepare(ctx_id=0, det_size=(512, 512))
+
+def get_device():
+    return torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+style_loss_fn = StyleTransferLoss().to(get_device())
+
+def train(datasetDir, learning_rate=0.0001, model_path=None, outputModelFolder='', saveModelEachSteps = 1, stopAtSteps=None, logDir=None, previewDir=None, saveAs_onnx = False, resolutions = [128], enableDataAugmentation = False):
+    device = get_device()
+    print(f"Using device: {device}")
+
+    model = StyleTransferModel().to(device)
+
+    if model_path is not None:
+        model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
+        print(f"Loaded model from {model_path}")
+
+        lastSteps = int(model_path.split('-')[-1].split('.')[0])
+        print(f"Resuming training from step {lastSteps}")
+    else:
+        lastSteps = 0
+
+    model.train()
+    model = model.to(device)
+
+    # Initialize optimizer
+    optimizer = optim.Adam(model.parameters(), lr=learning_rate)
+
+    # Initialize TensorBoard writer
+    if logDir is not None:
+        train_writer = SummaryWriter(os.path.join(logDir, "training"))
+        val_writer = SummaryWriter(os.path.join(logDir, "validation"))
+
+    steps = 0
+
+    image = os.listdir(datasetDir)
+
+    resolutionIndex = 0
+
+    # Training loop
+    while True:
+        start_time = datetime.now()
+        
+        resolution = resolutions[resolutionIndex%len(resolutions)]
+
+        targetFaceIndex = random.randint(0, len(image)-1)
+        sourceFaceIndex = random.randint(0, len(image)-1)
+
+        target_img=cv2.imread(f"{datasetDir}/{image[targetFaceIndex]}")
+        if enableDataAugmentation and steps % 2 == 0:
+            target_img = cv2.cvtColor(target_img, cv2.COLOR_BGR2GRAY)
+            target_img = cv2.cvtColor(target_img, cv2.COLOR_GRAY2BGR)
+        faces = faceAnalysis.get(target_img)
+
+        if targetFaceIndex != sourceFaceIndex:
+            source_img = cv2.imread(f"{datasetDir}/{image[sourceFaceIndex]}")
+            faces2 = faceAnalysis.get(source_img)
+        else:
+            faces2 = faces
+
+        if len(faces) > 0 and len(faces2) > 0:
+            new_aligned_face, _ = face_align.norm_crop2(target_img, faces[0].kps, img_size)
+            blob = Image.getBlob(new_aligned_face)
+            latent = Image.getLatent(faces2[0])
+        else:
+            continue
+
+        if targetFaceIndex != sourceFaceIndex:
+            input = {inswapperInferenceSession.get_inputs()[0].name: blob,
+                    inswapperInferenceSession.get_inputs()[1].name: latent}
+
+            expected_output = inswapperInferenceSession.run([inswapperInferenceSession.get_outputs()[0].name], input)[0]
+        else:
+            expected_output = blob
+
+        expected_output_tensor = torch.from_numpy(expected_output).to(device)
+
+        if resolution != 128:
+            new_aligned_face, _ = face_align.norm_crop2(target_img, faces[0].kps, resolution)
+            blob = Image.getBlob(new_aligned_face, (resolution, resolution))
+
+        latent_tensor = torch.from_numpy(latent).to(device)
+        target_input_tensor = torch.from_numpy(blob).to(device)
+
+        optimizer.zero_grad()
+        output = model(target_input_tensor, latent_tensor)
+
+        if (resolution != 128):
+            output = F.interpolate(output, size=(128, 128), mode='bilinear', align_corners=False)
+
+        content_loss, identity_loss = style_loss_fn(output, expected_output_tensor)
+
+        loss = content_loss
+
+        if identity_loss is not None:
+            loss +=identity_loss
+        
+        loss.backward()
+
+        optimizer.step()
+
+        steps += 1
+        totalSteps = steps + lastSteps
+
+        if logDir is not None:
+            train_writer.add_scalar("Loss/total", loss.item(), totalSteps)
+            train_writer.add_scalar("Loss/content_loss", content_loss.item(), totalSteps)
+
+            if identity_loss is not None:
+                train_writer.add_scalar("Loss/identity_loss", identity_loss.item(), totalSteps)
+
+        elapsed_time = datetime.now() - start_time
+
+        print(f"Total Steps: {totalSteps}, Step: {steps}, Loss: {loss.item():.4f}, Elapsed time: {elapsed_time}")
+
+        if steps % saveModelEachSteps == 0:
+            outputModelPath = f"reswapper-{totalSteps}.pth"
+            if outputModelFolder != '':
+                outputModelPath = f"{outputModelFolder}/{outputModelPath}"
+            saveModel(model, outputModelPath)
+
+            validation_total_loss, validation_content_loss, validation_identity_loss, swapped_face, swapped_face_256 = validate(outputModelPath)
+            if previewDir is not None:
+                cv2.imwrite(f"{previewDir}/{totalSteps}.jpg", swapped_face)
+                cv2.imwrite(f"{previewDir}/{totalSteps}_256.jpg", swapped_face_256)
+
+            if logDir is not None:
+                val_writer.add_scalar("Loss/total", validation_total_loss.item(), totalSteps)
+                val_writer.add_scalar("Loss/content_loss", validation_content_loss.item(), totalSteps)
+                if validation_identity_loss is not None:
+                    val_writer.add_scalar("Loss/identity_loss", validation_identity_loss.item(), totalSteps)
+
+            if saveAs_onnx :
+                ModelFormat.save_as_onnx_model(outputModelPath)
+
+        if stopAtSteps is not None and steps == stopAtSteps:
+            exit()
+
+        resolutionIndex += 1
+
+def saveModel(model, outputModelPath):
+    torch.save(model.state_dict(), outputModelPath)
+
+def load_model(model_path):
+    device = get_device()
+    model = StyleTransferModel().to(device)
+    model.load_state_dict(torch.load(model_path, map_location=device), strict=False)
+
+    model.eval()
+    return model
+
+def swap_face(model, target_face, source_face_latent):
+    device = get_device()
+
+    target_tensor = torch.from_numpy(target_face).to(device)
+    source_tensor = torch.from_numpy(source_face_latent).to(device)
+
+    with torch.no_grad():
+        swapped_tensor = model(target_tensor, source_tensor)
+
+    swapped_face = Image.postprocess_face(swapped_tensor)
+    
+    return swapped_face, swapped_tensor
+
+# test image
+test_img = ins_get_image('t1')
+
+test_faces = faceAnalysis.get(test_img)
+test_faces = sorted(test_faces, key = lambda x : x.bbox[0])
+test_target_face, _ = face_align.norm_crop2(test_img, test_faces[0].kps, img_size)
+test_target_face = Image.getBlob(test_target_face)
+test_l = Image.getLatent(test_faces[2])
+
+test_target_face_256, _ = face_align.norm_crop2(test_img, test_faces[0].kps, 256)
+test_target_face_256 = Image.getBlob(test_target_face_256, (256, 256))
+
+test_input = {inswapperInferenceSession.get_inputs()[0].name: test_target_face,
+        inswapperInferenceSession.get_inputs()[1].name: test_l}
+
+test_inswapperOutput = inswapperInferenceSession.run([inswapperInferenceSession.get_outputs()[0].name], test_input)[0]
+
+def validate(modelPath):
+    model = load_model(modelPath)
+    swapped_face, swapped_tensor= swap_face(model, test_target_face, test_l)
+    swapped_face_256, _= swap_face(model, test_target_face_256, test_l)
+
+    validation_content_loss, validation_identity_loss = style_loss_fn(swapped_tensor, torch.from_numpy(test_inswapperOutput).to(get_device()))
+
+    validation_total_loss = validation_content_loss
+    if validation_identity_loss is not None:
+        validation_total_loss += validation_identity_loss
+
+    return validation_total_loss, validation_content_loss, validation_identity_loss, swapped_face, swapped_face_256
+
+def main():
+    outputModelFolder = "model"
+    modelPath = None
+    # modelPath = f"{outputModelFolder}/reswapper-<step>.pth"
+
+    logDir = "training/log"
+    previewDir = "training/preview"
+    datasetDir = "FFHQ"
+
+    os.makedirs(outputModelFolder, exist_ok=True)
+    os.makedirs(previewDir, exist_ok=True)
+
+    train(
+        datasetDir=datasetDir,
+        model_path=modelPath,
+        learning_rate=0.0001,
+        # resolutions = [128, 256],
+        # enableDataAugmentation=True,
+        outputModelFolder=outputModelFolder,
+        saveModelEachSteps = 1000,
+        stopAtSteps = 70000,
+        logDir=f"{logDir}/{datetime.now().strftime('%Y%m%d %H%M%S')}",
+        previewDir=previewDir)
+                    
+if __name__ == "__main__":
+    main()