import gradio from huggingface_hub import Repository import os from utils.utils import norm_crop, estimate_norm, inverse_estimate_norm, transform_landmark_points, get_lm from networks.layers import AdaIN, AdaptiveAttention from tensorflow_addons.layers import InstanceNormalization import numpy as np import cv2 from scipy.ndimage import gaussian_filter from tensorflow.keras.models import load_model from options.swap_options import SwapOptions # . token = os.environ['model_fetch'] opt = SwapOptions().parse() retina_repo = Repository(local_dir="retina_model", clone_from="felixrosberg/retinaface_resnet50", private=True, use_auth_token=token, git_user="felixrosberg") from retina_model.models import * RetinaFace = load_model("retina_model/retinaface_res50.h5", custom_objects={"FPN": FPN, "SSH": SSH, "BboxHead": BboxHead, "LandmarkHead": LandmarkHead, "ClassHead": ClassHead}) arc_repo = Repository(local_dir="arcface_model", clone_from="felixrosberg/arcface_tf", private=True, use_auth_token=token) ArcFace = load_model("arcface_model/arc_res50.h5") ArcFaceE = load_model("arcface_model/arc_res50e.h5") g_repo = Repository(local_dir="g_model_c_hq", clone_from="felixrosberg/affa_config_c_hq", private=True, use_auth_token=token) G = load_model("g_model_c_hq/generator_t_28.h5", custom_objects={"AdaIN": AdaIN, "AdaptiveAttention": AdaptiveAttention, "InstanceNormalization": InstanceNormalization}) r_repo = Repository(local_dir="reconstruction_attack", clone_from="felixrosberg/reconstruction_attack", private=True, use_auth_token=token) R = load_model("reconstruction_attack/reconstructor_42.h5", custom_objects={"AdaIN": AdaIN, "AdaptiveAttention": AdaptiveAttention, "InstanceNormalization": InstanceNormalization}) permuter_repo = Repository(local_dir="identity_permuter", clone_from="felixrosberg/identitypermuter", private=True, use_auth_token=token, git_user="felixrosberg") from identity_permuter.id_permuter import identity_permuter IDP = identity_permuter(emb_size=32, min_arg=False) IDP.load_weights("identity_permuter/id_permuter.h5") blend_mask_base = np.zeros(shape=(256, 256, 1)) blend_mask_base[80:244, 32:224] = 1 blend_mask_base = gaussian_filter(blend_mask_base, sigma=7) def run_inference(target, source, slider, adv_slider, settings): try: source = np.array(source) target = np.array(target) # Prepare to load video if "anonymize" not in settings: source_a = RetinaFace(np.expand_dims(source, axis=0)).numpy()[0] source_h, source_w, _ = source.shape source_lm = get_lm(source_a, source_w, source_h) source_aligned = norm_crop(source, source_lm, image_size=256) source_z = ArcFace.predict(np.expand_dims(tf.image.resize(source_aligned, [112, 112]) / 255.0, axis=0)) else: source_z = None # read frame im = target im_h, im_w, _ = im.shape im_shape = (im_w, im_h) detection_scale = im_w // 640 if im_w > 640 else 1 faces = RetinaFace(np.expand_dims(cv2.resize(im, (im_w // detection_scale, im_h // detection_scale)), axis=0)).numpy() total_img = im / 255.0 for annotation in faces: lm_align = np.array([[annotation[4] * im_w, annotation[5] * im_h], [annotation[6] * im_w, annotation[7] * im_h], [annotation[8] * im_w, annotation[9] * im_h], [annotation[10] * im_w, annotation[11] * im_h], [annotation[12] * im_w, annotation[13] * im_h]], dtype=np.float32) # align the detected face M, pose_index = estimate_norm(lm_align, 256, "arcface", shrink_factor=1.0) im_aligned = (cv2.warpAffine(im, M, (256, 256), borderValue=0.0) - 127.5) / 127.5 if "adversarial defense" in settings: eps = adv_slider / 200 X = tf.convert_to_tensor(np.expand_dims(im_aligned, axis=0)) with tf.GradientTape() as tape: tape.watch(X) X_z = ArcFaceE(tf.image.resize(X * 0.5 + 0.5, [112, 112])) output = R([X, X_z]) loss = tf.reduce_mean(tf.abs(0 - output)) gradient = tf.sign(tape.gradient(loss, X)) adv_x = X + eps * gradient im_aligned = tf.clip_by_value(adv_x, -1, 1)[0] if "anonymize" in settings and "reconstruction attack" not in settings: """source_z = ArcFace.predict(np.expand_dims(tf.image.resize(im_aligned, [112, 112]) / 255.0, axis=0)) anon_ratio = int(512 * (slider / 100)) anon_vector = np.ones(shape=(1, 512)) anon_vector[:, :anon_ratio] = -1 np.random.shuffle(anon_vector) source_z *= anon_vector""" slider_weight = slider / 100 target_z = ArcFace.predict(np.expand_dims(tf.image.resize(im_aligned, [112, 112]) * 0.5 + 0.5, axis=0)) source_z = IDP.predict(target_z) source_z = slider_weight * source_z + (1 - slider_weight) * target_z if "reconstruction attack" in settings: source_z = ArcFaceE.predict(np.expand_dims(tf.image.resize(im_aligned, [112, 112]) * 0.5 + 0.5, axis=0)) # face swap if "reconstruction attack" not in settings: changed_face_cage = G.predict([np.expand_dims(im_aligned, axis=0), source_z]) changed_face = changed_face_cage[0] * 0.5 + 0.5 # get inverse transformation landmarks transformed_lmk = transform_landmark_points(M, lm_align) # warp image back iM, _ = inverse_estimate_norm(lm_align, transformed_lmk, 256, "arcface", shrink_factor=1.0) iim_aligned = cv2.warpAffine(changed_face, iM, im_shape, borderValue=0.0) # blend swapped face with target image blend_mask = cv2.warpAffine(blend_mask_base, iM, im_shape, borderValue=0.0) blend_mask = np.expand_dims(blend_mask, axis=-1) total_img = (iim_aligned * blend_mask + total_img * (1 - blend_mask)) else: changed_face_cage = R.predict([np.expand_dims(im_aligned, axis=0), source_z]) changed_face = changed_face_cage[0] * 0.5 + 0.5 # get inverse transformation landmarks transformed_lmk = transform_landmark_points(M, lm_align) # warp image back iM, _ = inverse_estimate_norm(lm_align, transformed_lmk, 256, "arcface", shrink_factor=1.0) iim_aligned = cv2.warpAffine(changed_face, iM, im_shape, borderValue=0.0) # blend swapped face with target image blend_mask = cv2.warpAffine(blend_mask_base, iM, im_shape, borderValue=0.0) blend_mask = np.expand_dims(blend_mask, axis=-1) total_img = (iim_aligned * blend_mask + total_img * (1 - blend_mask)) if "compare" in settings: total_img = np.concatenate((im / 255.0, total_img), axis=1) total_img = np.clip(total_img, 0, 1) total_img *= 255.0 total_img = total_img.astype('uint8') return total_img except Exception as e: print(e) return None description = "Performs subject agnostic identity transfer from a source face to all target faces. \n\n" \ "Implementation and demo of FaceDancer, accepted to WACV 2023. \n\n" \ "Pre-print: https://arxiv.org/abs/2210.10473 \n\n" \ "Code: https://github.com/felixrosberg/FaceDancer \n\n" \ "\n\n" \ "Options:\n\n" \ "-Compare returns the target image concatenated with the results.\n\n" \ "-Anonymize will ignore the source image and perform an identity permutation of target faces.\n\n" \ "-Reconstruction attack will attempt to invert the face swap or the anonymization.\n\n" \ "-Adversarial defense will add a permutation noise that disrupts the reconstruction attack.\n\n" \ "NOTE: There is no guarantees with the anonymization process currently.\n\n" \ "NOTE: source image with too high resolution may not work properly!" examples = [["assets/rick.jpg", "assets/musk.jpg", 100, 10, ["compare"]], ["assets/musk.jpg", "assets/musk.jpg", 100, 10, ["anonymize"]]] article = """ Demo is based of recent research from my Ph.D work. Results expects to be published in the coming months. """ iface = gradio.Interface(run_inference, [gradio.inputs.Image(shape=None, label='Target'), gradio.inputs.Image(shape=None, label='Source'), gradio.inputs.Slider(0, 100, default=100, label="Anonymization ratio (%)"), gradio.inputs.Slider(0, 100, default=100, label="Adversarial defense ratio (%)"), gradio.inputs.CheckboxGroup(["compare", "anonymize", "reconstruction attack", "adversarial defense"], label='Options')], gradio.outputs.Image(), title="Face Swap", description=description, examples=examples, article=article, layout="vertical") iface.launch()