import random import gradio as gr import imageio import numpy as np import onnx import onnxruntime as rt import huggingface_hub from numpy.random import RandomState from skimage import transform def get_inter(r1, r2): h_inter = max(min(r1[3], r2[3]) - max(r1[1], r2[1]), 0) w_inter = max(min(r1[2], r2[2]) - max(r1[0], r2[0]), 0) return h_inter * w_inter def iou(r1, r2): s1 = (r1[2] - r1[0]) * (r1[3] - r1[1]) s2 = (r2[2] - r2[0]) * (r2[3] - r2[1]) i = get_inter(r1, r2) return i / (s1 + s2 - i) def letterbox(im, new_shape=(640, 640), color=(0.5, 0.5, 0.5), stride=32): # Resize and pad image while meeting stride-multiple constraints shape = im.shape[:2] # current shape [height, width] # Scale ratio (new / old) r = min(new_shape[0] / shape[0], new_shape[1] / shape[1]) # Compute padding new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding dw /= 2 # divide padding into 2 sides dh /= 2 if shape != new_unpad: # resize im = transform.resize(im, (new_unpad[1], new_unpad[0])) top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1)) left, right = int(round(dw - 0.1)), int(round(dw + 0.1)) im_new = np.full((new_unpad[1] + top + bottom, new_unpad[0] + left + right, 3), color, dtype=np.float32) im_new[top:new_unpad[1] + top, left:new_unpad[0] + left] = im return im_new def nms(pred, conf_thres, iou_thres, max_instance=20): # pred (anchor_num, 5 + cls_num) nc = pred.shape[1] - 5 candidates = [list() for x in range(nc)] for x in pred: if x[4] < conf_thres: continue cls = np.argmax(x[5:]) p = x[4] * x[5 + cls] if conf_thres <= p: box = (x[0] - x[2] / 2, x[1] - x[3] / 2, x[0] + x[2] / 2, x[1] + x[3] / 2) # xywh2xyxy candidates[cls].append([p, box]) result = [list() for x in range(nc)] for i, candidate in enumerate(candidates): candidate = sorted(candidate, key=lambda a: a[0], reverse=True) candidate = candidate[:max_instance] for x in candidate: ok = True for r in result[i]: if iou(r[1], x[1]) > iou_thres: ok = False break if ok: result[i].append(x) return result class Model: def __init__(self): self.detector = None self.encoder = None self.g_synthesis = None self.g_mapping = None self.detector_stride = None self.detector_imgsz = None self.detector_class_names = None self.anime_seg = None self.w_avg = None self.load_models() def load_models(self): g_mapping_path = huggingface_hub.hf_hub_download("skytnt/fbanime-gan", "g_mapping.onnx") g_synthesis_path = huggingface_hub.hf_hub_download("skytnt/fbanime-gan", "g_synthesis.onnx") encoder_path = huggingface_hub.hf_hub_download("skytnt/fbanime-gan", "encoder.onnx") detector_path = huggingface_hub.hf_hub_download("skytnt/fbanime-gan", "waifu_dect.onnx") anime_seg_path = huggingface_hub.hf_hub_download("skytnt/anime-seg", "isnetis.onnx") providers = ['CPUExecutionProvider'] gpu_providers = ['CUDAExecutionProvider'] g_mapping = onnx.load(g_mapping_path) w_avg = [x for x in g_mapping.graph.initializer if x.name == "w_avg"][0] w_avg = np.frombuffer(w_avg.raw_data, dtype=np.float32)[np.newaxis, :] w_avg = w_avg.repeat(16, axis=0)[np.newaxis, :] self.w_avg = w_avg self.g_mapping = rt.InferenceSession(g_mapping_path, providers=gpu_providers + providers) self.g_synthesis = rt.InferenceSession(g_synthesis_path, providers=gpu_providers + providers) self.encoder = rt.InferenceSession(encoder_path, providers=providers) self.detector = rt.InferenceSession(detector_path, providers=providers) detector_meta = self.detector.get_modelmeta().custom_metadata_map self.detector_stride = int(detector_meta['stride']) self.detector_imgsz = 1088 self.detector_class_names = eval(detector_meta['names']) self.anime_seg = rt.InferenceSession(anime_seg_path, providers=providers) def get_img(self, w, noise=0): img = self.g_synthesis.run(None, {'w': w, "noise": np.asarray([noise], dtype=np.float32)})[0] return (img.transpose(0, 2, 3, 1) * 127.5 + 128).clip(0, 255).astype(np.uint8)[0] def get_w(self, z, psi1, psi2): return self.g_mapping.run(None, {'z': z, 'psi': np.asarray([psi1, psi2], dtype=np.float32)})[0] def remove_bg(self, img, s=1024): img0 = img img = (img / 255).astype(np.float32) h, w = h0, w0 = img.shape[:-1] h, w = (s, int(s * w / h)) if h > w else (int(s * h / w), s) ph, pw = s - h, s - w img_input = np.zeros([s, s, 3], dtype=np.float32) img_input[ph // 2:ph // 2 + h, pw // 2:pw // 2 + w] = transform.resize(img, (h, w)) img_input = np.transpose(img_input, (2, 0, 1)) img_input = img_input[np.newaxis, :] mask = self.anime_seg.run(None, {'img': img_input})[0][0] mask = np.transpose(mask, (1, 2, 0)) mask = mask[ph // 2:ph // 2 + h, pw // 2:pw // 2 + w] mask = transform.resize(mask, (h0, w0)) img0 = (img0 * mask + 255 * (1 - mask)).astype(np.uint8) return img0 def encode_img(self, img): img = transform.resize(((img / 255 - 0.5) / 0.5), (256, 256)).transpose(2, 0, 1)[np.newaxis, :].astype( np.float32) return self.encoder.run(None, {'img': img})[0] + self.w_avg def detect(self, im0, conf_thres, iou_thres, detail=False): if im0 is None: return [] img = letterbox((im0 / 255).astype(np.float32), (self.detector_imgsz, self.detector_imgsz), stride=self.detector_stride) # Convert img = img.transpose(2, 0, 1) img = img[np.newaxis, :] pred = self.detector.run(None, {'images': img})[0][0] dets = nms(pred, conf_thres, iou_thres) imgs = [] # Print results s = '%gx%g ' % img.shape[2:] # print string for i, det in enumerate(dets): n = len(det) s += f"{n} {self.detector_class_names[i]}{'s' * (n > 1)}, " # add to string if detail: print(s) waifu_rects = [] head_rects = [] body_rects = [] for i, det in enumerate(dets): for x in det: # Rescale boxes from img_size to im0 size wr = im0.shape[1] / img.shape[3] hr = im0.shape[0] / img.shape[2] x[1] = (int(x[1][0] * wr), int(x[1][1] * hr), int(x[1][2] * wr), int(x[1][3] * hr)) if i == 0: head_rects.append(x[1]) elif i == 1: body_rects.append(x[1]) elif i == 2: waifu_rects.append(x[1]) for j, waifu_rect in enumerate(waifu_rects): msg = f'waifu {j + 1} ' head_num = 0 body_num = 0 hr, br = None, None for r in head_rects: if get_inter(r, waifu_rect) / ((r[2] - r[0]) * (r[3] - r[1])) > 0.75: hr = r head_num += 1 if head_num != 1: if detail: print(msg + f'head num error: {head_num}') continue for r in body_rects: if get_inter(r, waifu_rect) / ((r[2] - r[0]) * (r[3] - r[1])) > 0.65: br = r body_num += 1 if body_num != 1: if detail: print(msg + f'body num error: {body_num}') continue bounds = (min(waifu_rect[0], hr[0], br[0]), min(waifu_rect[1], hr[1], br[1]), max(waifu_rect[2], hr[2], br[2]), max(waifu_rect[3], hr[3], br[3])) if (bounds[2] - bounds[0]) / (bounds[3] - bounds[1]) > 0.7: if detail: print(msg + "ratio out of limit") continue expand_pixel = (bounds[3] - bounds[1]) // 20 bounds = [max(bounds[0] - expand_pixel // 2, 0), max(bounds[1] - expand_pixel, 0), min(bounds[2] + expand_pixel // 2, im0.shape[1]), min(bounds[3] + expand_pixel, im0.shape[0]), ] # corp and resize w = bounds[2] - bounds[0] h = bounds[3] - bounds[1] bounds[3] += h % 2 h += h % 2 r = min(512 / w, 1024 / h) pw, ph = int(512 / r - w), int(1024 / r - h) bounds_tmp = (bounds[0] - pw // 2, bounds[1] - ph // 2, bounds[2] + pw // 2 + pw % 2, bounds[3] + ph // 2 + ph % 2) bounds = (max(0, bounds_tmp[0]), max(0, bounds_tmp[1]), min(im0.shape[1], bounds_tmp[2]), min(im0.shape[0], bounds_tmp[3])) dl = bounds[0] - bounds_tmp[0] dr = bounds[2] - bounds_tmp[2] dt = bounds[1] - bounds_tmp[1] db = bounds[3] - bounds_tmp[3] w = bounds_tmp[2] - bounds_tmp[0] h = bounds_tmp[3] - bounds_tmp[1] temp_img = np.full((h, w, 3), 255, dtype=np.uint8) temp_img[dt:h + db, dl:w + dr] = im0[bounds[1]:bounds[3], bounds[0]:bounds[2]] temp_img = transform.resize(temp_img, (1024, 512), preserve_range=True).astype(np.uint8) imgs.append(temp_img) return imgs def gen_video(self, w1, w2, noise, path, frame_num=10): video = imageio.get_writer(path, mode='I', fps=frame_num // 2, codec='libx264', bitrate='16M') lin = np.linspace(0, 1, frame_num) for i in range(0, frame_num): img = self.get_img(((1 - lin[i]) * w1) + (lin[i] * w2), noise) video.append_data(img) for i in reversed(range(0, frame_num)): img = self.get_img(((1 - lin[i]) * w1) + (lin[i] * w2), noise) video.append_data(img) video.close() def get_thumbnail(img): img_new = np.full((256, 384, 3), 200, dtype=np.uint8) img_new[:, 128:256] = transform.resize(img, (256, 128), preserve_range=True) return img_new def gen_fn(seed, random_seed, psi1, psi2, noise): if random_seed: seed = random.randint(0, 2 ** 32 - 1) z = RandomState(int(seed)).randn(1, 1024) w = model.get_w(z.astype(dtype=np.float32), psi1, psi2) img_out = model.get_img(w, noise) return img_out, seed, w, get_thumbnail(img_out) def encode_img_fn(img, noise): if img is None: return "please upload a image", None, None, None, None img = model.remove_bg(img) imgs = model.detect(img, 0.2, 0.03) if len(imgs) == 0: return "failed to detect anime character", None, None, None, None w = model.encode_img(imgs[0]) img_out = model.get_img(w, noise) return "success", imgs[0], img_out, w, get_thumbnail(img_out) def gen_video_fn(w1, w2, noise, frame): if w1 is None or w2 is None: return None model.gen_video(w1, w2, noise, "video.mp4", int(frame)) return "video.mp4" if __name__ == '__main__': model = Model() app = gr.Blocks() with app: gr.Markdown("# full-body anime GAN\n\n" "![visitor badge](https://visitor-badge.glitch.me/badge?page_id=skytnt.full-body-anime-gan)\n\n") with gr.Tabs(): with gr.TabItem("generate image"): with gr.Row(): with gr.Column(): gr.Markdown("generate image") with gr.Row(): gen_input1 = gr.Slider(minimum=0, maximum=2 ** 32 - 1, step=1, value=0, label="seed") gen_input2 = gr.Checkbox(label="Random", value=True) gen_input3 = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.7, label="truncation psi 1") gen_input4 = gr.Slider(minimum=0, maximum=1, step=0.01, value=1, label="truncation psi 2") gen_input5 = gr.Slider(minimum=0, maximum=1, step=0.01, value=1, label="noise strength") with gr.Group(): gen_submit = gr.Button("Generate", variant="primary") with gr.Column(): gen_output1 = gr.Image(label="output image") select_img_input_w1 = gr.Variable() select_img_input_img1 = gr.Variable() with gr.TabItem("encode image"): with gr.Row(): with gr.Column(): gr.Markdown("you'd better upload a standing full-body image") encode_img_input = gr.Image(label="input image") examples_data = [[f"examples/{x:02d}.jpg"] for x in range(1, 5)] encode_img_examples = gr.Dataset(components=[encode_img_input], samples=examples_data) with gr.Group(): encode_img_submit = gr.Button("Run", variant="primary") with gr.Column(): encode_img_output1 = gr.Textbox(label="output message") with gr.Row(): encode_img_output2 = gr.Image(label="detected") encode_img_output3 = gr.Image(label="encoded") select_img_input_w2 = gr.Variable() select_img_input_img2 = gr.Variable() with gr.TabItem("generate video"): with gr.Row(): with gr.Column(): gr.Markdown("generate video between 2 images") with gr.Row(): with gr.Column(): select_img1_dropdown = gr.Radio(label="Select image 1", value="current generated image", choices=["current generated image", "current encoded image"], type="index") with gr.Group(): select_img1_button = gr.Button("Select", variant="primary") select_img1_output_img = gr.Image(label="selected image 1") select_img1_output_w = gr.Variable() with gr.Column(): select_img2_dropdown = gr.Radio(label="Select image 2", value="current generated image", choices=["current generated image", "current encoded image"], type="index") with gr.Group(): select_img2_button = gr.Button("Select", variant="primary") select_img2_output_img = gr.Image(label="selected image 2") select_img2_output_w = gr.Variable() generate_video_frame = gr.Slider(minimum=10, maximum=30, step=1, label="frame", value=15) with gr.Group(): generate_video_button = gr.Button("Generate", variant="primary") with gr.Column(): generate_video_output = gr.Video(label="output video") gen_submit.click(gen_fn, [gen_input1, gen_input2, gen_input3, gen_input4, gen_input5], [gen_output1, gen_input1, select_img_input_w1, select_img_input_img1]) encode_img_submit.click(encode_img_fn, [encode_img_input, gen_input5], [encode_img_output1, encode_img_output2, encode_img_output3, select_img_input_w2, select_img_input_img2]) encode_img_examples.click(lambda x: x[0], [encode_img_examples], [encode_img_input]) select_img1_button.click(lambda i, img1, img2, w1, w2: (img1, w1) if i == 0 else (img2, w2), [select_img1_dropdown, select_img_input_img1, select_img_input_img2, select_img_input_w1, select_img_input_w2], [select_img1_output_img, select_img1_output_w]) select_img2_button.click(lambda i, img1, img2, w1, w2: (img1, w1) if i == 0 else (img2, w2), [select_img2_dropdown, select_img_input_img1, select_img_input_img2, select_img_input_w1, select_img_input_w2], [select_img2_output_img, select_img2_output_w]) generate_video_button.click(gen_video_fn, [select_img1_output_w, select_img2_output_w, gen_input5, generate_video_frame], [generate_video_output]) app.launch()