Updated app, added video generation and model files

app.py

@@ -1,3 +1,207 @@
+import os
+
+import torch
 import gradio as gr
 
-gr.Interface.load("spaces/eugenesiow/remove-bg").launch()
+import os
+import sys
+import numpy as np
+
+from e4e.models.psp import pSp
+from util import *
+from huggingface_hub import hf_hub_download
+
+import os
+import sys
+import tempfile
+import shutil
+from argparse import Namespace
+from pathlib import Path
+
+import dlib
+import numpy as np
+import torchvision.transforms as transforms
+from torchvision import utils
+from PIL import Image
+
+from model.sg2_model import Generator
+from generate_videos import generate_frames, video_from_interpolations, vid_to_gif
+
+model_dir = "models"
+os.makedirs(model_dir, exist_ok=True)
+
+models_and_paths = {"akhaliq/JoJoGAN_e4e_ffhq_encode": "e4e_ffhq_encode.pt",
+                    "akhaliq/jojogan_dlib": "shape_predictor_68_face_landmarks.dat",
+                    "akhaliq/jojogan-stylegan2-ffhq-config-f": f"{model_dir}/base.pt"}
+
+def get_models():
+    for repo_id, file_path in models_and_paths:
+        hf_hub_download(repo_id=repo_id, filename=file_path)
+
+    model_list = ['base'] + [Path(model_ckpt).stem for model_ckpt in os.listdir(model_dir) if not 'base' in model_ckpt]
+
+    return model_list
+
+model_list = get_models()
+
+class ImageEditor(object):
+    def __init__(self):
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        latent_size = 512
+        n_mlp = 8
+        channel_mult = 2
+        model_size = 1024
+
+        self.generators = {}
+
+        for model in model_list:
+            g_ema = Generator(
+                model_size, latent_size, n_mlp, channel_multiplier=channel_mult
+            ).to(self.device)
+
+            checkpoint = torch.load(f"models/{model}.pt")
+
+            g_ema.load_state_dict(checkpoint['g_ema'])
+
+            self.generators[model] = g_ema
+
+        self.experiment_args = {"model_path": "e4e_ffhq_encode.pt"}
+        self.experiment_args["transform"] = transforms.Compose(
+            [
+                transforms.Resize((256, 256)),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
+            ]
+        )
+        self.resize_dims = (256, 256)
+
+        model_path = self.experiment_args["model_path"]
+
+        ckpt = torch.load(model_path, map_location="cpu")
+        opts = ckpt["opts"]
+
+        opts["checkpoint_path"] = model_path
+        opts = Namespace(**opts)
+
+        self.e4e_net = pSp(opts)
+        self.e4e_net.eval()
+        self.e4e_net.cuda()
+
+        self.shape_predictor = dlib.shape_predictor(
+            models_and_paths["akhaliq/jojogan_dlib"]
+        )
+
+        print("setup complete")
+
+    def get_style_list(self):
+        style_list = ['all', 'list - enter below']
+
+        for key in self.generators:
+            style_list.append(key)
+
+        return style_list
+
+    def predict(
+        self,
+        input,              # Input image path
+        output_style,       # Which output style do you want to use?
+        style_list,         # Comma seperated list of models to use. Only accepts models from the output_style list
+        generate_video,     # Generate a video instead of an output image
+        with_editing,       # Apply latent space editing to the generated video
+        video_format        # Choose gif to display in browser, mp4 for higher-quality downloadable video
+    ):  
+
+        if output_style == 'all':
+            styles = model_list
+        elif output_style == 'list - enter below':
+            styles = style_list.split(",")
+            for style in styles:
+                if style not in model_list:
+                    raise ValueError(f"Encountered style '{style}' in the style_list which is not an available option.")
+        else:
+            styles = [output_style]
+
+        # @title Align image
+        input_image = self.run_alignment(str(input))
+        
+        input_image = input_image.resize(self.resize_dims)
+
+        img_transforms = self.experiment_args["transform"]
+        transformed_image = img_transforms(input_image)
+
+        with torch.no_grad():
+            images, latents = self.run_on_batch(transformed_image.unsqueeze(0))
+            result_image, latent = images[0], latents[0]
+
+        inverted_latent = latent.unsqueeze(0).unsqueeze(1)
+        out_dir = Path(tempfile.mkdtemp())
+        out_path = out_dir / "out.jpg"
+        
+        generators = [self.generators[style] for style in styles]
+
+        if not generate_video:
+            with torch.no_grad():
+                img_list = []
+                for g_ema in generators:
+                    img, _ = g_ema(inverted_latent, input_is_latent=True, truncation=1, randomize_noise=False)
+                    img_list.append(img)
+                
+                out_img = torch.cat(img_list, axis=0)
+                utils.save_image(out_img, out_path, nrow=int(np.sqrt(out_img.size(0))), normalize=True, scale_each=True, range=(-1, 1))
+
+            return Path(out_path)
+        
+        return self.generate_vid(generators, inverted_latent, out_dir, video_format, with_editing)
+
+    def generate_vid(self, generators, latent, out_dir, video_format, with_editing):      
+        np_latent = latent.squeeze(0).cpu().detach().numpy()
+        args = {
+                'fps': 24,
+                'target_latents': None,
+                'edit_directions': None,
+                'unedited_frames': 0 if with_editing else 40 * (len(generators) - 1)
+                }
+
+        args = Namespace(**args)
+        with tempfile.TemporaryDirectory() as dirpath:
+
+            generate_frames(args, np_latent, generators, dirpath)
+            video_from_interpolations(args.fps, dirpath)
+            
+            gen_path = Path(dirpath) / "out.mp4"
+            out_path = out_dir / f"out.{video_format}"
+
+            if video_format == 'gif':
+                vid_to_gif(gen_path, out_dir, scale=256, fps=args.fps)
+            else:
+                shutil.copy2(gen_path, out_path) 
+
+        return out_path
+
+    def run_alignment(self, image_path):
+        aligned_image = align_face(filepath=image_path, predictor=self.shape_predictor)
+        print("Aligned image has shape: {}".format(aligned_image.size))
+        return aligned_image
+
+    def run_on_batch(self, inputs):
+        images, latents = self.e4e_net(
+            inputs.to("cuda").float(), randomize_noise=False, return_latents=True
+        )
+        return images, latents
+
+editor = ImageEditor()
+
+title = "StyleGAN-NADA"
+description = "Gradio Demo for StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators (SIGGRAPH 2022). To use it, upload your image and select a target style. More information about the paper and training new models can be found below."
+
+article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2108.00946' target='_blank'>StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators</a> | <a href='https://stylegan-nada.github.io/' target='_blank'>Project Page</a> | <a href='https://github.com/rinongal/StyleGAN-nada' target='_blank'>Code</a></p> <center><img src='https://visitor-badge.glitch.me/badge?page_id=rinong_sgnada' alt='visitor badge'></center>"
+
+gr.Interface(editor.predict, [gr.inputs.Image(type="pil"), 
+                              gr.inputs.Dropdown(choices=editor.get_style_list(), type="value", default='base', label="Model"), 
+                              gr.inputs.Textbox(lines=1, placeholder=None, default="joker,anime,modigliani", label="Style List", optional=True),
+                              gr.inputs.Checkbox(default=False, label="Generate Video?", optional=False),
+                              gr.inputs.Checkbox(default=False, label="With Editing?", optional=False),
+                              gr.inputs.Radio(choices=["gif", "mp4"], type="value", default='mp4', label="Video Format")],
+                              gr.outputs.Image(type="file"), title=title, description=description, article=article, allow_flagging=False, allow_screenshot=False).launch()
+

generate_videos.py

@@ -0,0 +1,259 @@
+'''
+Tool for generating editing videos across different domains.
+
+Given a set of latent codes and pre-trained models, it will interpolate between the different codes in each of the target domains
+and combine the resulting images into a video.
+
+Example run command:
+
+python generate_videos.py --ckpt /model_dir/pixar.pt \
+                                 /model_dir/ukiyoe.pt \
+                                 /model_dir/edvard_munch.pt \
+                                 /model_dir/botero.pt \
+                          --out_dir /output/video/ \
+                          --source_latent /latents/latent000.npy \
+                          --target_latents /latents/
+
+'''
+
+import os
+import argparse
+
+import torch
+from torchvision import utils
+
+from model.sg2_model import Generator
+from tqdm import tqdm
+from pathlib import Path
+
+import numpy as np
+
+import subprocess
+import shutil
+import copy
+
+VALID_EDITS = ["pose", "age", "smile", "gender", "hair_length", "beard"]
+
+SUGGESTED_DISTANCES = {
+                       "pose": (3.0, -3.0),
+                       "smile": (2.0, -2.0),
+                       "age": (4.0, -4.0),
+                       "gender": (3.0, -3.0),
+                       "hair_length": (None, -4.0),
+                       "beard": (2.0, None)
+                      }
+                    
+def project_code(latent_code, boundary, distance=3.0):
+
+    if len(boundary) == 2:
+        boundary = boundary.reshape(1, 1, -1)
+
+    return latent_code + distance * boundary
+
+def generate_frames(args, source_latent, g_ema_list, output_dir):
+
+    alphas = np.linspace(0, 1, num=20)
+    
+    interpolate_func = interpolate_with_boundaries # default
+    if args.target_latents:                        # if provided with targets
+        interpolate_func = interpolate_with_target_latents
+    if args.unedited_frames:                       # if only interpolating through generators
+        interpolate_func = duplicate_latent
+
+    latents = interpolate_func(args, source_latent, alphas)
+
+    segments = len(g_ema_list) - 1
+    if segments:
+        segment_length = len(latents) / segments
+
+        g_ema = copy.deepcopy(g_ema_list[0])
+
+        src_pars = dict(g_ema.named_parameters())
+        mix_pars = [dict(model.named_parameters()) for model in g_ema_list]
+    else:
+        g_ema = g_ema_list[0]
+
+    print("Generating frames for video...")
+    for idx, latent in tqdm(enumerate(latents), total=len(latents)):
+
+        if segments:
+            mix_alpha = (idx % segment_length) * 1.0 / segment_length
+            segment_id = int(idx // segment_length)
+
+            for k in src_pars.keys():
+                src_pars[k].data.copy_(mix_pars[segment_id][k] * (1 - mix_alpha) + mix_pars[segment_id + 1][k] * mix_alpha)
+
+        if idx == 0 or segments or latent is not latents[idx - 1]:
+            w = torch.from_numpy(latent).float().cuda()
+
+            with torch.no_grad():
+                img, _ = g_ema([w], input_is_latent=True, truncation=1, randomize_noise=False)
+
+        utils.save_image(img, f"{output_dir}/{str(idx).zfill(3)}.jpg", nrow=1, normalize=True, scale_each=True, range=(-1, 1))
+
+def interpolate_forward_backward(source_latent, target_latent, alphas):
+    latents_forward  = [a * target_latent + (1-a) * source_latent for a in alphas] # interpolate from source to target
+    latents_backward = latents_forward[::-1]                                       # interpolate from target to source
+    return latents_forward + [target_latent] * 20 + latents_backward               # forward + short delay at target + return
+
+def duplicate_latent(args, source_latent, alphas):
+    return [source_latent for _ in range(args.unedited_frames)]
+
+def interpolate_with_boundaries(args, source_latent, alphas):
+    edit_directions = args.edit_directions or ['pose', 'smile', 'gender', 'age', 'hair_length']
+    
+    # interpolate latent codes with all targets
+
+    print("Interpolating latent codes...")
+
+    boundary_dir = Path(os.path.abspath(__file__)).parents[1].joinpath("editing", "interfacegan_boundaries")
+
+    boundaries_and_distances = []
+    for direction_type in edit_directions:
+        distances = SUGGESTED_DISTANCES[direction_type]
+        boundary = torch.load(os.path.join(boundary_dir, f'{direction_type}.pt')).cpu().detach().numpy()
+
+        for distance in distances:
+            if distance:
+                boundaries_and_distances.append((boundary, distance))
+
+    latents = []
+    for boundary, distance in boundaries_and_distances:
+        
+        target_latent = project_code(source_latent, boundary, distance)
+        latents.extend(interpolate_forward_backward(source_latent, target_latent, alphas)) 
+
+    return latents
+
+def interpolate_with_target_latents(args, source_latent, alphas):    
+    # interpolate latent codes with all targets
+
+    print("Interpolating latent codes...")
+    
+    latents = []
+    for target_latent_path in args.target_latents:
+        
+        if target_latent_path == args.source_latent:
+            continue
+
+        target_latent = np.load(target_latent_path, allow_pickle=True)
+
+        latents.extend(interpolate_forward_backward(source_latent, target_latent, alphas))
+
+    return latents
+
+def video_from_interpolations(fps, output_dir):
+
+    # combine frames to a video
+    command = ["ffmpeg", 
+               "-r", f"{fps}", 
+               "-i", f"{output_dir}/%03d.jpg", 
+               "-c:v", "libx264", 
+               "-vf", f"fps={fps}", 
+               "-pix_fmt", "yuv420p", 
+               f"{output_dir}/out.mp4"]
+    
+    subprocess.call(command)
+
+def merge_videos(output_dir, num_subdirs):
+
+    output_file = os.path.join(output_dir, "combined.mp4")
+
+    if num_subdirs == 1: # if we only have one video, just copy it over
+        shutil.copy2(os.path.join(output_dir, str(0), "out.mp4"), output_file)
+    else:                # otherwise merge using ffmpeg
+        command = ["ffmpeg"]
+        for dir in range(num_subdirs):
+            command.extend(['-i', os.path.join(output_dir, str(dir), "out.mp4")])
+        
+        sqrt_subdirs = int(num_subdirs ** .5)
+
+        if (sqrt_subdirs ** 2) != num_subdirs:
+            raise ValueError("Number of checkpoints cannot be arranged in a square grid")
+        
+        command.append("-filter_complex")
+
+        filter_string = ""
+        vstack_string = ""
+        for row in range(sqrt_subdirs):
+            row_str = ""
+            for col in range(sqrt_subdirs):
+                row_str += f"[{row * sqrt_subdirs + col}:v]"
+
+            letter = chr(ord('A')+row)
+            row_str += f"hstack=inputs={sqrt_subdirs}[{letter}];"
+            vstack_string += f"[{letter}]"
+
+            filter_string += row_str
+        
+        vstack_string += f"vstack=inputs={sqrt_subdirs}[out]"
+        filter_string += vstack_string
+
+        command.extend([filter_string, "-map", "[out]", output_file])
+
+        subprocess.call(command)
+
+def vid_to_gif(vid_path, output_dir, scale=256, fps=35):
+
+    command = ["ffmpeg", 
+               "-i", f"{vid_path}", 
+               "-vf", f"fps={fps},scale={scale}:-1:flags=lanczos,split[s0][s1];[s0]palettegen[p];[s1]fifo[s2];[s2][p]paletteuse",  
+               "-loop", "0",
+               f"{output_dir}/out.gif"]
+    
+    subprocess.call(command)
+
+
+if __name__ == '__main__':
+    device = 'cuda'
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--size', type=int, default=1024)
+    parser.add_argument('--ckpt', type=str, nargs="+", required=True, help="Path to one or more pre-trained generator checkpoints.")
+    parser.add_argument('--channel_multiplier', type=int, default=2)
+    parser.add_argument('--out_dir', type=str, required=True, help="Directory where output files will be placed")
+    parser.add_argument('--source_latent', type=str, required=True, help="Path to an .npy file containing an initial latent code")
+    parser.add_argument('--target_latents', nargs="+", type=str, help="A list of paths to .npy files containing target latent codes to interpolate towards, or a directory containing such .npy files.")
+    parser.add_argument('--force', '-f', action='store_true', help="Force run with non-empty directory. Image files not overwritten by the proccess may still be included in the final video")
+    parser.add_argument('--fps', default=35, type=int, help='Frames per second in the generated videos.')
+    parser.add_argument('--edit_directions', nargs="+", type=str, help=f"A list of edit directions to use in video generation (if not using a target latent directory). Available directions are: {VALID_EDITS}")
+    parser.add_argument('--unedited_frames', type=int, default=0, help="Used to generate videos with no latent editing. If set to a positive number and target_latents is not provided, will simply duplicate the initial frame <unedited_frames> times.")
+    
+    args = parser.parse_args()
+
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    if not args.force and os.listdir(args.out_dir):
+        print("Output directory is not empty. Either delete the directory content or re-run with -f.")
+        exit(0)
+
+    if args.target_latents and len(args.target_latents) == 1 and os.path.isdir(args.target_latents[0]):
+        args.target_latents = [os.path.join(args.target_latents[0], file_name) for file_name in os.listdir(args.target_latents[0]) if file_name.endswith(".npy")]
+        args.target_latents = sorted(args.target_latents)
+
+    args.latent = 512
+    args.n_mlp = 8
+
+    g_ema = Generator(
+        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier
+    ).to(device)
+
+    source_latent = np.load(args.source_latent, allow_pickle=True)
+
+    for idx, ckpt_path in enumerate(args.ckpt):
+        print(f"Generating video using checkpoint: {ckpt_path}")
+        checkpoint = torch.load(ckpt_path)
+
+        g_ema.load_state_dict(checkpoint['g_ema'])
+
+        output_dir = os.path.join(args.out_dir, str(idx))
+        os.makedirs(output_dir)
+
+        generate_frames(args, source_latent, [g_ema], output_dir)
+        video_from_interpolations(args.fps, output_dir)
+
+    merge_videos(args.out_dir, len(args.ckpt))
+
+    
+

model/sg2_model.py

@@ -0,0 +1,780 @@
+import math
+import random
+import functools
+import operator
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.autograd import Function
+
+from op import FusedLeakyReLU, fused_leaky_relu, upfirdn2d, conv2d_gradfix
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+
+def make_kernel(k):
+    k = torch.tensor(k, dtype=torch.float32)
+
+    if k.ndim == 1:
+        k = k[None, :] * k[:, None]
+
+    k /= k.sum()
+
+    return k
+
+
+class Upsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel) * (factor ** 2)
+        self.register_buffer("kernel", kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2 + factor - 1
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=self.factor, down=1, pad=self.pad)
+
+        return out
+
+
+class Downsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel)
+        self.register_buffer("kernel", kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=1, down=self.factor, pad=self.pad)
+
+        return out
+
+
+class Blur(nn.Module):
+    def __init__(self, kernel, pad, upsample_factor=1):
+        super().__init__()
+
+        kernel = make_kernel(kernel)
+
+        if upsample_factor > 1:
+            kernel = kernel * (upsample_factor ** 2)
+
+        self.register_buffer("kernel", kernel)
+
+        self.pad = pad
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, pad=self.pad)
+
+        return out
+
+
+class EqualConv2d(nn.Module):
+    def __init__(
+        self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(
+            torch.randn(out_channel, in_channel, kernel_size, kernel_size)
+        )
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+
+        self.stride = stride
+        self.padding = padding
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channel))
+
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        out = conv2d_gradfix.conv2d(
+            input,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},"
+            f" {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})"
+        )
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})"
+        )
+
+
+class ModulatedConv2d(nn.Module):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        style_dim,
+        demodulate=True,
+        upsample=False,
+        downsample=False,
+        blur_kernel=[1, 3, 3, 1],
+        fused=True,
+    ):
+        super().__init__()
+
+        self.eps = 1e-8
+        self.kernel_size = kernel_size
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.upsample = upsample
+        self.downsample = downsample
+
+        if upsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) - (kernel_size - 1)
+            pad0 = (p + 1) // 2 + factor - 1
+            pad1 = p // 2 + 1
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1), upsample_factor=factor)
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1))
+
+        fan_in = in_channel * kernel_size ** 2
+        self.scale = 1 / math.sqrt(fan_in)
+        self.padding = kernel_size // 2
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channel, in_channel, kernel_size, kernel_size)
+        )
+
+        self.modulation = EqualLinear(style_dim, in_channel, bias_init=1)
+
+        self.demodulate = demodulate
+        self.fused = fused
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}({self.in_channel}, {self.out_channel}, {self.kernel_size}, "
+            f"upsample={self.upsample}, downsample={self.downsample})"
+        )
+
+    def forward(self, input, style, is_s_code=False):
+        batch, in_channel, height, width = input.shape
+
+        if not self.fused:
+            weight = self.scale * self.weight.squeeze(0)
+
+            if is_s_code:
+                style = style[self.modulation]
+            else:
+                style = self.modulation(style)
+
+            if self.demodulate:
+                w = weight.unsqueeze(0) * style.view(batch, 1, in_channel, 1, 1)
+                dcoefs = (w.square().sum((2, 3, 4)) + 1e-8).rsqrt()
+
+            input = input * style.reshape(batch, in_channel, 1, 1)
+
+            if self.upsample:
+                weight = weight.transpose(0, 1)
+                out = conv2d_gradfix.conv_transpose2d(
+                    input, weight, padding=0, stride=2
+                )
+                out = self.blur(out)
+
+            elif self.downsample:
+                input = self.blur(input)
+                out = conv2d_gradfix.conv2d(input, weight, padding=0, stride=2)
+
+            else:
+                out = conv2d_gradfix.conv2d(input, weight, padding=self.padding)
+
+            if self.demodulate:
+                out = out * dcoefs.view(batch, -1, 1, 1)
+
+            return out
+
+        if is_s_code:
+            style = style[self.modulation]
+        else:
+            style = self.modulation(style)
+
+        style = style.view(batch, 1, in_channel, 1, 1)
+        weight = self.scale * self.weight * style
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + 1e-8)
+            weight = weight * demod.view(batch, self.out_channel, 1, 1, 1)
+
+        weight = weight.view(
+            batch * self.out_channel, in_channel, self.kernel_size, self.kernel_size
+        )
+
+        if self.upsample:
+            input = input.view(1, batch * in_channel, height, width)
+            weight = weight.view(
+                batch, self.out_channel, in_channel, self.kernel_size, self.kernel_size
+            )
+            weight = weight.transpose(1, 2).reshape(
+                batch * in_channel, self.out_channel, self.kernel_size, self.kernel_size
+            )
+            out = conv2d_gradfix.conv_transpose2d(
+                input, weight, padding=0, stride=2, groups=batch
+            )
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+            out = self.blur(out)
+
+        elif self.downsample:
+            input = self.blur(input)
+            _, _, height, width = input.shape
+            input = input.view(1, batch * in_channel, height, width)
+            out = conv2d_gradfix.conv2d(
+                input, weight, padding=0, stride=2, groups=batch
+            )
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        else:
+            input = input.view(1, batch * in_channel, height, width)
+            out = conv2d_gradfix.conv2d(
+                input, weight, padding=self.padding, groups=batch
+            )
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        return out
+
+
+class NoiseInjection(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.zeros(1))
+
+    def forward(self, image, noise=None):
+        if noise is None:
+            batch, _, height, width = image.shape
+            noise = image.new_empty(batch, 1, height, width).normal_()
+
+        return image + self.weight * noise
+
+
+class ConstantInput(nn.Module):
+    def __init__(self, channel, size=4):
+        super().__init__()
+
+        self.input = nn.Parameter(torch.randn(1, channel, size, size))
+
+    def forward(self, input, is_s_code=False):
+        if not is_s_code:
+            batch = input.shape[0]
+        else: 
+            batch = next(iter(input.values())).shape[0]
+
+        out = self.input.repeat(batch, 1, 1, 1)
+
+        return out
+
+
+class StyledConv(nn.Module):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        style_dim,
+        upsample=False,
+        blur_kernel=[1, 3, 3, 1],
+        demodulate=True,
+    ):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=upsample,
+            blur_kernel=blur_kernel,
+            demodulate=demodulate,
+        )
+
+        self.noise = NoiseInjection()
+        # self.bias = nn.Parameter(torch.zeros(1, out_channel, 1, 1))
+        # self.activate = ScaledLeakyReLU(0.2)
+        self.activate = FusedLeakyReLU(out_channel)
+
+    def forward(self, input, style, noise=None, is_s_code=False):
+        out = self.conv(input, style, is_s_code=is_s_code)
+        out = self.noise(out, noise=noise)
+        # out = out + self.bias
+        out = self.activate(out)
+
+        return out
+
+
+class ToRGB(nn.Module):
+    def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        if upsample:
+            self.upsample = Upsample(blur_kernel)
+
+        self.conv = ModulatedConv2d(in_channel, 3, 1, style_dim, demodulate=False)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, input, style, skip=None, is_s_code=False):
+        out = self.conv(input, style, is_s_code=is_s_code)
+        out = out + self.bias
+
+        if skip is not None:
+            skip = self.upsample(skip)
+
+            out = out + skip
+
+        return out
+
+
+class Generator(nn.Module):
+    def __init__(
+        self,
+        size,
+        style_dim,
+        n_mlp,
+        channel_multiplier=2,
+        blur_kernel=[1, 3, 3, 1],
+        lr_mlp=0.01,
+    ):
+        super().__init__()
+
+        self.size = size
+
+        self.style_dim = style_dim
+
+        layers = [PixelNorm()]
+
+        for i in range(n_mlp):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=lr_mlp, activation="fused_lrelu"
+                )
+            )
+
+        self.style = nn.Sequential(*layers)
+
+        self.channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        self.input = ConstantInput(self.channels[4])
+        self.conv1 = StyledConv(
+            self.channels[4], self.channels[4], 3, style_dim, blur_kernel=blur_kernel
+        )
+        self.to_rgb1 = ToRGB(self.channels[4], style_dim, upsample=False)
+
+        self.log_size = int(math.log(size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+
+        self.convs = nn.ModuleList()
+        self.upsamples = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channel = self.channels[4]
+
+        for layer_idx in range(self.num_layers):
+            res = (layer_idx + 5) // 2
+            shape = [1, 1, 2 ** res, 2 ** res]
+            self.noises.register_buffer(f"noise_{layer_idx}", torch.randn(*shape))
+
+        for i in range(3, self.log_size + 1):
+            out_channel = self.channels[2 ** i]
+
+            self.convs.append(
+                StyledConv(
+                    in_channel,
+                    out_channel,
+                    3,
+                    style_dim,
+                    upsample=True,
+                    blur_kernel=blur_kernel,
+                )
+            )
+
+            self.convs.append(
+                StyledConv(
+                    out_channel, out_channel, 3, style_dim, blur_kernel=blur_kernel
+                )
+            )
+
+            self.to_rgbs.append(ToRGB(out_channel, style_dim))
+
+            in_channel = out_channel
+
+        self.n_latent = self.log_size * 2 - 2
+
+
+        self.modulation_layers = [self.conv1.conv.modulation, self.to_rgb1.conv.modulation] + \
+                                 [layer.conv.modulation for layer in self.convs]            + \
+                                 [layer.conv.modulation for layer in self.to_rgbs]
+
+    def make_noise(self):
+        device = self.input.input.device
+
+        noises = [torch.randn(1, 1, 2 ** 2, 2 ** 2, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2 ** i, 2 ** i, device=device))
+
+        return noises
+
+    def mean_latent(self, n_latent):
+        latent_in = torch.randn(
+            n_latent, self.style_dim, device=self.input.input.device
+        )
+        latent = self.style(latent_in).mean(0, keepdim=True)
+
+        return latent
+
+    def get_latent(self, input):
+        return self.style(input)
+
+    def get_s_code(self, styles, input_is_latent):
+
+        if not input_is_latent:
+            styles = [self.style(s) for s in styles]
+
+        s_codes = [{layer: layer(s) for layer in self.modulation_layers} for s in styles] * len(styles)
+
+        return s_codes    
+
+    def forward(
+        self,
+        styles,
+        return_latents=False,
+        inject_index=None,
+        truncation=1,
+        truncation_latent=None,
+        input_is_latent=False,
+        input_is_s_code=False,
+        noise=None,
+        randomize_noise=True,
+    ):
+        if not input_is_s_code:
+            return self.forward_with_w(styles, return_latents, inject_index, truncation, truncation_latent, input_is_latent, noise, randomize_noise)
+        
+        return self.forward_with_s(styles, return_latents, noise, randomize_noise)
+
+    def forward_with_w(
+            self,
+            styles,
+            return_latents=False,
+            inject_index=None,
+            truncation=1,
+            truncation_latent=None,
+            input_is_latent=False,
+            noise=None,
+            randomize_noise=True,
+        ):
+            if not input_is_latent:
+                styles = [self.style(s) for s in styles]
+
+            if noise is None:
+                if randomize_noise:
+                    noise = [None] * self.num_layers
+                else:
+                    noise = [
+                        getattr(self.noises, f"noise_{i}") for i in range(self.num_layers)
+                    ]
+
+            if truncation < 1:
+                style_t = []
+
+                for style in styles:
+                    style_t.append(
+                        truncation_latent + truncation * (style - truncation_latent)
+                    )
+
+                styles = style_t
+
+            if len(styles) < 2:
+                inject_index = self.n_latent
+
+                if styles[0].ndim < 3:
+                    latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+
+                else:
+                    latent = styles[0]
+
+            else:
+                if inject_index is None:
+                    inject_index = random.randint(1, self.n_latent - 1)
+
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+                latent2 = styles[1].unsqueeze(1).repeat(1, self.n_latent - inject_index, 1)
+
+                latent = torch.cat([latent, latent2], 1)
+
+            out = self.input(latent)
+            out = self.conv1(out, latent[:, 0], noise=noise[0])
+
+            skip = self.to_rgb1(out, latent[:, 1])
+
+            i = 1
+            for conv1, conv2, noise1, noise2, to_rgb in zip(
+                self.convs[::2], self.convs[1::2], noise[1::2], noise[2::2], self.to_rgbs
+            ):
+                out = conv1(out, latent[:, i], noise=noise1)
+                out = conv2(out, latent[:, i + 1], noise=noise2)
+                skip = to_rgb(out, latent[:, i + 2], skip)
+
+                i += 2
+
+            image = skip
+
+            if return_latents:
+                return image, latent
+
+            else:
+                return image, None
+
+    def forward_with_s(
+            self,
+            styles,
+            return_latents=False,
+            noise=None,
+            randomize_noise=True,
+        ):
+
+            if noise is None:
+                if randomize_noise:
+                    noise = [None] * self.num_layers
+                else:
+                    noise = [
+                        getattr(self.noises, f"noise_{i}") for i in range(self.num_layers)
+                    ]
+
+            out = self.input(styles, is_s_code=True)
+            out = self.conv1(out, styles, is_s_code=True, noise=noise[0])
+
+            skip = self.to_rgb1(out, styles, is_s_code=True)
+
+            i = 1
+            for conv1, conv2, noise1, noise2, to_rgb in zip(
+                self.convs[::2], self.convs[1::2], noise[1::2], noise[2::2], self.to_rgbs
+            ):
+                out = conv1(out, styles, is_s_code=True, noise=noise1)
+                out = conv2(out, styles, is_s_code=True, noise=noise2)
+                skip = to_rgb(out, styles, skip, is_s_code=True)
+
+                i += 2
+
+            image = skip
+
+            if return_latents:
+                return image, styles
+
+            else:
+                return image, None
+
+class ConvLayer(nn.Sequential):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        downsample=False,
+        blur_kernel=[1, 3, 3, 1],
+        bias=True,
+        activate=True,
+    ):
+        layers = []
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            layers.append(Blur(blur_kernel, pad=(pad0, pad1)))
+
+            stride = 2
+            self.padding = 0
+
+        else:
+            stride = 1
+            self.padding = kernel_size // 2
+
+        layers.append(
+            EqualConv2d(
+                in_channel,
+                out_channel,
+                kernel_size,
+                padding=self.padding,
+                stride=stride,
+                bias=bias and not activate,
+            )
+        )
+
+        if activate:
+            layers.append(FusedLeakyReLU(out_channel, bias=bias))
+
+        super().__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_channel, out_channel, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        self.conv1 = ConvLayer(in_channel, in_channel, 3)
+        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True)
+
+        self.skip = ConvLayer(
+            in_channel, out_channel, 1, downsample=True, activate=False, bias=False
+        )
+
+    def forward(self, input):
+        out = self.conv1(input)
+        out = self.conv2(out)
+
+        skip = self.skip(input)
+        out = (out + skip) / math.sqrt(2)
+
+        return out
+
+
+class Discriminator(nn.Module):
+    def __init__(self, size, channel_multiplier=2, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        convs = [ConvLayer(3, channels[size], 1)]
+
+        log_size = int(math.log(size, 2))
+
+        in_channel = channels[size]
+
+        for i in range(log_size, 2, -1):
+            out_channel = channels[2 ** (i - 1)]
+
+            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
+
+            in_channel = out_channel
+
+        self.convs = nn.Sequential(*convs)
+
+        self.stddev_group = 4
+        self.stddev_feat = 1
+
+        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
+        self.final_linear = nn.Sequential(
+            EqualLinear(channels[4] * 4 * 4, channels[4], activation="fused_lrelu"),
+            EqualLinear(channels[4], 1),
+        )
+
+    def forward(self, input):
+        out = self.convs(input)
+
+        batch, channel, height, width = out.shape
+        group = min(batch, self.stddev_group)
+        stddev = out.view(
+            group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
+        )
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, height, width)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+
+        out = out.view(batch, -1)
+        out = self.final_linear(out)
+
+        return out
+