app.py

import os
import random

import torch
import gradio as gr

from e4e.models.psp import pSp
from util import *
from huggingface_hub import hf_hub_download

import tempfile
from argparse import Namespace
import shutil

import dlib
import numpy as np
import torchvision.transforms as transforms
from torchvision import utils

from model.sg2_model import Generator
from generate_videos import generate_frames, video_from_interpolations, project_code_by_edit_name
from styleclip.styleclip_global import project_code_with_styleclip, style_tensor_to_style_dict

import clip

model_dir = "models"
os.makedirs(model_dir, exist_ok=True)

model_repos = {"e4e": ("akhaliq/JoJoGAN_e4e_ffhq_encode", "e4e_ffhq_encode.pt"),
               "dlib": ("akhaliq/jojogan_dlib", "shape_predictor_68_face_landmarks.dat"),
               "sc_fs3": ("rinong/stylegan-nada-models", "fs3.npy"),
               "base": ("akhaliq/jojogan-stylegan2-ffhq-config-f", "stylegan2-ffhq-config-f.pt"),
               "sketch": ("rinong/stylegan-nada-models", "sketch.pt"),
               "heat_miser": ("mjdolan/stylegan-nada-models", "heat.pt"),
               "santa": ("mjdolan/stylegan-nada-models", "santa.pt"),
               "jesus": ("mjdolan/stylegan-nada-models", "jesus.pt"),
               "mariah": ("mjdolan/stylegan-nada-models", "mariah.pt"),
               "claymation": ("mjdolan/stylegan-nada-models", "claymation.pt")
               }

def get_models():
    os.makedirs(model_dir, exist_ok=True)

    model_paths = {}

    for model_name, repo_details in model_repos.items():
        download_path = hf_hub_download(repo_id=repo_details[0], filename=repo_details[1])
        model_paths[model_name] = download_path

    return model_paths

model_paths = 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 = {}

        self.model_list = [name for name in model_paths.keys() if name not in ["e4e", "dlib", "sc_fs3"]]

        for model in self.model_list:
            g_ema = Generator(
                model_size, latent_size, n_mlp, channel_multiplier=channel_mult
            ).to(self.device)

            checkpoint = torch.load(model_paths[model], map_location=self.device)

            g_ema.load_state_dict(checkpoint['g_ema'])

            self.generators[model] = g_ema

        self.experiment_args = {"model_path": model_paths["e4e"]}
        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.device)
        self.e4e_net.eval()

        self.shape_predictor = dlib.shape_predictor(
            model_paths["dlib"]
        )

        self.styleclip_fs3 = torch.from_numpy(np.load(model_paths["sc_fs3"])).to(self.device)

        self.clip_model, _ = clip.load("ViT-B/32", device=self.device)

        print("setup complete")

    def get_style_list(self):
        style_list = []

        for key in self.generators:
            style_list.append(key)

        return style_list

    def invert_image(self, input_image):
        input_image = self.run_alignment(str(input_image))
        
        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)

        return inverted_latent
    
    def get_generators_for_styles(self, output_styles, loop_styles=False):

        if "base" in output_styles:              # always start with base if chosen
            output_styles.insert(0, output_styles.pop(output_styles.index("base")))
        if loop_styles:
            output_styles.append(output_styles[0])

        return [self.generators[style] for style in output_styles]

    def _pack_edits(func):
        def inner(self, 
                  pose_slider,
                  smile_slider, 
                  gender_slider, 
                  age_slider, 
                  hair_slider,
                  *args):

            edit_choices = {
                            "pose": pose_slider,
                            "smile": smile_slider,
                            "gender": gender_slider,
                            "age": age_slider,
                            "hair_length": hair_slider}

            return func(self, *args, edit_choices)

        return inner

    def get_target_latents(self, source_latent, edit_choices, generators):

        target_latents = []

        np_source_latent = source_latent.squeeze(0).cpu().detach().numpy()

        for attribute_name in ["pose", "smile", "gender", "age", "hair_length"]:
            strength = edit_choices[attribute_name]
            if strength != 0.0:
                projected_code_np = project_code_by_edit_name(np_source_latent, attribute_name, strength)
                target_latents.append(torch.from_numpy(projected_code_np).float().to(self.device))

        
        # if edit type is none or if all sliders were set to 0
        if not target_latents:
            target_latents = [source_latent.squeeze(0), ] * max((len(generators) - 1), 1)
        
        return target_latents

    @_pack_edits
    def edit_image(self, input, output_styles, edit_choices):
        return self.predict(input, output_styles, edit_choices=edit_choices)

    @_pack_edits
    def edit_video(self, input, output_styles, loop_styles, edit_choices):
        return self.predict(input, output_styles, generate_video=True, loop_styles=loop_styles, edit_choices=edit_choices)

    def predict(
        self,
        input,                  # Input image path
        output_styles,          # Style checkbox options.
        generate_video = False, # Generate a video instead of an output image
        loop_styles    = False, # Loop back to the initial style
        edit_choices   = None,  # Optional dictionary with edit choice arguments
    ):  

        if edit_choices is None:
            edit_choices = {"edit_type": "None"}

        # @title Align image
        out_dir = tempfile.mkdtemp()
        
        inverted_latent = self.invert_image(input)
        generators = self.get_generators_for_styles(output_styles, loop_styles)

        target_latents = self.get_target_latents(inverted_latent, edit_choices, generators)

        if not generate_video:
            output_paths = []

            with torch.no_grad():
                for g_ema in generators:
                    latent_for_gen = random.choice(target_latents)

                    img, _ = g_ema([latent_for_gen], input_is_latent=True, truncation=1, randomize_noise=False)

                    output_path = os.path.join(out_dir, f"out_{len(output_paths)}.jpg")
                    utils.save_image(img, output_path, nrow=1, normalize=True, range=(-1, 1))

                    output_paths.append(output_path)

            return output_paths

        return self.generate_vid(generators, inverted_latent, target_latents, out_dir)

    def generate_vid(self, generators, source_latent, target_latents, out_dir):    

        fps = 24

        with tempfile.TemporaryDirectory() as dirpath:
            generate_frames(source_latent, target_latents, generators, dirpath)
            video_from_interpolations(fps, dirpath)
            
            gen_path = os.path.join(dirpath, "out.mp4")
            out_path = os.path.join(out_dir, "out.mp4")

            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(self.device).float(), randomize_noise=False, return_latents=True
        )
        return images, latents

editor = ImageEditor()

blocks = gr.Blocks()

with blocks:
    gr.Markdown("<h1><center>StyleGAN-NADA</center></h1>")
    gr.Markdown(
        "<h4 style='font-size: 110%;margin-top:.5em'>Inference demo for StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators (SIGGRAPH 2022).</h4>"
    )
    gr.Markdown(
        "<h4 style='font-size: 110%;margin-top:.5em'>Usage</h4><div>Upload an image of your face, pick your desired output styles, and apply StyleGAN-based editing.</div>"
        "<div>Choose the edit image tab to create static images in all chosen styles. Choose the video tab in order to interpolate between all chosen styles</div><div>(To make it easier on the servers, we've limited video length. If you add too many styles (we recommend no more than 3!), they'll pass in the blink of an eye! 🤗)</div>"
    )
    with gr.Row():
        with gr.Column():
            input_img = gr.inputs.Image(type="filepath", label="Input image")
            
        with gr.Column():    
            style_choice = gr.inputs.CheckboxGroup(choices=editor.get_style_list(), type="value", label="Choose your styles!")

    with gr.Row():
        with gr.Column():
            img_button = gr.Button("Edit Image")
            img_output = gr.Gallery(label="Output Images")


        with gr.Column():
                gr.Markdown("Move the sliders to make the chosen attribute stronger (e.g. the person older) or leave at 0 to disable editing.")
                gr.Markdown("If multiple options are provided, they will be used randomly between images (or sequentially for a video), <u>not</u> together.")
                gr.Markdown("Please note that some directions may be entangled. For example, hair length adjustments are likely to also modify the perceived gender.")

                gr.Markdown("For more information about InterFaceGAN, please visit <a href='https://github.com/genforce/interfacegan' target='_blank'>the official repository</a>")

                pose_slider   = gr.Slider(label="Pose", minimum=-1, maximum=1, value=0, step=0.05)
                smile_slider  = gr.Slider(label="Smile", minimum=-1, maximum=1, value=0, step=0.05)
                gender_slider = gr.Slider(label="Perceived Gender", minimum=-1, maximum=1, value=0, step=0.05)
                age_slider    = gr.Slider(label="Age", minimum=-1, maximum=1, value=0, step=0.05)
                hair_slider   = gr.Slider(label="Hair Length", minimum=-1, maximum=1, value=0, step=0.05)

                ig_edit_choices = [pose_slider, smile_slider, gender_slider, age_slider, hair_slider]

    img_button.click(fn=editor.edit_image, inputs=ig_edit_choices + [input_img, style_choice], outputs=img_output)

    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.Markdown(article)

blocks.launch(enable_queue=True)