app.py

import sys
import torch
import pickle
import cv2
import gradio as gr
import numpy as np

from PIL import Image
from collections import defaultdict
from glob import glob

from matplotlib import pyplot as plt
from matplotlib import animation

from easydict import EasyDict as edict
from huggingface_hub import hf_hub_download

sys.path.append("./rome/")
sys.path.append('./DECA')

from rome.infer import Infer
from rome.src.utils.processing import process_black_shape, tensor2image
from rome.src.utils.visuals import mask_errosion

# loading models ---- create model repo
default_modnet_path = hf_hub_download('Pie31415/rome', 'modnet_photographic_portrait_matting.ckpt')
default_model_path = hf_hub_download('Pie31415/rome', 'rome.pth')

# parser configurations
args = edict({
    "save_dir": ".",
    "save_render": True,
    "model_checkpoint": default_model_path,
    "modnet_path": default_modnet_path,
    "random_seed": 0,
    "debug": False,
    "verbose": False,
    "model_image_size": 256,
    "align_source": True,
    "align_target": False,
    "align_scale": 1.25,
    "use_mesh_deformations": False,
    "subdivide_mesh": False,
    "renderer_sigma": 1e-08,
    "renderer_zfar": 100.0,
    "renderer_type": "soft_mesh",
    "renderer_texture_type": "texture_uv",
    "renderer_normalized_alphas": False,
    "deca_path": "DECA",
    "rome_data_dir": "rome/data",
    "autoenc_cat_alphas": False,
    "autoenc_align_inputs": False,
    "autoenc_use_warp": False,
    "autoenc_num_channels": 64,
    "autoenc_max_channels": 512,
    "autoenc_num_groups": 4,
    "autoenc_num_bottleneck_groups": 0,
    "autoenc_num_blocks": 2,
    "autoenc_num_layers": 4,
    "autoenc_block_type": "bottleneck",
    "neural_texture_channels": 8,
    "num_harmonic_encoding_funcs": 6,
    "unet_num_channels": 64,
    "unet_max_channels": 512,
    "unet_num_groups": 4,
    "unet_num_blocks": 1,
    "unet_num_layers": 2,
    "unet_block_type": "conv",
    "unet_skip_connection_type": "cat",
    "unet_use_normals_cond": True,
    "unet_use_vertex_cond": False,
    "unet_use_uvs_cond": False,
    "unet_pred_mask": False,
    "use_separate_seg_unet": True,
    "norm_layer_type": "gn",
    "activation_type": "relu",
    "conv_layer_type": "ws_conv",
    "deform_norm_layer_type": "gn",
    "deform_activation_type": "relu",
    "deform_conv_layer_type": "ws_conv",
    "unet_seg_weight": 0.0,
    "unet_seg_type": "bce_with_logits",
    "deform_face_tightness": 0.0001,
    "use_whole_segmentation": False,
    "mask_hair_for_neck": False,
    "use_hair_from_avatar": False,
    "use_scalp_deforms": True,
    "use_neck_deforms": True,
    "use_basis_deformer": False,
    "use_unet_deformer": True,
    "pretrained_encoder_basis_path": "",
    "pretrained_vertex_basis_path": "",
    "num_basis": 50,
    "basis_init": "pca",
    "num_vertex": 5023,
    "train_basis": True,
    "path_to_deca": "DECA",
    "path_to_linear_hair_model": "data/linear_hair.pth",  # N/A
    "path_to_mobile_model": "data/disp_model.pth",  # N/A
    "n_scalp": 60,
    "use_distill": False,
    "use_mobile_version": False,
    "deformer_path": "data/rome.pth",
    "output_unet_deformer_feats": 32,
    "use_deca_details": False,
    "use_flametex": False,
    "upsample_type": "nearest",
    "num_frequencies": 6,
    "deform_face_scale_coef": 0.0,
    "device": "cuda"
})

# download FLAME and DECA pretrained
generic_model_path = hf_hub_download('Pie31415/rome', 'generic_model.pkl')
deca_model_path = hf_hub_download('Pie31415/rome', 'deca_model.tar')

with open(generic_model_path, 'rb') as f:
    ss = pickle.load(f, encoding='latin1')

    with open('./DECA/data/generic_model.pkl', 'wb') as out:
        pickle.dump(ss, out)

with open(deca_model_path, "rb") as input:
    with open('./DECA/data/deca_model.tar', "wb") as out:
        for line in input:
            out.write(line)

# load ROME inference model
infer = Infer(args)

def image_inference(
    source_img: gr.inputs.Image = None,
    driver_img: gr.inputs.Image = None
):
    out = infer.evaluate(source_img, driver_img, crop_center=False)
    res = tensor2image(torch.cat([out['source_information']['data_dict']['source_img'][0].cpu(),
                                  out['source_information']['data_dict']['target_img'][0].cpu(),
                                  out['render_masked'].cpu(), out['pred_target_shape_img'][0].cpu()], dim=2))
    return res[..., ::-1]

def extract_frames(
    driver_vid: gr.inputs.Video = None
):
    image_frames = []
    vid = cv2.VideoCapture(driver_vid) # path to mp4

    while True:
        success, img = vid.read()

        if not success: break

        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        pil_img = Image.fromarray(img)
        image_frames.append(pil_img)
    
    return image_frames

def video_inference(
    source_img: gr.inputs.Image = None,
    driver_vid: gr.inputs.Video = None
):
    image_frames = extract_frames(driver_vid)

    resulted_imgs = defaultdict(list)

    mask_hard_threshold = 0.5
    N = len(image_frames)
    for i in range(0, N, 4): # frame limits
        new_out = infer.evaluate(source_img, image_frames[i])

        mask_pred = (new_out['pred_target_unet_mask'].cpu() > mask_hard_threshold).float()
        mask_pred = mask_errosion(mask_pred[0].float().numpy() * 255)
        render = new_out['pred_target_img'].cpu() * (mask_pred) + (1 - mask_pred)
            
        normals = process_black_shape(((new_out['pred_target_normal'][0].cpu() + 1) / 2 * mask_pred + (1 - mask_pred) ) )
        normals[normals==0.5]=1.

        resulted_imgs['res_normal'].append(tensor2image(normals))
        resulted_imgs['res_mesh_images'].append(tensor2image(new_out['pred_target_shape_img'][0]))
        resulted_imgs['res_renders'].append(tensor2image(render[0]))

    video = np.array(resulted_imgs['res_renders'])

    fig = plt.figure()
    im = plt.imshow(video[0,:,:,::-1])
    plt.axis('off')
    plt.close() # this is required to not display the generated image

    def init():
        im.set_data(video[0,:,:,::-1])

    def animate(i):
        im.set_data(video[i,:,:,::-1])
        return im

    anim = animation.FuncAnimation(fig, animate, init_func=init, frames=video.shape[0], interval=30)
    anim.save("avatar.gif", dpi=300, writer = animation.PillowWriter(fps=24))
    
    return "avatar.gif"

description = """<p style='text-align: center'> Create a personal avatar from just a single image using ROME. <br> <a href='https://arxiv.org/abs/2206.08343' target='_blank'>Paper</a> | <a href='https://samsunglabs.github.io/rome' target='_blank'>Project Page</a> | <a href='https://github.com/SamsungLabs/rome' target='_blank'>Github</a> </p>"""
quote = """<blockquote style='text-align: center'> [The] system creates realistic mesh-based avatars from a single <strong>source</strong> photo. These avatars are rigged, i.e., they can be driven by the animation parameters from a different <strong>driving</strong> frame. </blockquote>"""

with gr.Blocks() as demo:
    gr.Markdown("# **<p align='center'>ROME: Realistic one-shot mesh-based head avatars</p>**")
    gr.HTML(value="<img src='file/media/tease.gif' alt='Teaser' style='display: block; margin: auto;'>")
    gr.Markdown(description)
    gr.HTML(quote)

    with gr.Tab("Image Inference"):
        with gr.Row():
            source_img = gr.Image(type="pil", label="Source image", show_label=True)
            driver_img =  gr.Image(type="pil", label="Driver image", show_label=True)
        image_output = gr.Image(label="Rendered avatar")
        image_button = gr.Button("Predict")
    with gr.Tab("Video Inference"):
        with gr.Row():
            source_img2 = gr.Image(type="pil", label="Source image", show_label=True)
            driver_vid = gr.Video(label="Driver video")
        video_output = gr.Image(label="Rendered GIF avatar")
        video_button = gr.Button("Predict")

    gr.Examples(
        examples=[
            ["./examples/lincoln.jpg", "./examples/taras2.jpg"],
            ["./examples/lincoln.jpg", "./examples/taras1.jpg"]
        ],
        inputs=[source_img, driver_img],
        outputs=[image_output],
        fn=image_inference,
        cache_examples=True
    )

    image_button.click(image_inference, inputs=[source_img, driver_img], outputs=image_output)
    video_button.click(video_inference, inputs=[source_img2, driver_vid], outputs=video_output)

demo.launch()