app.py

# -*- coding: utf-8 -*-
# Author: Gaojian Wang@ZJUICSR; TongWu@ZJUICSR
# --------------------------------------------------------
# This source code is licensed under the Attribution-NonCommercial 4.0 International License.
# You can find the license in the LICENSE file in the root directory of this source tree.
# --------------------------------------------------------

import sys
import os
os.system(f'pip install grad-cam')
os.system(f'pip install dlib')
import dlib
import argparse
import numpy as np
from PIL import Image
import cv2
import torch
from huggingface_hub import hf_hub_download
import gradio as gr

import models_vit
from util.datasets import build_dataset
from engine_finetune import test_two_class, test_multi_class
import matplotlib.pyplot as plt
from torchvision import transforms
import traceback
from pytorch_grad_cam import (
    GradCAM, ScoreCAM,
    XGradCAM, EigenCAM
)
from pytorch_grad_cam import GuidedBackpropReLUModel
from pytorch_grad_cam.utils.image import show_cam_on_image, preprocess_image


def reshape_transform(tensor, height=14, width=14):
    result = tensor[:, 1:, :].reshape(tensor.size(0), height, width, tensor.size(2))
    result = result.transpose(2, 3).transpose(1, 2)
    return result


def get_args_parser():
    parser = argparse.ArgumentParser('FSFM3C fine-tuning&Testing for image classification', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int, help='Batch size per GPU')
    parser.add_argument('--epochs', default=50, type=int)
    parser.add_argument('--accum_iter', default=1, type=int, help='Accumulate gradient iterations')
    parser.add_argument('--model', default='vit_large_patch16', type=str, metavar='MODEL',
                        help='Name of model to train')
    parser.add_argument('--input_size', default=224, type=int, help='images input size')
    parser.add_argument('--normalize_from_IMN', action='store_true', help='cal mean and std from imagenet')
    parser.set_defaults(normalize_from_IMN=True)
    parser.add_argument('--apply_simple_augment', action='store_true', help='apply simple data augment')
    parser.add_argument('--drop_path', type=float, default=0.1, metavar='PCT', help='Drop path rate')
    parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM', help='Clip gradient norm')
    parser.add_argument('--weight_decay', type=float, default=0.05, help='weight decay')
    parser.add_argument('--lr', type=float, default=None, metavar='LR', help='learning rate')
    parser.add_argument('--blr', type=float, default=1e-3, metavar='LR', help='base learning rate')
    parser.add_argument('--layer_decay', type=float, default=0.75, help='layer-wise lr decay')
    parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR', help='lower lr bound')
    parser.add_argument('--warmup_epochs', type=int, default=5, metavar='N', help='epochs to warmup LR')
    parser.add_argument('--color_jitter', type=float, default=None, metavar='PCT', help='Color jitter factor')
    parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1', metavar='NAME', help='Use AutoAugment policy')
    parser.add_argument('--smoothing', type=float, default=0.1, help='Label smoothing')
    parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT', help='Random erase prob')
    parser.add_argument('--remode', type=str, default='pixel', help='Random erase mode')
    parser.add_argument('--recount', type=int, default=1, help='Random erase count')
    parser.add_argument('--resplit', action='store_true', default=False,
                        help='Do not random erase first augmentation split')
    parser.add_argument('--mixup', type=float, default=0, help='mixup alpha')
    parser.add_argument('--cutmix', type=float, default=0, help='cutmix alpha')
    parser.add_argument('--cutmix_minmax', type=float, nargs='+', default=None, help='cutmix min/max ratio')
    parser.add_argument('--mixup_prob', type=float, default=1.0, help='Probability of performing mixup or cutmix')
    parser.add_argument('--mixup_switch_prob', type=float, default=0.5, help='Probability of switching to cutmix')
    parser.add_argument('--mixup_mode', type=str, default='batch', help='How to apply mixup/cutmix params')
    parser.add_argument('--finetune', default='', help='finetune from checkpoint')
    parser.add_argument('--global_pool', action='store_true')
    parser.set_defaults(global_pool=True)
    parser.add_argument('--cls_token', action='store_false', dest='global_pool',
                        help='Use class token for classification')
    parser.add_argument('--data_path', default='/datasets01/imagenet_full_size/061417/', type=str, help='dataset path')
    parser.add_argument('--nb_classes', default=1000, type=int, help='number of the classification types')
    parser.add_argument('--output_dir', default='', help='path where to save')
    parser.add_argument('--log_dir', default='', help='path where to tensorboard log')
    parser.add_argument('--device', default='cuda', help='device to use for training / testing')
    parser.add_argument('--seed', default=0, type=int)
    parser.add_argument('--resume', default='', help='resume from checkpoint')
    parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='start epoch')
    parser.add_argument('--eval', action='store_true', help='Perform evaluation only')
    parser.set_defaults(eval=True)
    parser.add_argument('--dist_eval', action='store_true', default=False, help='Enabling distributed evaluation')
    parser.add_argument('--num_workers', default=10, type=int)
    parser.add_argument('--pin_mem', action='store_true', help='Pin CPU memory in DataLoader')
    parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
    parser.set_defaults(pin_mem=True)
    parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes')
    parser.add_argument('--local_rank', default=-1, type=int)
    parser.add_argument('--dist_on_itp', action='store_true')
    parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
    return parser


def load_model(select_skpt):
    global ckpt, device, model, checkpoint
    if select_skpt not in CKPT_NAME:
        return gr.update(), "Select a correct model"
    ckpt = select_skpt
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    args.nb_classes = CKPT_CLASS[ckpt]
    model = models_vit.__dict__[CKPT_MODEL[ckpt]](
        num_classes=args.nb_classes,
        drop_path_rate=args.drop_path,
        global_pool=args.global_pool,
    ).to(device)

    args.resume = os.path.join(CKPT_SAVE_PATH, CKPT_PATH[ckpt])
    if os.path.isfile(args.resume) == False:
        hf_hub_download(local_dir=CKPT_SAVE_PATH,
                        local_dir_use_symlinks=False,
                        repo_id='Wolowolo/fsfm-3c',
                        filename=CKPT_PATH[ckpt])
    args.resume = os.path.join(CKPT_SAVE_PATH, CKPT_PATH[ckpt])
    checkpoint = torch.load(args.resume, map_location=device)
    model.load_state_dict(checkpoint['model'], strict=False)
    model.eval()
    global cam
    cam = GradCAM(model=model,
                  target_layers=[model.blocks[-1].norm1],
                  reshape_transform=reshape_transform
                  )
    return gr.update(), f"[Loaded Model Successfully:] {args.resume}] "


def get_boundingbox(face, width, height, minsize=None):
    x1, y1, x2, y2 = face.left(), face.top(), face.right(), face.bottom()
    size_bb = int(max(x2 - x1, y2 - y1) * 1.3)
    if minsize and size_bb < minsize:
        size_bb = minsize
    center_x, center_y = (x1 + x2) // 2, (y1 + y2) // 2
    x1, y1 = max(int(center_x - size_bb // 2), 0), max(int(center_y - size_bb // 2), 0)
    size_bb = min(width - x1, size_bb)
    size_bb = min(height - y1, size_bb)
    return x1, y1, size_bb


def extract_face(frame):
    face_detector = dlib.get_frontal_face_detector()
    image = np.array(frame.convert('RGB'))
    faces = face_detector(image, 1)
    if faces:
        face = faces[0]
        x, y, size = get_boundingbox(face, image.shape[1], image.shape[0])
        cropped_face = image[y:y + size, x:x + size]
        return Image.fromarray(cropped_face)
    return None


def get_frame_index_uniform_sample(total_frame_num, extract_frame_num):
    return np.linspace(0, total_frame_num - 1, num=extract_frame_num, dtype=int).tolist()


def extract_face_from_fixed_num_frames(src_video, dst_path, num_frames=None):
    video_capture = cv2.VideoCapture(src_video)
    total_frames = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
    frame_indices = get_frame_index_uniform_sample(total_frames, num_frames) if num_frames else range(total_frames)
    for frame_index in frame_indices:
        video_capture.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
        ret, frame = video_capture.read()
        if not ret:
            continue
        image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
        img = extract_face(image)
        if img:
            img = img.resize((224, 224), Image.BICUBIC)
            save_img_name = f"frame_{frame_index}.png"
            img.save(os.path.join(dst_path, '0', save_img_name))
    video_capture.release()
    return frame_indices


class TargetCategory:
    def __init__(self, category_index):
        self.category_index = category_index

    def __call__(self, output):
        return output[self.category_index]


def preprocess_image_cam(pil_img,
                         mean=[0.5482207536697388, 0.42340534925460815, 0.3654651641845703],
                         std=[0.2789176106452942, 0.2438540756702423, 0.23493893444538116]):
    img_np = np.array(pil_img)
    img_np = img_np.astype(np.float32) / 255.0
    img_np = (img_np - mean) / std
    img_np = np.transpose(img_np, (2, 0, 1))
    img_np = np.expand_dims(img_np, axis=0)
    return img_np


def FSFM3C_image_detection(image):
    frame_path = os.path.join(FRAME_SAVE_PATH, str(len(os.listdir(FRAME_SAVE_PATH))))
    os.makedirs(frame_path, exist_ok=True)
    os.makedirs(os.path.join(frame_path, '0'), exist_ok=True)
    img = extract_face(image)
    if img is None:
        return 'No face detected, please upload a clear face!'
    img = img.resize((224, 224), Image.BICUBIC)
    img.save(os.path.join(frame_path, '0', "frame_0.png"))
    args.data_path = frame_path
    args.batch_size = 1
    dataset_val = build_dataset(is_train=False, args=args)
    sampler_val = torch.utils.data.SequentialSampler(dataset_val)
    data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=args.batch_size,
                                                  num_workers=args.num_workers, pin_memory=args.pin_mem,
                                                  drop_last=False)

    if CKPT_CLASS[ckpt] > 2:
        frame_preds_list, video_pred_list = test_multi_class(data_loader_val, model, device)
        class_names = ['Real or Bonafide', 'Deepfake', 'Diffusion or AIGC generated', 'Spoofing or Presentation-attack']
        avg_video_pred = np.mean(video_pred_list, axis=0)
        max_prob_index = np.argmax(avg_video_pred)
        max_prob_class = class_names[max_prob_index]
        probabilities = [f"{class_names[i]}: {prob * 100:.1f}%" for i, prob in enumerate(avg_video_pred)]
        image_results = f"The largest face in this image may be {max_prob_class} with probability: \n [{', '.join(probabilities)}]"

        # Generate CAM heatmap for the detected class
        use_cuda = torch.cuda.is_available()
        input_tensor = preprocess_image(img,
                                        mean=[0.5482207536697388, 0.42340534925460815, 0.3654651641845703],
                                        std=[0.2789176106452942, 0.2438540756702423, 0.23493893444538116])
        if use_cuda:
            input_tensor = input_tensor.cuda()

        # Dynamically determine the target category based on the maximum probability class
        category_names_to_index = {
            'Real or Bonafide': 0,
            'Deepfake': 1,
            'Diffusion or AIGC generated': 2,
            'Spoofing or Presentation-attack': 3
        }
        target_category = TargetCategory(category_names_to_index[max_prob_class])

        cam = GradCAM(model=model,
                      target_layers=[model.blocks[-1].norm1],
                      reshape_transform=reshape_transform
                      )
        grayscale_cam = cam(input_tensor=input_tensor, targets=[target_category], aug_smooth=False, eigen_smooth=True)
        grayscale_cam = 1 - grayscale_cam[0, :]
        img = np.array(img)
        if img.shape[2] == 4:
            img = img[:, :, :3]
        img = img.astype(np.float32) / 255.0
        visualization = show_cam_on_image(img, grayscale_cam)
        visualization = cv2.cvtColor(visualization, cv2.COLOR_RGB2BGR)

        # Add text overlay to the heatmap
        # text = f"Detected: {max_prob_class}"
        # cv2.putText(visualization, text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
        cam_path = os.path.join(CAM_SAVE_PATH, str(len(os.listdir(CAM_SAVE_PATH))))
        os.makedirs(cam_path, exist_ok=True)
        os.makedirs(os.path.join(cam_path, '0'), exist_ok=True)
        output_path = os.path.join(cam_path, "output_heatmap.png")
        cv2.imwrite(output_path, visualization)
        return image_results, output_path, probabilities[max_prob_index]

    if CKPT_CLASS[ckpt] == 2:
        frame_preds_list, video_pred_list = test_two_class(data_loader_val, model, device)
        if ckpt == 'DfD-Checkpoint_Fine-tuned_on_FF++':
            prob = sum(video_pred_list) / len(video_pred_list)
            label = "Deepfake" if prob <= 0.5 else "Real"
            prob = prob if label == "Real" else 1 - prob
        if ckpt == 'FAS-Checkpoint_Fine-tuned_on_MCIO':
            prob = sum(video_pred_list) / len(video_pred_list)
            label = "Spoofing" if prob <= 0.5 else "Bonafide"
            prob = prob if label == "Bonafide" else 1 - prob
        image_results = f"The largest face in this image may be {label} with probability {prob * 100:.1f}%"
        return image_results, None, None


def FSFM3C_video_detection(video, num_frames):
    try:
        frame_path = os.path.join(FRAME_SAVE_PATH, str(len(os.listdir(FRAME_SAVE_PATH))))
        os.makedirs(frame_path, exist_ok=True)
        os.makedirs(os.path.join(frame_path, '0'), exist_ok=True)
        frame_indices = extract_face_from_fixed_num_frames(video, frame_path, num_frames=num_frames)
        args.data_path = frame_path
        args.batch_size = num_frames
        dataset_val = build_dataset(is_train=False, args=args)
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)
        data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=args.batch_size,
                                                      num_workers=args.num_workers, pin_memory=args.pin_mem,
                                                      drop_last=False)

        if CKPT_CLASS[ckpt] > 2:
            frame_preds_list, video_pred_list = test_multi_class(data_loader_val, model, device)
            class_names = ['Real or Bonafide', 'Deepfake', 'Diffusion or AIGC generated',
                           'Spoofing or Presentation-attack']
            avg_video_pred = np.mean(video_pred_list, axis=0)
            max_prob_index = np.argmax(avg_video_pred)
            max_prob_class = class_names[max_prob_index]
            probabilities = [f"{class_names[i]}: {prob * 100:.1f}%" for i, prob in enumerate(avg_video_pred)]

            frame_results = {f"frame_{frame_indices[i]}": [f"{class_names[j]}: {prob * 100:.1f}%" for j, prob in
                                                           enumerate(frame_preds_list[i])] for i in
                             range(len(frame_indices))}
            video_results = (
                f"The largest face in this image may be {max_prob_class} with probability: \n [{', '.join(probabilities)}]\n \n"
                f"The frame-level detection results ['frame_index': 'probabilities']: \n{frame_results}")
            return video_results

        if CKPT_CLASS[ckpt] == 2:
            frame_preds_list, video_pred_list = test_two_class(data_loader_val, model, device)
            if ckpt == 'DfD-Checkpoint_Fine-tuned_on_FF++':
                prob = sum(video_pred_list) / len(video_pred_list)
                label = "Deepfake" if prob <= 0.5 else "Real"
                prob = prob if label == "Real" else 1 - prob
                frame_results = {f"frame_{frame_indices[i]}": f"{(frame_preds_list[i]) * 100:.1f}%" for i in
                                 range(len(frame_indices))} if label == "Real" else {
                    f"frame_{frame_indices[i]}": f"{(1 - frame_preds_list[i]) * 100:.1f}%" for i in
                    range(len(frame_indices))}

            if ckpt == 'FAS-Checkpoint_Fine-tuned_on_MCIO':
                prob = sum(video_pred_list) / len(video_pred_list)
                label = "Spoofing" if prob <= 0.5 else "Bonafide"
                prob = prob if label == "Bonafide" else 1 - prob
                frame_results = {f"frame_{frame_indices[i]}": f"{(frame_preds_list[i]) * 100:.1f}%" for i in
                                 range(len(frame_indices))} if label == "Bonafide" else {
                    f"frame_{frame_indices[i]}": f"{(1 - frame_preds_list[i]) * 100:.1f}%" for i in
                    range(len(frame_indices))}

            video_results = (f"The largest face in this image may be {label} with probability {prob * 100:.1f}%\n \n"
                             f"The frame-level detection results ['frame_index': 'real_face_probability']: \n{frame_results}")
            return video_results
    except Exception as e:
        return f"Error occurred. Please provide a clear face video or reduce the number of frames."


# Paths and Constants
P = os.path.abspath(__file__)
FRAME_SAVE_PATH = os.path.join(os.path.dirname(P), 'frame')
CAM_SAVE_PATH = os.path.join(os.path.dirname(P), 'cam')
CKPT_SAVE_PATH = os.path.join(os.path.dirname(P), 'checkpoints')
os.makedirs(FRAME_SAVE_PATH, exist_ok=True)
os.makedirs(CAM_SAVE_PATH, exist_ok=True)
os.makedirs(CKPT_SAVE_PATH, exist_ok=True)
CKPT_NAME = [
    '✨Unified-detector_v1_Fine-tuned_on_4_classes',
    'DfD-Checkpoint_Fine-tuned_on_FF++',
    'FAS-Checkpoint_Fine-tuned_on_MCIO',
]
CKPT_PATH = {
    '✨Unified-detector_v1_Fine-tuned_on_4_classes': 'finetuned_models/Unified-detector/v1_Fine-tuned_on_4_classes/checkpoint-min_train_loss.pth',
    'DfD-Checkpoint_Fine-tuned_on_FF++': 'finetuned_models/FF++_c23_32frames/checkpoint-min_val_loss.pth',
    'FAS-Checkpoint_Fine-tuned_on_MCIO': 'finetuned_models/MCIO_protocol/Both_MCIO/checkpoint-min_val_loss.pth',
}
CKPT_CLASS = {
    '✨Unified-detector_v1_Fine-tuned_on_4_classes': 4,
    'DfD-Checkpoint_Fine-tuned_on_FF++': 2,
    'FAS-Checkpoint_Fine-tuned_on_MCIO': 2
}
CKPT_MODEL = {
    '✨Unified-detector_v1_Fine-tuned_on_4_classes': 'vit_base_patch16',
    'DfD-Checkpoint_Fine-tuned_on_FF++': 'vit_base_patch16',
    'FAS-Checkpoint_Fine-tuned_on_MCIO': 'vit_base_patch16',
}

with gr.Blocks(css=".custom-label { font-weight: bold !important; font-size: 16px !important; }") as demo:
    gr.HTML(
        "<h1 style='text-align: center;'>🦱 Real Facial Image&Video Detection <br> Against Face Forgery (Deepfake/Diffusion) and Spoofing (Presentation-attacks)</h1>")
    gr.Markdown(
        "<b>☉ Powered by the fine-tuned ViT models that is pre-trained from [FSFM-3C](https://fsfm-3c.github.io/)</b> <br> "
        "<b>☉ We do not and cannot access or store the data you have uploaded!</b> <br> "
        "<b>☉ Release (Continuously updating [by [Gaojian Wang/汪高健](https://scholar.google.com/citations?user=tpP4cFQAAAAJ&hl=zh-CN&oi=ao), [Tong Wu/吴桐](https://github.com/Coco-T-T), [Xingtang Luo/罗兴塘](https://github.com/Rox-C)]) </b> <br> <b>[V1.0] 2025/02/22-Current🎉</b>: "
        "1) Updated <b>[✨Unified-detector_v1] for Physical-Digital Face Attack&Forgery Detection, a ViT-B/16-224 (FSFM Pre-trained) detector that could identify Real&Bonafide, Deepfake, Diffusion&AIGC, Spooing&Presentation-attacks facial images or videos </b> ; 2) Provided the selection of the number of video frames (uniformly sampling 1-32 frames, more frames may time-consuming for this page without GPU acceleration); 3) Fixed some errors of V0.1. <br>"
        "<b>[V0.1] 2024/12-2025/02/21</b>: "
        "Create this page with basic detectors [DfD-Checkpoint_Fine-tuned_on_FF++, FAS-Checkpoint_Fine-tuned_on_MCIO] that follow the paper implementation. <br> ")
    gr.Markdown(
        "- Please <b>provide a facial image or video(<100s)</b>, and <b>select the model</b> for detection: <br> <b>[SUGGEST] [✨Unified-detector_v1_Fine-tuned_on_4_classes]</b> a (FSFM Pre-trained) ViT-B/16-224 for Both Real/Deepfake/Diffusion/Spoofing facial images&videos Detection <br> <b>[DfD-Checkpoint_Fine-tuned_on_FF++]</b> for deepfake detection, FSFM ViT-B/16-224 fine-tuned on the FF++_c23 train&val sets (4 manipulations, 32 frames per video) <br> <b>[FAS-Checkpoint_Fine-tuned_on_MCIO]</b> for face anti-spoofing, FSFM ViT-B/16-224 fine-tuned on the MCIO datasets (2 frames per video)")

    with gr.Row():
        ckpt_select_dropdown = gr.Dropdown(
            label="Select the Model for Detection ⬇️",
            elem_classes="custom-label",
            choices=['Choose Model Here 🖱️'] + CKPT_NAME + ['continuously updating...'],
            multiselect=False,
            value='Choose Model Here 🖱️',
            interactive=True,
        )
        model_loading_status = gr.Textbox(label="Model Loading Status")
    with gr.Row():
        with gr.Column(scale=5):
            gr.Markdown(
                "### Image Detection (Fast Try: copying image from [whichfaceisreal](https://whichfaceisreal.com/))")
            image = gr.Image(label="Upload/Capture/Paste your image", type="pil")
            image_submit_btn = gr.Button("Submit")
            output_results_image = gr.Textbox(label="Detection Result")

            with gr.Row():
                output_heatmap = gr.Image(label="Grad_CAM")
                output_max_prob_class = gr.Textbox(label="Detected Class")
        with gr.Column(scale=5):
            gr.Markdown("### Video Detection")
            video = gr.Video(label="Upload/Capture your video")
            frame_slider = gr.Slider(minimum=1, maximum=32, step=1, value=32, label="Number of Frames for Detection")
            video_submit_btn = gr.Button("Submit")
            output_results_video = gr.Textbox(label="Detection Result")

    gr.HTML(
        '<div style="display: flex; justify-content: center; gap: 20px; margin-bottom: 20px;">'
        '<a href="https://mapmyvisitors.com/web/1bxvi" title="Visit tracker">'
        '<img src="https://mapmyvisitors.com/map.png?d=FYhBoxLDEaFAxdfRzk5TuchYOBGrnSa98Ky59EkEEpY&cl=ffffff">'
        '</a>'
        '</div>'
    )

    ckpt_select_dropdown.change(
        fn=load_model,
        inputs=[ckpt_select_dropdown],
        outputs=[ckpt_select_dropdown, model_loading_status],
    )
    image_submit_btn.click(
        fn=FSFM3C_image_detection,
        inputs=[image],
        outputs=[output_results_image, output_heatmap, output_max_prob_class],
    )
    video_submit_btn.click(
        fn=FSFM3C_video_detection,
        inputs=[video, frame_slider],
        outputs=[output_results_video],
    )

if __name__ == "__main__":
    args = get_args_parser()
    args = args.parse_args()
    ckpt = '✨Unified-detector_v1_Fine-tuned_on_4_classes'
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    args.nb_classes = CKPT_CLASS[ckpt]
    model = models_vit.__dict__[CKPT_MODEL[ckpt]](
        num_classes=args.nb_classes,
        drop_path_rate=args.drop_path,
        global_pool=args.global_pool,
    ).to(device)
    args.resume = os.path.join(CKPT_SAVE_PATH, CKPT_PATH[ckpt])
    if os.path.isfile(args.resume) == False:
        hf_hub_download(local_dir=CKPT_SAVE_PATH,
                        local_dir_use_symlinks=False,
                        repo_id='Wolowolo/fsfm-3c',
                        filename=CKPT_PATH[ckpt])
    args.resume = os.path.join(CKPT_SAVE_PATH, CKPT_PATH[ckpt])
    checkpoint = torch.load(args.resume, map_location=device)
    model.load_state_dict(checkpoint['model'], strict=False)
    model.eval()

    gr.close_all()
    demo.queue()
    demo.launch()