app.py

import os
import torch
from comfy.model_management import CPUState  # Импорт из того же файла

# Отключаем CUDA, чтобы избежать инициализации
os.environ["CUDA_VISIBLE_DEVICES"] = ""
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = ""

# Принудительно устанавливаем CPU режим
import comfy.model_management
comfy.model_management.cpu_state = CPUState.CPU

import random
import sys
from typing import Sequence, Mapping, Any, Union
from PIL import Image
from huggingface_hub import hf_hub_download
import spaces

import subprocess, sys

import gradio
import gradio_client
import gradio as gr

import imageio
import tempfile
import numpy as np

print("gradio version:", gradio.__version__)
print("gradio_client version:", gradio_client.__version__)

hf_hub_download(repo_id="ezioruan/inswapper_128.onnx", filename="inswapper_128.onnx", local_dir="models/insightface")
hf_hub_download(repo_id="martintomov/comfy", filename="facerestore_models/GPEN-BFR-512.onnx", local_dir="models")

hf_hub_download(repo_id="facefusion/models-3.3.0", filename="hyperswap_1a_256.onnx", local_dir="models/hyperswap")
hf_hub_download(repo_id="facefusion/models-3.3.0", filename="hyperswap_1b_256.onnx", local_dir="models/hyperswap")
hf_hub_download(repo_id="facefusion/models-3.3.0", filename="hyperswap_1c_256.onnx", local_dir="models/hyperswap")

hf_hub_download(repo_id="martintomov/comfy", filename="facedetection/yolov5l-face.pth", local_dir="models")
###hf_hub_download(repo_id="darkeril/collection", filename="detection_Resnet50_Final.pth", local_dir="models/facedetection")
hf_hub_download(repo_id="gmk123/GFPGAN", filename="parsing_parsenet.pth", local_dir="models/facedetection")

hf_hub_download(repo_id="MonsterMMORPG/tools", filename="1k3d68.onnx", local_dir="models/insightface/models/buffalo_l")
hf_hub_download(repo_id="MonsterMMORPG/tools", filename="2d106det.onnx", local_dir="models/insightface/models/buffalo_l")
hf_hub_download(repo_id="maze/faceX", filename="det_10g.onnx", local_dir="models/insightface/models/buffalo_l")
hf_hub_download(repo_id="typhoon01/aux_models", filename="genderage.onnx", local_dir="models/insightface/models/buffalo_l")
hf_hub_download(repo_id="maze/faceX", filename="w600k_r50.onnx", local_dir="models/insightface/models/buffalo_l")

hf_hub_download(repo_id="vladmandic/insightface-faceanalysis", filename="buffalo_l.zip", local_dir="models/insightface/models/buffalo_l")


def get_value_at_index(obj: Union[Sequence, Mapping], index: int) -> Any:
    """Returns the value at the given index of a sequence or mapping.

    If the object is a sequence (like list or string), returns the value at the given index.
    If the object is a mapping (like a dictionary), returns the value at the index-th key.

    Some return a dictionary, in these cases, we look for the "results" key

    Args:
        obj (Union[Sequence, Mapping]): The object to retrieve the value from.
        index (int): The index of the value to retrieve.

    Returns:
        Any: The value at the given index.

    Raises:
        IndexError: If the index is out of bounds for the object and the object is not a mapping.
    """
    try:
        return obj[index]
    except KeyError:
        return obj["result"][index]


def find_path(name: str, path: str = None) -> str:
    """
    Recursively looks at parent folders starting from the given path until it finds the given name.
    Returns the path as a Path object if found, or None otherwise.
    """
    # If no path is given, use the current working directory
    if path is None:
        path = os.getcwd()

    # Check if the current directory contains the name
    if name in os.listdir(path):
        path_name = os.path.join(path, name)
        print(f"{name} found: {path_name}")
        return path_name

    # Get the parent directory
    parent_directory = os.path.dirname(path)

    # If the parent directory is the same as the current directory, we've reached the root and stop the search
    if parent_directory == path:
        return None

    # Recursively call the function with the parent directory
    return find_path(name, parent_directory)


def add_comfyui_directory_to_sys_path() -> None:
    """
    Add 'ComfyUI' to the sys.path
    """
    comfyui_path = find_path("ComfyUI")
    if comfyui_path is not None and os.path.isdir(comfyui_path):
        sys.path.append(comfyui_path)
        print(f"'{comfyui_path}' added to sys.path")


def add_extra_model_paths() -> None:
    """
    Parse the optional extra_model_paths.yaml file and add the parsed paths to the sys.path.
    """
    try:
        from main import load_extra_path_config
    except ImportError:
        print(
            "Could not import load_extra_path_config from main.py. Looking in utils.extra_config instead."
        )
        from utils.extra_config import load_extra_path_config

    extra_model_paths = find_path("extra_model_paths.yaml")

    if extra_model_paths is not None:
        load_extra_path_config(extra_model_paths)
    else:
        print("Could not find the extra_model_paths config file.")


add_comfyui_directory_to_sys_path()
add_extra_model_paths()


def import_custom_nodes() -> None:
    """Find all custom nodes in the custom_nodes folder and add those node objects to NODE_CLASS_MAPPINGS

    This function sets up a new asyncio event loop, initializes the PromptServer,
    creates a PromptQueue, and initializes the custom nodes.
    """
    import asyncio
    import execution
    from nodes import init_extra_nodes
    import server

    # Creating a new event loop and setting it as the default loop
    loop = asyncio.new_event_loop()
    asyncio.set_event_loop(loop)

    # Creating an instance of PromptServer with the loop
    server_instance = server.PromptServer(loop)
    execution.PromptQueue(server_instance)

    # Initializing custom nodes
    # Запускаем корутину и ждём её завершения
    loop.run_until_complete(init_extra_nodes())

import_custom_nodes()
from nodes import NODE_CLASS_MAPPINGS

# --- Глобальная загрузка моделей (один раз при старте) ---
loadimage = NODE_CLASS_MAPPINGS["LoadImage"]()
reactorfaceswap = NODE_CLASS_MAPPINGS["ReActorFaceSwap"]()
saveimage = NODE_CLASS_MAPPINGS["SaveImage"]()

#@spaces.GPU
def generate_image_v1(source_image, target_image, target_index, swap_model, face_restore_model, restore_strength):

    with torch.inference_mode():

        # ==============================
        # ===== GIF PROCESSING =========
        # ==============================
        if target_image.lower().endswith(".gif"):

            reader = imageio.get_reader(target_image)
            frames = []
            durations = []

            for i, frame in enumerate(reader):
                frames.append(frame)

                meta = reader.get_meta_data(index=i)
                duration = meta.get("duration", 100)  # ms
                durations.append(duration)

            output_frames = []

            loadimage_source = loadimage.load_image(image=source_image)

            for frame in frames:

                temp_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
                Image.fromarray(frame).save(temp_file.name)

                loadimage_target = loadimage.load_image(image=temp_file.name)

                result = reactorfaceswap.execute(
                    enabled=True,
                    swap_model=swap_model,
                    facedetection="YOLOv5l",
                    face_restore_model=face_restore_model,
                    face_restore_visibility=restore_strength,
                    codeformer_weight=0.5,
                    detect_gender_input="no",
                    detect_gender_source="no",
                    input_faces_index=str(target_index),
                    source_faces_index="0",
                    console_log_level=1,
                    input_image=get_value_at_index(loadimage_target, 0),
                    source_image=get_value_at_index(loadimage_source, 0),
                )

                swapped = get_value_at_index(result, 0)[0]

                # ===== FIX FLOAT32 =====
                if isinstance(swapped, torch.Tensor):
                    swapped = swapped.cpu().numpy()

                if swapped.dtype != np.uint8:
                    swapped = (swapped * 255.0).clip(0, 255).astype(np.uint8)
                # =======================

                output_frames.append(swapped)

                os.unlink(temp_file.name)

            output_path = "output/swapped.gif"

            pil_frames = [Image.fromarray(f) for f in output_frames]

            pil_frames[0].save(
                output_path,
                save_all=True,
                append_images=pil_frames[1:],
                duration=durations,   # 🔥 giữ nguyên duration từng frame
                loop=0
            )

            return output_path

        # ==============================
        # ===== IMAGE PROCESSING =======
        # ==============================
        else:

            loadimage_source = loadimage.load_image(image=source_image)
            loadimage_target = loadimage.load_image(image=target_image)

            result = reactorfaceswap.execute(
                enabled=True,
                swap_model=swap_model,
                facedetection="YOLOv5l",
                face_restore_model=face_restore_model,
                face_restore_visibility=restore_strength,
                codeformer_weight=0.5,
                detect_gender_input="no",
                detect_gender_source="no",
                input_faces_index=str(target_index),
                source_faces_index="0",
                console_log_level=1,
                input_image=get_value_at_index(loadimage_target, 0),
                source_image=get_value_at_index(loadimage_source, 0),
            )

            save_result = saveimage.save_images(
                filename_prefix="ComfyUI",
                images=get_value_at_index(result, 0),
            )

            saved_path = f"output/{save_result['ui']['images'][0]['filename']}"
            return saved_path


def generate_image(source_image, target_image, target_index,
                   swap_model, face_restore_model, restore_strength):

    os.makedirs("output", exist_ok=True)

    with torch.inference_mode():

        # ===============================
        # LOAD SOURCE (chỉ load 1 lần)
        # ===============================
        loadimage_source = loadimage.load_image(image=source_image)
        source_tensor = get_value_at_index(loadimage_source, 0)

        # =========================================
        # FUNCTION: SAFE TENSOR → UINT8 CONVERSION
        # =========================================
        def tensor_to_uint8(img):

            if isinstance(img, torch.Tensor):
                img = img.detach().cpu().float().numpy()

            img = np.array(img)

            # Nếu range 0-1 → scale lên
            if img.max() <= 1.0:
                img = img * 255.0

            img = np.clip(img, 0, 255).astype(np.uint8)

            return img

        # ===============================
        # ===== GIF PROCESSING ==========
        # ===============================
        if target_image.lower().endswith(".gif"):

            reader = imageio.get_reader(target_image)

            frames = []
            durations = []

            for i, frame in enumerate(reader):

                # Convert chắc chắn về RGB
                frame_rgb = Image.fromarray(frame).convert("RGB")
                frames.append(np.array(frame_rgb))

                meta = reader.get_meta_data(index=i)
                durations.append(meta.get("duration", 100))

            reader.close()

            output_frames = []

            for frame in frames:

                # Tạo file tạm
                with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp:
                    Image.fromarray(frame).save(tmp.name)
                    temp_path = tmp.name

                loadimage_target = loadimage.load_image(image=temp_path)
                target_tensor = get_value_at_index(loadimage_target, 0)

                result = reactorfaceswap.execute(
                    enabled=True,
                    swap_model=swap_model,
                    facedetection="YOLOv5l",
                    face_restore_model=face_restore_model,
                    face_restore_visibility=restore_strength,
                    codeformer_weight=0.5,
                    detect_gender_input="no",
                    detect_gender_source="no",
                    input_faces_index=str(target_index),
                    source_faces_index="0",
                    console_log_level=1,
                    input_image=target_tensor,
                    source_image=source_tensor,
                )

                swapped = get_value_at_index(result, 0)[0]
                swapped = tensor_to_uint8(swapped)

                output_frames.append(swapped)

                os.remove(temp_path)

            # ===== SAVE GIF (QUANTIZE FIX) =====
            output_path = "output/swapped.gif"

            pil_frames = []

            for f in output_frames:
                img = Image.fromarray(f).convert("RGB")

                # Quantize chất lượng cao hơn
                img = img.quantize(
                    method=Image.FASTOCTREE,
                    kmeans=0
                )

                pil_frames.append(img)

            pil_frames[0].save(
                output_path,
                save_all=True,
                append_images=pil_frames[1:],
                duration=durations,
                loop=0,
                optimize=False
            )

            return output_path

        # ===============================
        # ===== IMAGE PROCESSING ========
        # ===============================
        else:

            loadimage_target = loadimage.load_image(image=target_image)
            target_tensor = get_value_at_index(loadimage_target, 0)

            result = reactorfaceswap.execute(
                enabled=True,
                swap_model=swap_model,
                facedetection="YOLOv5l",
                face_restore_model=face_restore_model,
                face_restore_visibility=restore_strength,
                codeformer_weight=0.5,
                detect_gender_input="no",
                detect_gender_source="no",
                input_faces_index=str(target_index),
                source_faces_index="0",
                console_log_level=1,
                input_image=target_tensor,
                source_image=source_tensor,
            )

            swapped = get_value_at_index(result, 0)[0]
            swapped = tensor_to_uint8(swapped)

            output_path = "output/swapped.png"
            Image.fromarray(swapped).save(output_path)

            return output_path

if __name__ == "__main__":
    with gr.Blocks() as app:
        with gr.Row():
            with gr.Column():
                # Вложенный Row для групп (группы расположены горизонтально)
                with gr.Row():
                    # Первая группа
                    with gr.Group():
                        source_image = gr.Image(label="Source (Face)", type="filepath")
                        swap_model = gr.Dropdown(
                            choices=["inswapper_128.onnx", "hyperswap_1a_256.onnx", "hyperswap_1b_256.onnx", "hyperswap_1c_256.onnx"],
                            value="hyperswap_1b_256.onnx",
                            label="Swap Model"
                        )
                        face_restore_model = gr.Dropdown(
                            choices=["none", "GPEN-BFR-512.onnx"], 
                            value="none", 
                            label="Face Restore Model"
                        )
                        restore_strength = gr.Slider(
                            minimum=0, 
                            maximum=1, 
                            step=0.05, 
                            value=0.7, 
                            label="Face Restore Strength"
                        )
                    
                    # Вторая группа
                    with gr.Group():
                        target_image = gr.Image(label="Target (Body)", type="filepath")
                        target_index = gr.Dropdown(
                            choices=[0, 1, 2, 3, 4], 
                            value=0, 
                            label="Target Face Index"
                        )
                        gr.Markdown("Index_0 = Largest Face. To switch for another target face - switch to Index_1, Index_2, e.t.c")
                        generate_btn = gr.Button("Generate")  # Кнопка генерации
            
            with gr.Column():
                output_image = gr.Image(label="Generated Image")  # Вывод изображения
                
        with gr.Row():
            with gr.Column(scale=1):
                gr.Markdown("***Hyperswap_1b_256.onnx is the best (in most cases) - but sometimes model produce FAIL swap (do not do any swapping). It's known inner bug.")
                gr.Markdown("***Hyperswap models do not need Face Restorer - use it with None. Inswapper_128 need Face Restorer - use it with GPEN-BFR-512 at strength 0.7-0.8.")
                gr.Markdown("*** This Space uses only CPU. You have unlimited usage in HF Spaces on CPU.")
                gr.Markdown("*** For avoiding queue - duplicate this space to your account (it's free). Top right corner - Three dots - Duplicate this Space. Make them Private. Enjoy!")
                gr.Markdown(
                    "***ComfyUI Reactor Fast Face Swap Hyperswap running directly on Gradio. - "
                    "[How to convert your any ComfyUI workflow to Gradio]"
                    "(https://huggingface.co/blog/run-comfyui-workflows-on-spaces)"
                )
                
                # Связываем клик кнопки с функцией
                generate_btn.click(
                    fn=generate_image,
                    inputs=[source_image, target_image, target_index, swap_model, face_restore_model, restore_strength],
                    outputs=[output_image]
                )
    
    app.launch(share=True)