utils/gradio_utils.py

from streaming_svd_inference import StreamingSVD
from lib.farancia import IImage

import datetime
from pathlib import Path
import os
import ast
from typing import Tuple 
import numpy as np
from PIL import Image


def get_uuid(asset: str,cache: str) -> Path:
    """
    Generate a unique filename based on the current timestamp and save it in the specified root folder.
    Root Folder will be under environment variable GRADIO_TEMP_DIR, if specified, otherwise the current working directory.    
    Args:
        root_fol (str): The root folder where the file will be saved. 

    Returns:
        Path: The path to the saved file.

    """
    file_name = "_".join(
        "_".join(str(datetime.datetime.now()).split('.')).split(" "))+".mp4"
    file = Path(cache) / asset / file_name
    if not file.parent.exists():
        file.parent.mkdir(parents=True)
    print(f"Saving file to {file}")
    return file


def retrieve_intermediate_data(video_file: str) -> Tuple[list[int],list[int],Image.Image]:
    """
    Retrieve intermediate data related to a video file, including expansion size, original size, and outpainted image.

    Args:
        video_file (str): The path to the video file with "__cropped__" in its name.

    Returns:
        Tuple[list[int], list[int], Image.Image]: A tuple containing the expansion size, original size, and outpainted image.

    Raises:
        AssertionError: If the video file path is not a string or does not contain "__cropped__" in its name.

    """
    assert isinstance(video_file,str) and "__cropped__" in video_file,f"File {video_file} is missing __cropped__ keyword"

    video_file_expanded = video_file.replace(
        "__cropped__", "__expanded__")

    # get the expansion size to obtain 16:9 aspect ratio
    expanded_size = ast.literal_eval(Path(video_file_expanded.replace(
        "__expanded__", "__meta_expanded_size__").replace("mp4", "txt")).read_text())
    # get the original size
    orig_size = ast.literal_eval(Path(video_file_expanded.replace(
        "__expanded__", "__meta_orig_size__").replace("mp4", "txt")).read_text())
    # get the outpainted image
    scaled_outpainted_image = IImage.open(video_file_expanded.replace(
        "__expanded__", "__anchor__").replace("mp4", "png")).numpy()
    return expanded_size, orig_size, scaled_outpainted_image


def save_intermediate_data(video: np.ndarray, user_image: np.ndarray, video_path: Path, expanded_size: list[int], fps: int, scaled_outpainted_image: Image.Image):
    """
    Save intermediate data related to the generated video, including resolution information and scaled outpainted image.

    Args:
        video (np.ndarray): The generated video.
        user_image (np.ndarray): The user image used for generating the video.
        video_path (Path): The path to the generated video file.
        expanded_size (list[int]): The expansion size information.
        fps (int): The frames per second of the video.
        scaled_outpainted_image (Image.Image): The scaled outpainted image.

    """
    # save resolution of outpainting (before scaling)
    meta = video_path.parent / \
        ("__meta_expanded_size__"+video_path.name.replace("mp4", "txt"))
    meta.write_text(str(expanded_size))

    # save original resolution of user image
    meta = video_path.parent / \
        ("__meta_orig_size__"+video_path.name.replace("mp4", "txt"))
    meta.write_text(str([user_image.shape[1], user_image.shape[0]]))

    # save scaled outpainted first frame
    anchor = video_path.parent / \
        ("__anchor__"+video_path.name.replace("mp4", "png"))
    IImage(scaled_outpainted_image).save(anchor)

    # save video generated from outpainted image
    video_path_expanded = video_path.parent / \
        ("__expanded__" + video_path.name)
    IImage(video, vmin=0, vmax=255).setFps(fps).save(video_path_expanded)


def image_to_video_gradio(img: np.ndarray, streaming_svd: StreamingSVD, gradio_cache: str, fps: int =24, asset: str="first_stage",  **kwargs: dict) -> str:
    """
    Convert an image to a video using the provided streaming_svd object and perform additional processing steps.

    Args:
        img: The input image to convert to video.
        streaming_svd: The object used for converting the image to video.
        fps (int, optional): The frames per second of the output video (default is 24).
        root_fol (str, optional): The root folder where the video will be saved (default is "first_stage").
        **kwargs: Additional keyword arguments to pass to the streaming_svd object.

    Returns:
        str: The path to the saved cropped video file.

    Note: We save several additional files to hard-drive using a path derived from the cropped video file.
        * image-to-video result using outpainted image (key = __cropped__ )
        * the size of the outpainted image (key = __meta_expanded_size__ )
        * the size of the input image (key = __meta_orig_size__ )
        * the input image (key = __anchor__ )
    """

    video, scaled_outpainted_image, expanded_size = streaming_svd.image_to_video(img,  **kwargs)
    video_path = get_uuid(asset,cache=gradio_cache)
    video_path_cropped = video_path.parent / ("__cropped__" + video_path.name)
    IImage(video, vmin=0, vmax=255).resize(expanded_size[::-1]).crop(
        (0, 0, img.shape[1], img.shape[0])).setFps(fps).save(video_path_cropped)
    save_intermediate_data(video=video, video_path=video_path, expanded_size=expanded_size,
                           fps=fps, user_image=img, scaled_outpainted_image=scaled_outpainted_image)
    return video_path_cropped.as_posix()


def image_to_video_vfi_gradio(img: np.ndarray, streaming_svd: StreamingSVD, gradio_cache: str, fps: int =24, asset: str="first_stage", num_frames: int=None,  **kwargs: dict) -> str:
    """
    Convert an image to a video using the provided streaming_svd object and perform additional processing steps. Then applies VFI

    Args:
        img: The input image to convert to video.
        streaming_svd: The object used for converting the image to video.
        fps (int, optional): The frames per second of the output video (default is 24).
        root_fol (str, optional): The root folder where the video will be saved (default is "first_stage").
        **kwargs: Additional keyword arguments to pass to the streaming_svd object.

    Returns:
        str: The path to the saved cropped video file.

    Note: We save several additional files to hard-drive using a path derived from the cropped video file.
        * image-to-video result using outpainted image (key = __cropped__ )
        * the size of the outpainted image (key = __meta_expanded_size__ )
        * the size of the input image (key = __meta_orig_size__ )
        * the input image (key = __anchor__ )
    """

    video, scaled_outpainted_image, expanded_size = streaming_svd.image_to_video(img, num_frames = (num_frames+1) // 2,  **kwargs)
    video = streaming_svd.interpolate_video(video, dest_num_frames=num_frames)

    video_path = get_uuid(asset,cache=gradio_cache)
    video_path_cropped = video_path.parent / ("__cropped__" + video_path.name)
    IImage(video, vmin=0, vmax=255).resize(expanded_size[::-1]).crop(
        (0, 0, img.shape[1], img.shape[0])).setFps(fps).save(video_path_cropped)
    save_intermediate_data(video=video, video_path=video_path, expanded_size=expanded_size,
                           fps=fps, user_image=img, scaled_outpainted_image=scaled_outpainted_image)
    return video_path_cropped.as_posix()


def text_to_image_gradio(prompt: str, streaming_svd: StreamingSVD, **kwargs: dict) -> np.ndarray:
    """
    Generate an image from the provided text prompt using the specified streaming_svd object.

    Args:
        prompt (str): The text prompt used to generate the image.
        streaming_svd (StreamingSVD): The object used for converting the text to an image.
        **kwargs (dict): Additional keyword arguments to pass to the streaming_svd object.

    Returns:
        np.ndarray: The generated image based on the text prompt.

    """
    return streaming_svd.text_to_image(prompt, **kwargs)


def enhance_video_vfi_gradio(img: np.ndarray, video : str, expanded_size: list[int], num_frames: int,gradio_cache:str, streaming_svd: StreamingSVD, fps: int = 24, asset="second_stage", orig_size: list[int] = None, **kwargs: dict) -> str:
    """
    Enhance a video by applying our proposed enhancement (including randomized blending) to the video.

    Args:
        img (np.ndarray): The input image used for enhancing the video.
        video (str): The path to the input video to be enhanced.
        expanded_size (list[int]): The size to which the video will be expanded.
        streaming_svd (StreamingSVD): The object used for enhancing the video.
        fps (int, optional): The frames per second of the output video (default is 24).
        root_fol (str, optional): The root folder where the enhanced video will be saved (default is "second_stage_preview").
        orig_size (list[int], optional): The original size of the image (default is None).
        **kwargs (dict): Additional keyword arguments to pass to the streaming_svd object for enhancement.

    Returns:
        str: The path to the saved enhanced video file.

    """
    video_enh = streaming_svd.enhance_video(image=img, video=video, num_frames=(num_frames+1) // 2, **kwargs)
    video_int = streaming_svd.interpolate_video(video_enh, dest_num_frames=num_frames)
    video_path = get_uuid(asset, cache=gradio_cache)
    IImage(video_int, vmin=0, vmax=255).resize(
        expanded_size[::-1]).crop((0, 0, orig_size[0], orig_size[1])).setFps(fps).save(video_path)
    return video_path.as_posix()