## Upscale
Automatically upscale and interpolate video files.

### Table of Contents
- [Usage](#usage)
- [Configuration](#configuration)
- [Stats](#stats)
  - [Real-ESRGAN](#real-esrgan)
  - [RIFE](#rife)
  - [SVT-AV1](#svt-av1)
- [Requirements](#requirements)
  - [Requirements for 1080p, 1080p@48, 4k](#requirements-for-1080p-1080p48-4k)
  - [Requirements for 4k48fps](#requirements-for-4k48fps)
- [Installation](#installation)
  - [Windows](#windows)
  - [Arch Linux](#arch-linux)

## Usage
1. Preparing video file (MKVToolNix)
    - Make sure video file does NOT have an image attachment
    - Video file should have ONE `ass` subtitle file
    - Name video file similar to the following: `Your Title - 01.mkv`
2. Place `.mkv` file inside `0-Source` folder
3. Execute `python Start.py`

## Configuration
The `Start.py` file has two options INSIDE, edit in text editor:
- `INTERPOLATE_4K`: `True` or `False` (wether or not 4k48fps should be outputted)
- `MAX_INPUT_WIDTH`: `'720'` (downscales the video to this width, as 1080p upscaling would be insane)

## Stats

_Note: The VRAM usage is the one reported in `nvidia-smi` and is system wide. Subtract ~1GB for OS._

#### Real-ESRGAN
| Input    | Output    |  VRAM  | RAM    | GPU          | FPS       | OS       |
| :------: | :-------: | :----: | :----: | :----------: | :-------: | :------: |
| 1280x720 | 2120x2880 | ~4.4GB | ~1.8GB | 3090 (@360W) | ~13.56fps | Windows  |
| 1280x720 | 2120x2880 | ~4.4GB | ~2.4GB | 3090 (@350W) | ~13.73fps | Linux 🐧 |

#### RIFE
| Input           | Output          | VRAM   | RAM    | GPU          | FPS    | OS       |
| :-------------: | :-------------: | :----: | :----: | :----------: | :----: | :------: |
| 1080p 23.976fps | 1080p 47.952fps | ~6.6GB | ~4.0GB | 3090 (@360W) | ~60fps | Windows  |
| 1080p 23.976fps | 1080p 47.952fps | ~7.6GB | ~8.4GB | 3090 (@350W) | ~64fps | Linux 🐧 |
| 2160p 23.976fps | 2160p 47.952fps | <span style="color:red">~19.4GB</span> | ~8.0GB | 3090 (@360W) | ~18fps | Windows  |
| 2160p 23.976fps | dnf | n/a    | <span style="color:red">~31.0GB (run out of memory)</span> | 3090 (@360W) | n/a    | Linux 🐧 |

#### SVT-AV1
Settings: CRF 26, Preset 6. See [encodeCDN.py](/utils/encodeCDN.py).

| Input     | RAM     | FPS     | CPU                 | OS       |
| :-------: | :-----: | :-----: | :-----------------: | :------: |
| 1920x1080 | ~3.9GB  | ~76fps  | Ryzen 7900x (@130W) | Windows  |
| 1920x1080 | ~4.8GB  | ~80fps  | Ryzen 7900x (@130W) | Linux 🐧 |
| 3840x2160 | <span style="color:gold">~12.3GB</span> | ~35fps  | Ryzen 7900x (@130W) | Windows |
| 3840x2160 | <span style="color:gold">~13.0GB</span> | ~37fps  | Ryzen 7900x (@130W) | Linux 🐧 |

## Requirements
- **OS**: Windows or Linux
- **Python**: >=3.12 (Windows) or Python 3.13 (Linux)
- **RAM**: Min: 16GB - Recommended: 32GB

#### Requirements for 1080p, 1080p@48, 4k
- **GPU**: NVIDIA GPU with 8GB+ of VRAM

#### Requirements for 4k48fps
- **GPU**: NVIDIA GPU with 24GB of VRAM
- **RAM** (Linux): 48GB (there is a memory leak in VapourSynth)

## Installation

#### Windows
1. Download and install [Python 3.12](https://www.python.org/ftp/python/3.12.8/python-3.12.8-amd64.exe) - Check option to add to path
2. Download and install [Imagemagick](https://imagemagick.org/script/download.php#windows) - Check option to install legacy tools (Required for creating player previews `thumbs.vtt` and `sprite.jpg`)
3. Download and install [MKVToolNix](https://mkvtoolnix.download/downloads.html#windows) - **Add to system path environment manually**
4. Download [ffmpeg](https://www.gyan.dev/ffmpeg/builds/#release-builds) - Either place it inside the MKVToolNix install directory or create a new directory somewhere and **add it to system path environment**
5. Download and install [VapourSynth R72](https://github.com/vapoursynth/vapoursynth/releases) - Check options to install vsrepo, avisynth support not required
6. Install `vsrepo` packages, in Terminal (cmd, NOT Powershell!):
    ```cmd
    :: FFmpeg Source Plugin
    vsrepo.py install ffms2

    :: (optional) Miscfilters Plugin, used for scenedetection in vsrife
    :: Unsure if it even works
    vsrepo.py install misc
    ```
7. Install `pip` packages
    ```cmd
    pip install packaging setuptools wheel
    pip install pymediainfo python-dateutil
    pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
    pip install vsrealesrgan vsrife
    ```
8. Install models used by `vsrealesrgan` and `vsrife`
    ```ps
    python -m vsrealesrgan
    python -m vsrife
    ```

#### Arch Linux
Sadly there is some sort of memory leak in vsrife or vapoursynth under Linux.

If you have 32GB of RAM it should suffice for everything EXCEPT 4k48p, there you probably need 48GB or 64GB of RAM.

```bash
# System Packages
sudo pacman -S python vapoursynth ffms2 ffmpeg mkvtoolnix-cli mediainfo imagemagick

# Setup virtual environment (in upscale dirctory)
python -m venv .venv
source .venv/bin/activate

# Install pip dependencies
pip install typing-extensions packaging setuptools wheel
pip install pymediainfo python-dateutil
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip install vsrealesrgan vsrife

# Install models
python -m vsrealesrgan
python -m vsrife
```