- Introduction
- Features
- Hardware Requirements
- Benchmarks
- Cloning
- Building
- Canary build
- Credits
- Custom Models
- FAQ
REAL Video Enhancer is a redesigned and enhanced version of the original Rife ESRGAN App for Linux. This program offers convenient access to frame interpolation and upscaling functionalities on Windows, Linux and MacOS, and is an alternative to outdated software like Flowframes or enhancr.
- NEW! Windows support. !!! NOTICE !!! The bin can be detected as a trojan. This is a false positive caused by pyinstaller.
- Support for Ubuntu 20.04+ on Executable and Flatpak. (libxcb-cursor0 Required to launch on x11!)
- Discord RPC support for Discord system package and Discord flatpak.
- Scene change detection to preserve sharp transitions.
- Preview that shows latest frame that has been rendered.
- TensorRT and NCNN for efficient inference across many GPUs.
- DEPRICATED MacOS support as of 2.0, this will not be returning due to changes made by apple in MacOS 15.1. Read more here.
| Minimum | Recommended | |
|---|---|---|
| CPU | Dual Core x64 bit | Quad Core x64 bit |
| GPU | Vulkan 1.3 capable device | Nvidia RTX GPU (20 series and up) |
| RAM | 8 GB | 16 GB |
| Storage | 1 GB free (NCNN install only) | 10 GB free (TensorRT install) |
| Operating System | Windows 10/11 64bit | Any modern Linux distro (Ubuntu 20.04+) |
Benchmarks done with 1920x1080 video, default settings.
| RX 6650 XT | |
|---|---|
| rife-v4.6 | 31 fps |
| rife-v4.7 - v4.9 | 28 fps |
| rife-v4.10 - v4.15 | 23 fps |
| rife-v4.16-lite | 31 fps |
| RTX 3080 | |
|---|---|
| rife-v4.6 | 81 fps |
| rife-v4.7 - v4.9 | 65 fps |
| rife-v4.10 - v4.15 | 55 fps |
| rife-v4.22 | 50 fps |
| rife-v4.22-lite | 63 fps |
| RTX 3080 | |
|---|---|
| rife-v4.6 | 270 fps |
| rife-v4.7 - v4.9 | 204 fps |
| rife-v4.10 - v4.15 | 166 fps |
| rife-v4.22 | 140 fps |
| rife-v4.22-lite | 192 fps |
git clone https://github.com/TNTwise/REAL-Video-Enhancer
python3 build.py --build_exe
https://github.com/TNTwise/REAL-Video-Enhancer/releases/tag/prerelease
| Person | For | Link |
|---|---|---|
| NevermindNilas | Some backend and reference code and working with me on many projects | https://github.com/NevermindNilas/ |
| Styler00dollar | RIFE models (4.1-4.5, 4.7-4.12-lite), Sudo Shuffle Span and benchmarking | https://github.com/styler00dollar |
| HolyWu | TensorRT engine generation code, inference optimizations, and RIFE jagged lines fixes | https://github.com/HolyWu/ |
| Rick Astley | Amazing music | https://www.youtube.com/watch?v=dQw4w9WgXcQ |
| Software Used | For | Link |
|---|---|---|
| FFmpeg | Multimedia framework for handling video, audio, and other media files | https://ffmpeg.org/ |
| PyTorch | Neural Network Inference (CUDA/ROCm/TensorRT) | https://pytorch.org/ |
| NCNN | Neural Network Inference (Vulkan) | https://github.com/tencent/ncnn |
| RIFE | Real-Time Intermediate Flow Estimation for Video Frame Interpolation | https://github.com/hzwer/Practical-RIFE |
| rife-ncnn-vulkan | Video frame interpolation implementation using NCNN and Vulkan | https://github.com/nihui/rife-ncnn-vulkan |
| rife ncnn vulkan python | Python bindings for RIFE NCNN Vulkan implementation | https://github.com/media2x/rife-ncnn-vulkan-python |
| GMFSS | GMFlow based Anime VFI | https://github.com/98mxr/GMFSS_Fortuna |
| GIMM | Motion Modeling Realistic VFI | https://github.com/GSeanCDAT/GIMM-VFI |
| ncnn python | Python bindings for NCNN Vulkan framework | https://pypi.org/project/ncnn |
| Real-ESRGAN | Upscaling | https://github.com/xinntao/Real-ESRGAN |
| SPAN | Upscaling | https://github.com/hongyuanyu/SPAN |
| Spandrel | CUDA upscaling model architecture support | https://github.com/chaiNNer-org/spandrel |
| cx_Freeze | Tool for creating standalone executables from Python scripts (Linux build) | https://github.com/marcelotduarte/cx_Freeze |
| PyInstaller | Tool for creating standalone executables from Python scripts (Windows/Mac builds) | https://github.com/pyinstaller/pyinstaller |
| Feather Icons | Open source icons library | https://github.com/feathericons/feather |
| PySceneDetect | Transition detection library for python | https://github.com/Breakthrough/PySceneDetect/ |
| Python Standalone Builds | Backend inference using portable python, helps when porting to different platforms. | https://github.com/indygreg/python-build-standalone |
| Model | Author | Link |
|---|---|---|
| 4x-SPANkendata | Crustaceous D | 4x-SPANkendata |
| 4x-ClearRealityV1 | Kim2091 | 4x-ClearRealityV1 |
| 4x-Nomos8k-SPAN series | Helaman | 4x-Nomos8k-SPAN series |
| 2x-OpenProteus | SiroSky | OpenProteus |
| 2x-AnimeJaNai V3 Sharp | The Database | https://github.com/the-database/mpv-upscale-2x_animejanai |
- Q: What does this program attempt to accomplish?
- A: Fast, efficient and easily accessable video interpolation (Ex: 24->48FPS) and video upscaling (Ex: 1920->3840)
- Q: What backend should I use?
- A:
Modern Nvidia (20 series and up), TensorRT is recommended.
Older Nvidia (10 and 16 series), CUDA is recommended.
Oldest Nvidia (900 series and below), NCNN is recommended.
Modern AMD Linux (6000 seies and up), ROCm is experimental.
Other Cards (AMD/Intel) NCNN is the only backend currently working.
- A:
- Q: Why is it failing to recognize installed backends?
- A: REAL Video Enhancer uses PIP and portable python for inference, this can sometimes have issues installing. Please attempt reinstalling the app before creating an issue.
- Q: Why does it take so long to begin inference?
- A: TensorRT uses advanced optimization at the beginning of inference based on your device, this is only done once per resolution of video inputed.
- Q: Why does the optimization and inference fail?
- A: The most common way an optimization can fail is Limited VRAM There is no fix to this except using CUDA or NCNN instead.
- Q: Why am I getting (Insert Error here)?
- A: ROCM is buggy, please take a look at ROCm Help.
- Q: Why am I getting (Insert Vulkan Error here)?
- A: This usually is an OOM (Out Of Memory) error, this can indicate a weak iGPU or very old GPU, I recommeding trying out the Colab Notebook instead.
