# Real ESRGAN Application Deployment

This simply denote what I do when I am doploying the *Real ESRGAN* to my WSL distro.

> Reference \:
> * Github Repository
> https://github.com/xinntao/Real-ESRGAN
> * Real-ESRGAN超分演算法效果
> https://blog.csdn.net/bornfree5511/article/details/141547609

# Deploying Hardware

## Hardware List

**CPU** : *E5-2696v3*

**GPUs** :
* *Nvidia Quadro T400*
* *Nvidia Tesla P40*

**RAM** : *non-ECC 2666 16 \* 8 128GB \(run with 2133\)*

**MB** : ASUS X99-AII

**Disks** :
* *Team T-FORCE Z44A7Q 2TB M.2 PCIe Gen4*
* *MSI SPATIUM M480 PRO 4TB Gen4 PCIe SSD*

**Power** : *Seasonic X-series 1050W*

## Important Notes

### CPU

CPU is a very cheap \(2025\/07\/20\) old Xeon which is publish in 2015. E5-2696v3 is the OEM version of E5-2699v3.

    1[|                         2.0%]  10[                          0.0%]    19[|                         2.6%]  28[||||                     10.0%]
    2[||||||                   16.4%]  11[|||                       4.6%]    20[                          0.0%]  29[|                         2.6%]
    3[|||||||||||              30.9%]  12[||                        3.3%]    21[|||                       7.9%]  30[                          0.0%]
    4[|                         1.3%]  13[||                        3.3%]    22[                          0.0%]  31[||                        5.3%]
    5[||||||||                 19.6%]  14[||                        3.9%]    23[|                         0.7%]  32[||                        3.9%]
    6[|                         2.6%]  15[|||||                    14.6%]    24[|                         2.6%]  33[|                         2.6%]
    7[||||                      9.3%]  16[|                         3.3%]    25[|                         2.0%]  34[                          0.0%]
    8[|                         2.6%]  17[||||                      7.3%]    26[||                        3.3%]  35[||||                      8.6%]
    9[|||||||||||||||||||||||||80.5%]  18[|||||                    13.8%]    27[||||                      9.9%]  36[||                        3.3%]
  Mem[||||||                                                  5.21G/110G]   Tasks: 33, 413 thr; 3 running
  Swp[                                                             0K/0K]   Load average: 2.08 2.46 2.56 
                                                                            Uptime: 21:04:59

### GPU

I use Nvidia Tesla P40 to do the computation and notice that despite having 24G of VRAM, the compute power is way lower than my main server Nvidia RTX A5000 * 2 with Nvlink.

Every 2.0s: nvidia-smi                                                                                         DESKTOP-4O8R2UE: Tue Jul 22 17:59:05 2025

Tue Jul 22 17:59:05 2025
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.147.01   Driver Version: 529.19       CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA T400         On   | 00000000:01:00.0 Off |                  N/A |
| 38%   47C    P8    N/A /  31W |    503MiB /  2048MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  Tesla P40           On   | 00000000:03:00.0 Off |                  Off |
| N/A   54C    P0    72W / 250W |    975MiB / 24576MiB |     24%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A     25430      C   /python3.10                     N/A      |
|    1   N/A  N/A     25430      C   /python3.10                     N/A      |
+-----------------------------------------------------------------------------+

Apparently, the GPU is not fully utilized when running `realesr-animevideov3`. However, with other models, it can be fully utilized e.g. `RealESRGAN_x4plus_anime_6B`.

### Disk

# Environment Setting

## WSL Configuration

For WSL system, the following command should be added to `.bashrc`.

export PATH="/usr/local/cuda/bin:$PATH"

export PATH="/usr/lib/wsl/lib:$PATH"

export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"

export LD_LIBRARY_PATH="/usr/lib/wsl/lib":$LD_LIBRARY_PATH

export CUDA_HOME="/usr/local/cuda"

export NUMBA_CUDA_DRIVER="/usr/lib/wsl/lib/libcuda.so.1"

export MESA_D3D12_DEFAULT_ADAPTER_NAME="NVIDIA Tesla P40"
export CUDA_VISIBLE_DEVICES=0

Notice that I have to set `CUDA_VISIBLE_DEVICES` since I don't want Quadro T400 to be seen by WSL.

In order to use lower version of python, I use `Ubuntu 22.04` for my WSL distro. That is the merite of using WSL since you can keep multiple distro spinning and they can share resources without slicing the system hardware.

## Python Configuration

Notice that the python version should not be too high and the Pytorch version has to be lower even if the requirement does not specify a higher bound. The following version of packages is what I found that need to be changed with default install on `Ubuntu 22.04`

* `Python 3.10.12`
* `Pytorch 2.0.1`
* `torchvision 0.15.2`
* `numpy 1.26.4`
* `ffmpeg-python 0.2.0`

# Usage \& Outputs

## Help page of the Python Script

$ python inference_realesrgan_video.py -h
/home/erebus/Real_ESRGAN_venv/.venv/lib/python3.10/site-packages/torchvision/transforms/functional_tensor.py:5: UserWarning: The torchvision.transforms.functional_tensor module is deprecated in 0.15 and will be **removed in 0.17**. Please don't rely on it. You probably just need to use APIs in torchvision.transforms.functional or in torchvision.transforms.v2.functional.
  warnings.warn(
usage: inference_realesrgan_video.py [-h] [-i INPUT] [-n MODEL_NAME] [-o OUTPUT] [-dn DENOISE_STRENGTH]
                                     [-s OUTSCALE] [--suffix SUFFIX] [-t TILE] [--tile_pad TILE_PAD]
                                     [--pre_pad PRE_PAD] [--face_enhance] [--fp32] [--fps FPS]
                                     [--ffmpeg_bin FFMPEG_BIN] [--extract_frame_first]
                                     [--num_process_per_gpu NUM_PROCESS_PER_GPU]
                                     [--alpha_upsampler ALPHA_UPSAMPLER] [--ext EXT]

options:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Input video, image or folder
  -n MODEL_NAME, --model_name MODEL_NAME
                        Model names: realesr-animevideov3 | RealESRGAN_x4plus_anime_6B | RealESRGAN_x4plus |
                        RealESRNet_x4plus | RealESRGAN_x2plus | realesr-general-x4v3Default:realesr-
                        animevideov3
  -o OUTPUT, --output OUTPUT
                        Output folder
  -dn DENOISE_STRENGTH, --denoise_strength DENOISE_STRENGTH
                        Denoise strength. 0 for weak denoise (keep noise), 1 for strong denoise ability. Only
                        used for the realesr-general-x4v3 model
  -s OUTSCALE, --outscale OUTSCALE
                        The final upsampling scale of the image
  --suffix SUFFIX       Suffix of the restored video
  -t TILE, --tile TILE  Tile size, 0 for no tile during testing
  --tile_pad TILE_PAD   Tile padding
  --pre_pad PRE_PAD     Pre padding size at each border
  --face_enhance        Use GFPGAN to enhance face
  --fp32                Use fp32 precision during inference. Default: fp16 (half precision).
  --fps FPS             FPS of the output video
  --ffmpeg_bin FFMPEG_BIN
                        The path to ffmpeg
  --extract_frame_first
  --num_process_per_gpu NUM_PROCESS_PER_GPU
  --alpha_upsampler ALPHA_UPSAMPLER
                        The upsampler for the alpha channels. Options: realesrgan | bicubic
  --ext EXT             Image extension. Options: auto | jpg | png, auto means using the same extension as
                        inputs

## Sample Output \#1

python inference_realesrgan_video.py -i "../inputs/1.mp4" -o "../HR/1.mp4" -n RealESRGAN_x4plus_anime_6B


inference: 44544frame [4:32:00,  2.73frame/s]  

## Sample Output \#2

$ python inference_realesrgan_video.py -i "../inputs/2.mp4" -o "../HR/2.mp4" -n realesr-animevideov3
/home/erebus/Real_ESRGAN_venv/.venv/lib/python3.10/site-packages/torchvision/transforms/functional_tensor.py:5: UserWarning: The torchvision.transforms.functional_tensor module is deprecated in 0.15 and will be **removed in 0.17**. Please don't rely on it. You probably just need to use APIs in torchvision.transforms.functional or in torchvision.transforms.v2.functional.
  warnings.warn(
Downloading: "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth" to /home/erebus/Real_ESRGAN_venv/Real-ESRGAN/weights/realesr-animevideov3.pth

100%|███████████████████████████████████████████████████████████████████████| 2.39M/2.39M [00:00<00:00, 8.34MB/s]
inference: 44540frame [1:28:06,  8.43frame/s]                                                           

## Summary

Apparently, different model will takeout different size of vram memory and utilization of GPU.

# Python Script for Running with a Batch of Files in a Folder

[Github Repository](https://github.com/Chen-KaiTsai/Batch-Real-ESRGAN-Video-with-Other-CV-Tools)