I need to incorporate a small amount of 1080p HD shots in a 2160p 4K composition. I'm not impressed by default upscaling techniques like lanczos/bicubic/trilinear; it seems like my television does a better job of upscaling 1080p in real time. It may be time to learn some new tricks.

The FFmpeg manual mention a technique called Super Resolution, and it comes in two forms:

  • SRCNN: Super-Resolution Convolutional Neural Network model
  • ESPCN: Efficient Sub-Pixel Convolutional Neural Network model

Usually I can find many online sources for figuring out how to do something with FFmpeg. But I can't seem to find any tutorials on how to use this. I don't really understand it. I need to train (how?) a model (from what?), or get a pre-trained model (where?). I'd like to know how to get from 1080p to 2160p using this technique, assuming I know nothing about it.

4 Answers 4


I realise this question is pretty old now but it still comes up quite high in search results so I would like to document how I got the "sr" to work (August 2020) in case it can help someone else. Before proceeding it is worth saying that I was not overly impressed by the result. I felt on my test videos that lanczos did an ever so slightly better job. Bear in mind that my test was upscaling from 640x360 flv file to 1920x1080 mp4. I have not tested many other videos and I also have quite a lot to learn about SR like if I feed it different training material will I get a better result.

What you will need

  • A CUDA enabled GPU with a capability of at least 6.0. I worked all the way through this once with a GeForce GT710 (capability of 3.5) only to get to the final ffmpeg upscaling for it to tell me I needed a capability of 6.0. It did proceed to use the CPU but was very slow. I then went through it again with a GeForce GTX 1050TI 4GB card and saw a noticable performance gain.
  • At least 70GB free disk space. The training scripts download about 42GB worth of videos and 9GB worth of images. The rest is used up generating the model.
  • Patience. Generating the datasets took a good few hours and training just one model took nearly the whole day.
  • I had a couple of spare computers (one server and one pc) so I was able to build this from scratch and let it run. First I tried it on Ubuntu Server 20.04 and 18.04. Final run through with the new GPU was on Xubuntu 20.04. The only reason I went with Xubuntu in the end was because I was also investigating a new USB capture device for VHS but the instructions below will work with Server.



My starting point is a fresh install of Xubuntu 20.04. I have a GeForce GTX 1050TI GPU that is not my main GPU and there is no monitor connected to it. I do not think it should matter though if it is your main GPU or if there is a monitor connected.

FROM https://www.tensorflow.org/install/gpu

Add NVIDIA package repositories

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.1.243-1_amd64.deb
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub
sudo dpkg -i cuda-repo-ubuntu1804_10.1.243-1_amd64.deb
sudo apt-get update
wget http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
sudo apt install ./nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
sudo apt-get update

Install NVIDIA driver

sudo apt-get install --no-install-recommends nvidia-driver-450

REBOOT, then check that GPUs are visible using the command: nvidia-smi. Do not proceed if you do not see a table output. If you have a message saying there is no device then you need to troubleshoot this stage. For example in my first build the BIOS was set to use the onboard graphics card first and changing it to use the offboard card got it working. I didn't have this problem in the last build on a different PC.

Install development and runtime libraries (~4GB). Here we are installing both CUDA 10.0, 10.1. The training scripts and filter are a couple of years old now and we therefore need to accommodate them.

sudo apt-get install --no-install-recommends \
    cuda-10-0 \
    libcudnn7=  \

Install TensorRT. Requires that libcudnn7 is installed above.

sudo apt-get install --no-install-recommends libnvinfer6=6.0.1-1+cuda10.0 \
    libnvinfer-dev=6.0.1-1+cuda10.0 \

And now the 10.1 versions

sudo apt-get install --no-install-recommends \
    cuda-10-1 \
    libcudnn7=  \

sudo apt-get install --no-install-recommends libnvinfer6=6.0.1-1+cuda10.1 \
    libnvinfer-dev=6.0.1-1+cuda10.1 \

FROM https://www.tensorflow.org/install/pip#system-install

The guide above suggests to install venv here but this was not a concern for me so I left it out. If you are familiar with Python then go ahead and use venv.

sudo apt install python3-dev python3-pip
sudo pip3 install --upgrade tensorflow

Add some libraries temporarily to LD_LIBRARY_PATH. We will make this permanent a little later.

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib:/usr/local/cuda-10.0/lib64:/usr/local/cuda-10.1/lib64:/usr/local/cuda-10.2/lib64:/usr/local/cuda/extras/CUPTI/lib64

You will notice I am adding cuda 10.2 here as well but we did not install it. For some reason one of the library files (libcublas) when installing 10.1 is taken from 10.2.

We can verify with the below (check for errors). There is a lot of output, mainly information and you want to check that it has seen your GPU and has successfully opened all the libraries.

python3 -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

Add LD_LIBRARY_PATH to ld.so.conf.d

sudo vi /etc/ld.so.conf.d/cuda-additional.conf

Add the lines


The other 2 paths are automatically added for us during install.

FROM https://ffmpeg.org/ffmpeg-filters.html#sr-1 Install the TensorFlow for C library. If you follow the link given at the above it shows you Tensorflow version 2.3.0 (or later), however, as mentioned earlier the filter code is old and requires version 1.15.0. So:

wget https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-linux-x86_64-1.15.0.tar.gz
sudo tar -C /usr/local -xzf libtensorflow-gpu-linux-x86_64-1.15.0.tar.gz

FROM https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu Now we fetch and compile ffmpeg. So install some dependencies first:

sudo apt-get update -qq && sudo apt-get -y install \
  autoconf \
  automake \
  build-essential \
  cmake \
  git-core \
  libass-dev \
  libfreetype6-dev \
  libgnutls28-dev \
  libsdl2-dev \
  libtool \
  libva-dev \
  libvdpau-dev \
  libvorbis-dev \
  libxcb1-dev \
  libxcb-shm0-dev \
  libxcb-xfixes0-dev \
  pkg-config \
  texinfo \
  wget \
  yasm \

mkdir -p ~/ffmpeg_sources ~/bin
sudo apt-get install nasm libx264-dev libx265-dev libnuma-dev libvpx-dev libfdk-aac-dev libmp3lame-dev libopus-dev libunistring-dev

cd ~/ffmpeg_sources && \
wget -O ffmpeg-snapshot.tar.bz2 https://ffmpeg.org/releases/ffmpeg-snapshot.tar.bz2 && \
tar xjvf ffmpeg-snapshot.tar.bz2 && \
cd ffmpeg && \
PATH="$HOME/bin:$PATH" PKG_CONFIG_PATH="$HOME/ffmpeg_build/lib/pkgconfig" ./configure \
  --prefix="$HOME/ffmpeg_build" \
  --pkg-config-flags="--static" \
  --extra-cflags="-I$HOME/ffmpeg_build/include" \
  --extra-ldflags="-L$HOME/ffmpeg_build/lib" \
  --extra-libs="-lpthread -lm" \
  --bindir="$HOME/bin" \
  --enable-gpl \
  --enable-gnutls \
  --enable-libass \
  --enable-libfdk-aac \
  --enable-libfreetype \
  --enable-libmp3lame \
  --enable-libopus \
  --enable-libvorbis \
  --enable-libvpx \
  --enable-libx264 \
  --enable-libx265 \
  --enable-libtensorflow \
  --enable-nonfree && \
PATH="$HOME/bin:$PATH" make && \
make install && \
hash -r

Notice that I have left out libaom and added libtensorflow. This has created new binaries under ~/bin and added it to your profile and path so it can be called from anywhere. If you do not want to replace your existing ffmpeg then be sure to change the bindir and paths as required.

Reload our profile

source ~/.profile

FROM https://github.com/XueweiMeng/sr/tree/sr_dnn_native

So now we are almost at the stage where we can start to generate and train the models, however, the scripts are old so we need to modify them somewhat. Get them first (you can specify any path to download it to or get it manually with GUI:

cd ~
wget https://github.com/XueweiMeng/sr/archive/sr_dnn_native.zip
unzip sr_dnn_native.zip
rm sr_dnn_native.zip
cd sr-sr_dnn_native

Now edit all python files that have "import tensorflow as tf", i.e.


After the imports in each script add the line:

tf = tf.compat.v1

For example using datasets/prepare_dataset.py:

import os
import argparse
from tqdm import tqdm
import cv2
import numpy as np
import json
import tensorflow as tf
from PIL import Image

tf = tf.compat.v1

class SceneChangeDetector:

There is one other depreciated function we need to take care of and that is imresize.

Edit the file datasets/prepare_dataset.py

Replace line 8: from scipy.misc import imresize

With: from PIL import Image

Replace line 127: frame_lr = imresize(frames[k], (lr_h, lr_w), interp='bicubic')

With: frame_lr = np.array(Image.fromarray(frames[k]).resize(size=(lr_h, lr_w)))

We can finally start generating. Install some dependencies.

sudo pip3 install Pillow tqdm opencv-python

Generate (expect it to take a good few hours and you will likely see a lot of information output and warnings of almost depreciated functions, these are ok):

sh generate_datasets.sh

This will download 11 videos totalling around 43GB and images totalling around 9GB and can take a few hours to run depending on hardware.

And now for the training (expect these to take about a day each). The models after this will be found in the extracted ~/sr-sr_dnn_native folder if that is where you saved the Git zip file to:


sh train_srcnn.sh
python3 generate_header_and_model.py --model=srcnn --ckpt_path=logdir/srcnn_batch_32_lr_1e-3_decay_adam/train


sh train_espcn.sh
python3 generate_header_and_model.py --model=espcn --ckpt_path=logdir/espcn_batch_32_lr_1e-3_decay_adam/train

Once the training is complete you can use the filter. It seems as though ESPCN only upscales by a factor of 2 whereas with SRCNN you can specify 2, 3, or 4. I moved the models generated into the same folder as my videos for ease of use but you can simply point to the models wherever they are.


ffmpeg -i <input_video> -vf sr=dnn_backend=tensorflow:scale_factor=3:model=srcnn.pb -q 15 -preset slow <output_video>
ffmpeg -i <input_video> -vf sr=dnn_backend=tensorflow:model=espcn.pb -q 15 -preset slow <output_video>

Here is an interesting video of this in use and was probably one of the only places I have seen anyone use this. In it he suggests it can only work on a single plane but I did not find that to be the case. His results actually appear to be quite good:

This is the ffmpeg command he refers to in the video at 14:44:

ffmpeg -i my_video_540p.mp4 -filter_complex "format=pix_fmts=yuv420p,extractplanes=y+u+v[y][u][v];[y] sr=dnn_backend=tensorflow:scale_factor=2:model=espcn.model [y_scale];[u] scale=960:-2 [u_scale];[v] scale=960:-2 [v_scale];[y_scale][u_scale][v_scale] mergeplanes=0x001020:yuv420p [merged]" -map "[merged]" -c:v libx264 -crf 18 -pix_fmt yuv420p my_video_1080p.mp4
  • 1
    I am happy with your extensive write-up, and I have accepted your answer after reading this. Also I'm a bit sad that you made a lot of effort yet still observe results similar lanczos at best. Great first contribution! Thank you.
    – Redsandro
    Commented Sep 3, 2020 at 11:28
  • 1
    Thank you. I am not giving up on it just yet because looking at "datasets/prepare_dataset.py" from line 166 and also in "train.py" there are many options that we can play with (passed in as parameters) and I believe with the right training I could achieve better results. I'll post back if I manage to get something amazing.
    – Mark
    Commented Sep 3, 2020 at 12:40
  • 1
    Using a linux box with an old nvidia driver (418), I decided to install to a docker image nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04. This meant skipping everything until libtensorflow installation and installing libtensorflow 1.12.3 rather than 1.15.0 (see tensorflow.org/install/source#gpu). Additionally, HighVoltageRocknRoll's repo is more up to date than XueweiMeng's, so I used the former, and since libtensorflow was precompiled for higher compute capability (6) than my GPU (5), I compiled it via tensorflow/tools/ci_build/linux/libtensorflow_gpu.sh from the Tensorflow repo.
    – Witiko
    Commented Sep 14, 2020 at 1:45
  • I really wish this were a container image Commented Sep 16, 2020 at 1:58
  • @mikebabcock Mikuláš and I have prepared a Docker image, see my answer.
    – Witiko
    Commented Oct 10, 2020 at 18:49

Since Google Summer of Code 2018, FFMpeg has supported the sr filter for applying super-resolution methods based on convolutional neural networks. However, as you have discovered, few super-resolution tutorials exist, and compiling FFMpeg with proper libraries and preparing models for super-resolution requires expert knowledge.

To make super-resolution in FFMpeg easier, Mikuláš, Dirk and I have taken the excellent answer of Mark, and we used it to prepare a Docker image with FFMpeg and Libtensorflow. We also wrote a tutorial that explains step-by-step how to use the Docker image for super-resolution in FFMpeg: https://github.com/MIR-MU/ffmpeg-tensorflow#ffmpeg-with-libtensorflow.

After installation, super-resolution in FFMpeg becomes as easy as this:

$ wget https://media.xiph.org/video/derf/y4m/flower_cif.y4m
$ alias ffmpeg-tensorflow='docker run --rm --gpus all -u $(id -u):$(id -g) -v "$PWD":/data -w /data -it miratmu/ffmpeg-tensorflow'
$ ffmpeg-tensorflow -i flower_cif.y4m -filter_complex '
>   [0:v] format=pix_fmts=yuv420p, extractplanes=y+u+v [y][u][v];
>   [y] sr=dnn_backend=tensorflow:scale_factor=2:model=espcn.pb [y_scaled];
>   [u] scale=iw*2:ih*2 [u_scaled];
>   [v] scale=iw*2:ih*2 [v_scaled];
>   [y_scaled][u_scaled][v_scaled] mergeplanes=0x001020:yuv420p [merged]
> ' -map [merged] -sws_flags lanczos \
> -c:v libx264 -crf 17 -preset ultrafast -tune film \
> -c:a copy \
> -y flower_cif_2x.mp4

Compare upscaling using Lanczos (left) with the ESPCN super-resolution model (right):

Comparison of Lanczos and ESPCN


Although Mark has provided great answer, you don't need to train models from scratch, all Super-Resolution models that ffmpeg supports have pre-trained checkpoints: https://github.com/XueweiMeng/sr/tree/sr_dnn_native all you need to do is to convert it to frozen graph (.pb) file and then use it in ffmpeg.

  • 1
    Upvoted. Could you elaborate the "all you need to do" part so that from reading the answer alone, the reader will know what to do?
    – Redsandro
    Commented Apr 7, 2021 at 9:33
  • 3
    Sure, no problem. 1. Download repo with models github.com/XueweiMeng/sr/tree/sr_dnn_native 2. Install all requirements using pip install, for me it was only tensorflow and numpy packages. 3. Run script to generate model you need: python3 generate_header_and_model.py --model=espcn --ckpt_path=checkpoints/espcn/ 4. 2 model files will be generated - espcn.model and espcn.pb (you can check that model works by evaluating .pb model in tensorflow) 5. Use one of these models in ffmpeg with sr filter (or dnn_processing filter), as stated in documentation.
    – Alex
    Commented Apr 7, 2021 at 17:55
  • Note: you might need to install TensorFlow C library and configure ffmpeg if you want to use tensorflow dnn_backend.
    – Alex
    Commented Apr 7, 2021 at 18:00
  • @Alex Very interesting! We provide a Docker image with precompiled tensorflow-capable FFMpeg and .pb files in github.com/MIR-MU/ffmpeg-tensorflow. However, I haven't found much documentation about these .model and .pb files. Are there sources that you would recommend for reading up on the common formats for storing computational graphs?
    – Witiko
    Commented Aug 28, 2021 at 20:03
  • 1
    These guys have ~70 Mb prtrained tflite model , they claim it is optimized to work better with different resolutions. They tiled GAN I guess. Once I tiled ESR-GAN it doesnt show any difference in speed or quality however it introduces some blurred lines at tiling border. So be careful collabora.com/news-and-blog/blog/2020/09/21/… Commented Feb 27, 2022 at 5:55

If all you need to do is superscale an image, going through the trouble of training neural networks is re-inventing the wheel. By all means, if you're studying computer science and are interested in AI/ML, I'd encourage you to look into it, but to just superscale an image, you don't need to train a neural network. There are tools available.

In DaVinci Resolve, for instance, all you need to do is right-click on a clip in the media pool, select "clip attributes...," then at the bottom of the "video" tab, change "Super Scale: None" to 2x, 3x, or 4x. I'm not sure if this feature is available in the free version, but I wouldn't be surprised if it is, because all they've done is integrated the open source ML bits of FFMPEG for you, trained the models for you, and bundled it as a feature. I'm sure Adobe has an equivalent feature, though I haven't used it.

  • 1
    I don't mind using a pre-trained model, but I would like to know how to do this using ffmpeg exclusively so I can incorporate it in a pipeline (e.g. kdenlive). No commercial software.
    – Redsandro
    Commented Jan 21, 2020 at 12:57
  • According to this it's not in the free version (which is DaVinci Resolve, as opposed to DaVinci Resolve Studio). Given that, and given DaVince Resolve Studio costs $300 dollars, and especially given what you claim about them basing their SR on FFmpeg (I would definitely be interested in a source for that claim), FFmpeg is currently the best free tool for the purpose out there. Commented Oct 17, 2021 at 20:50
  • @HashimAziz That’s about R16. R17 has been out for a while. I don’t know if superscale is one of the features that got “unlocked” this time around, but BMD always promotes some previously studio-only features to the free version with every update cycle, so there’s no harm in trying it out. Plus, R18 shouldn’t be too far away. Commented Oct 17, 2021 at 21:31
  • @JasonConrad R19 is now in beta, and Super Scale remains exclusive to (non-free) Resolve Studio.
    – mysteryegg
    Commented May 9 at 22:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.