---
license: mit
language:
- en
tags:
- code
pretty_name: upscaler
size_categories:
- 100K<n<1M
task_categories:
- image-to-image
---

![Thumbnail](thumbnail.jpg)

# Dataset Card for Latent Diffusion Super Sampling

Image datasets for building image/video upscaling networks.

This repository contains implementation of training and inference code for models trained on the following works:

## Part 1: Trained Sub-Pixel Convolutional Network for Upscaling on 5000 individual 720p-4K and 1080p-4K image pairs

### References: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network, Shi et al

Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D. and Wang, Z., 2016.

### Results:

720p images tested: 100 <br>
 Average PSNR: 40.44 dB <br>
 
1080p images tested: 100 <br>
 Average PSNR: 43.05 dB <br>

This outperforms the model in the proposed architecture (average is 28.09 dB) <br>

Refer points 5,6,7 in Section "Dataset Description" for further information <br>

## Part 2: Trained Convolutional Neural Network Video Frame Interpolation via Spatially-adaptive Separable Convolution for realtime 4K, 1080p and 720p videos

### References: Video Frame Interpolation via Adaptive Separable Convolution, Niklaus et al

sniklaus@pdx.edu, mtlong@cs.pdx.edu, fliu@cs.pdx.edu

### Results:

720p Model: <br>
•PSNR: 28.35 <br>
•SSIM: 0.78 <br>

1080p Model: <br>
•PSNR: 29.67 <br>
•SSIM: 0.84 <br>

4K Model: <br>
•PSNR: 33.74 <br>
•SSIM: 0.83 <br>

Refer points 8,9,10 in Section "Dataset Description" for further information <br>

## Part 3: Latent Diffusion Super Sampling coming soon!!!

Stay tuned>>>>>>>>>>>>>>>

## Dataset Details

Consists of 300,000 ground truth 720p and 1080p frames with corresponding 4K output frames

### Dataset Description


1. 4K_part1: Contains first part of 4K images
2. 4K_part2: Contains second part of 4K images
3. 720p: Contains 100,000 ground truth 720p images
4. 1080p: Contains 100,000 ground truth 1080p images
5. Additionally, you will find 2 ESPCN (Efficient Sub Pixel Convolution Network) PyTorch models and a Jupyter Notebook (ESPCN.ipynb), which you can use for retraining or inference.
6. Selected Super Resolution 5000 contains 5000 randomly picked image triplets for 4K, 1080p and 720p images.
7. Super Resolution Test 100 serves as the test dataset for the above training set.
8. In the latest update, 3 FIASC (Frame Interpolation via Adaptive Separable Convolutional) PyTorch models and a Jupyter Notebook (FIASC.ipynb) have been added to be used for retraining or inference.
9. Frame Interpolation Training contains 6416 frames used for training the models, each respectively for 4K, 1080p and 720p.
10. Frame Interpolation Testing contains 1309 frames used for evaluating the models, each respectively for 4K, 1080p and 720p.
 
### Dataset Sources

YouTube

## Uses

Diffusion networks, CNNs, Optical Flow Accelerators, etc.

## Dataset Structure

1. All images are in .jpg format
2. Images are named in the following format: *resolution_globalframenumber.jpg*
3. Resolution refers to either of 3: *720p, 1080p or 4K*
4. Globalframenumber is the frame number of the image under the respective resolution. *eg: 4K_10090.jpg*

### Curation Rationale

1. To build a real-time upscaling network using latent diffusion supersampling.
2. Design algorithms for increasing temporal resolution (framerate up-conversion) of videos in real-time.

## Dataset Card Authors

Alosh Denny

## Dataset Card Contact

aloshdenny@gmail.com