A subjective and objective video quality dataset for video super-resolution

This repository contains all the data related to the paper "How Accurate are Video Quality Models for Diffusion-Based Video Super-Resolution?"

[Github] [Dataset] [HuggingFace] [arXiv Preprint]

Dataset

The subjective and metric results are organized in the following way:

• The ./metrics/ directory contains all raw outputs from the evaluated quality metrics, stored as .json files.
• The file subjective.csv holds the subjective scores collected from the user study.
• An aggregated version of all results is available in dataset.json for easier access and analysis.

The corresponding video files are provided as lossless transcodes:

Directory	Size	Videos	Description
Reference	~13 GB	6	Original pristine 2160p source videos used for full-reference quality models
Degraded	~4 GB	36	Downscaled (360p / 720p) and encoded (AV1 / DCVC-RT / Uncompressed) inputs for the upscalers
Upscaled	~300 GB	222	Processed 2160p video sequences shown to and rated by participants (including sources)

The files can be downloaded through the link AVT-VQDB-UHD-1-VSR or by using the included script:

chmod +x ./download.sh
./download.sh

Alternatively the files can also be loaded from HuggingFace:

from datasets import load_dataset
dataset = load_dataset("benjaminherb/AVT-VQDB-UHD-1-VSR")

Citation

Please cite the following paper if you use the data provided in this repository.

@inproceedings{herb_avtvqdbuhd1vsr_2026,
    title     = {How Accurate are Video Quality Models for Diffusion-Based Video Super-Resolution?},
    booktitle = {18th International Conference on Quality of Multimedia Experience (QoMEX)},
    author    = {Benjamin Herb and Steve Göring and Alexander Raake and Rakesh Rao Ramachandra Rao},
    year      = {2026},
    address   = {Cardiff, UK},
    note      = {To appear}