Enhancing Perceptual Quality in Video Super-Resolution through Temporally-Consistent Detail Synthesis using Diffusion Models
Accepted to ECCV 2024
In this paper, we address the problem of enhancing perceptual quality in video super-resolution (VSR) using Diffusion Models (DMs) while ensuring temporal consistency among frames. We present StableVSR, a VSR method based on DMs that can significantly enhance the perceptual quality of upscaled videos by synthesizing realistic and temporally-consistent details. We introduce the Temporal Conditioning Module (TCM) into a pre-trained DM for single image super-resolution to turn it into a VSR method. TCM uses the novel Temporal Texture Guidance, which provides it with spatially-aligned and detail-rich texture information synthesized in adjacent frames. This guides the generative process of the current frame toward high-quality and temporally-consistent results. In addition, we introduce the novel Frame-wise Bidirectional Sampling strategy to encourage the use of information from past to future and vice-versa. This strategy improves the perceptual quality of the results and the temporal consistency across frames. We demonstrate the effectiveness of StableVSR in enhancing the perceptual quality of upscaled videos while achieving better temporal consistency compared to existing state-of-the-art methods for VSR.
The code is based on Python 3.8.17, CUDA 11, and diffusers.
conda create -n stablevsr python=3.8.17 -y
git clone https://github.com/claudiom4sir/StableVSR.git
cd StableVSR
conda activate stablevsr
pip install -r requirements.txt
Download the REDS dataset from here (sharp + low-resolution).
Data are expected to be in the format root/hr/sequences/frames and root/lr/sequences/frames.
Pretrained models are available here. If you run the train or test code, you don't need to download them explicitly as they are fetched with .from_pretrained('claudiom4sir/StableVSR').
Adjust the dataroot options in dataset/config_reds.yaml. Then, adjust the options in train.sh. Use the following command to start training:
bash ./train.sh
python test.py --in_path YOUR_PATH_TO_LR_SEQS --out_path YOUR_OUTPUT_PATH --num_inference_steps 50 --controlnet_ckpt YOUR_PATH_TO_CONTROLNET_CKPT_FOLDER
python eval.py --gt_path YOUR_PATH_TO_GT_SEQS --out_path YOUR_OUTPUT_PATH
Training with the provided configuration requires about 17GB GPU. Evaluation on REDS (320x180 -> 1280x720) about 14.5 GB.
000.mp4
011.mp4
015.mp4
020.mp4
@article{rota2023enhancing,
title={Enhancing Perceptual Quality in Video Super-Resolution through Temporally-Consistent Detail Synthesis using Diffusion Models},
author={Rota, Claudio and Buzzelli, Marco and van de Weijer, Joost},
journal={arXiv preprint arXiv:2311.15908},
year={2023}
}
If you have any questions, please contact me at claudio.rota@unimib.it