Live2K

Official implementation of Perceiving Better Moments: Cover Frame Reselection and Enhancement for Live Photos with the Live2K Dataset, accepted to ECCV 2026.

Live Photos contain a high-quality cover image and a short burst of video frames. In real smartphones, these two parts are produced by different imaging pipelines: the cover image receives full computational photography processing, while the video frames are usually lower-resolution, compressed, and less color-consistent. When users choose another video frame as the cover, the selected frame often looks much worse than the original cover.

Live Photo imaging pipelines for high-quality cover photos and lower-quality video frames.

This repository studies Live Photo Cover Frame Reselection and Enhancement (LPRE). Given a user-reselected low-quality frame, its adjacent video frames, and the original high-quality cover image as reference, the goal is to reconstruct a high-quality replacement cover frame with improved detail, color, and dynamic range.

LPRE improves a directly reselected cover frame with reference-guided enhancement.

Highlights

• We formulate the LPRE task for Live Photo cover frame reselection and enhancement.
• We build Live2K, a real-world dataset containing 2,042 Live Photos.
• We provide a unified one-stage baseline with:
  • multi-frame temporal fusion,
  • cover-guided color and appearance enhancement,
  • reference-guided super-resolution.

Dataset

Example scenes from the Live2K dataset.

The Live2K dataset can be downloaded from Hugging Face:

The original Live2K dataset can also be downloaded from Baidu Netdisk:

Download Live2K Dataset

Extraction code:

sprq

The dataset loader expects each Live Photo sample to be stored as one subfolder:

```text
Live2K_root/
├── 000001/
│   ├── gt.png
│   ├── ref.png
│   └── lq_sequence/
│       ├── 000.png
│       ├── 001.png
│       ├── ...
│       └── 008.png
├── 000002/
│   ├── gt.png
│   ├── ref.png
│   └── lq_sequence/
│       ├── 000.png
│       ├── ...
│       └── 008.png
└── ...

Where:

• gt.png: high-quality target frame.
• ref.png: original high-quality cover image used as the reference.
• lq_sequence/: nine low-quality adjacent video frames. The code sorts all .png files in this folder, so filenames should sort in temporal order.

After preparing the data, update the dataroot fields in:

options/train/Apple/train_sr_tsa.yml
options/train/OPPO/train_sr_tsa.yml
options/test/Apple/test.yml
options/test/OPPO/test.yml

Repository Structure

Live2K/
├── data/                         # Dataset and dataloader code
├── models/
│   ├── archs/                    # LPENet and network components
│   ├── losses/                   # Training losses
│   ├── train_model.py            # Training model wrapper
│   └── test_model.py             # Testing model wrapper
├── options/
│   ├── train/Apple/train_sr_tsa.yml
│   ├── train/OPPO/train_sr_tsa.yml
│   ├── test/Apple/test.yml
│   ├── test/OPPO/test.yml
│   └── test/test_speed.yml
├── pretrained/                   # Put pretrained weights here
├── train.py
├── test.py
├── test_speed.py
├── train.sh
├── test.sh
└── requirement.txt

Installation

Create a clean environment:

conda create -n live2k python=3.11 -y
conda activate live2k

Install PyTorch and torchvision according to your CUDA version. For example:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Then install the remaining dependencies:

pip install -r requirement.txt

Note: the default training configs use optim_g.type: Muon. The current development environment uses a standalone muon.py module that has no pip package metadata. To train with Muon, install or vendor a compatible MuonWithAuxAdam implementation. Otherwise, change optim_g.type in the training config to Adam.

Pretrained Models

Put pretrained weights in pretrained/. The provided test configs expect:

pretrained/iPhone.pth
pretrained/oppo.pth

You can change path.pretrain_network_g in the corresponding test config if your weights are stored elsewhere.

Training

Train on the Apple split:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \
  --use-env --nproc_per_node=2 --master_port=1145 train.py \
  -opt options/train/Apple/train_sr_tsa.yml --launcher pytorch

Train on the OPPO split:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \
  --use-env --nproc_per_node=2 --master_port=1145 train.py \
  -opt options/train/OPPO/train_sr_tsa.yml --launcher pytorch

Training outputs are saved under:

checkpoint/experiments/<experiment_name>/

Testing

Test the OPPO model:

CUDA_VISIBLE_DEVICES=0 python test.py -opt options/test/OPPO/test.yml

Test the Apple model:

CUDA_VISIBLE_DEVICES=0 python test.py -opt options/test/Apple/test.yml

Run speed evaluation:

CUDA_VISIBLE_DEVICES=0 python test_speed.py -opt options/test/test_speed.yml

Results are saved under:

results/results/<experiment_name>/

Citation

If this project is useful for your research, please cite:

@misc{lou2026perceiving,
  title  = {Perceiving Better Moments: Cover Frame Reselection and Enhancement for Live Photos with the Live2K Dataset},
  author = {Lou, Junyu and Chen, Kai and You, Weiyi and Zeng, Hui and Zhang, Lei and Gu, Shuhang},
  year   = {2026},
  note   = {Project page: https://github.com/CVL-UESTC/Live2K}
}

Keywords

Live Photo, image enhancement, image super-resolution, cover frame reselection, Live2K.