S2CFormer Reimplementation

This is an UNOFFICIAL re-implementation of S2CFormer in Pytorch. The original paper is below:

Y. Chen, Q. Li, B. He, D. Feng, R. Wu, Q. Wang, L. Song, G. Lu, and W. Zhang, "S2CFormer: Revisiting the RD-Latency Trade-off in Transformer-based Learned Image Compression," 2025, arXiv:2502.00700.

Installation

git clone http://github.com/tokkiwa/S2CFormer.git --recursive
cd S2CFormer
pip install compressai timm

Note that this code depends on the ELIC repository as a submodule. This ELIC is originally from JiangWeiBeta's implementation, but I added a slight modification for version compatibility.

If you did not clone with --recursive, please run:

git submodule update --init --recursive

Trained models

I trained S2CFormer-Large of four different bitrates on MLIC-Train Dataset for ~2M iterations, following the standard training procedure described in the paper. Note that LR is set to 1e-4, then decayed to 1e-5 at 1.73M iterations, and to 1e-6 at 1.86M iterations.

Checkpoints are available here.

Trainable parameters

Model	Params (This Code)	Official Paper
S2CFormerIdentity	61,525,548	64.63M
S2CFormerConv	64,659,500	66.60M
S2CFormerAttention	67,492,652	68.42M
S2CFormerHybridS	64,638,508	68.00M
S2CFormerHybridM	69,383,212	72.73M
S2CFormerHybridL	76,500,268	79.83M

Quick eval (Kodak)

python3 eval.py \
  --cuda \
  --checkpoint /path/to/hybrid_l_0.013S2CFormerHybridL_checkpoint_best.noctx0.pth.tar \
  --data /path/to/kodak \
  --model_type hybridl \
  --real

Citation

@article{arxiv:2502.00700,
  author  = {Yunuo Chen and Qian Li and Bing He and Donghui Feng and Ronghua Wu and Qi Wang and Li Song and Guo Lu and Wenjun Zhang},
  title   = {S2CFormer: Revisiting the RD-Latency Trade-off in Transformer-based Learned Image Compression},
  journal = {arXiv preprint arXiv:2502.00700},
  year    = {2025},
  url     = {https://arxiv.org/pdf/2502.00700v3}
}