LC-Mamba

June 9, 2026 · View on GitHub

Environment Setup

Conda Environment

conda create -n LC_Mamba
conda activate LC_Mamba

PyTorch Installation

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

Library Installation

pip install -r requirements.txt
cd kernels/selective_scan && pip install .

Checkpoints

Checkpoints can be downloaded from the following link:
Download Checkpoints

Please place the checkpoints in ./ckpt/.

Additional quantitative comparison across benchmarks with Test-Time Augmentation (TTA) disabled
Method	Vimeo90K (PSNR/SSIM)	UCF101 (PSNR/SSIM)	Xiph 2K (PSNR/SSIM)	Xiph 4K (PSNR/SSIM)	M.B. (IE)	SNU-FILM Easy (PSNR/SSIM)	SNU-FILM Medium (PSNR/SSIM)	SNU-FILM Hard (PSNR/SSIM)	SNU-FILM Extreme (PSNR/SSIM)	Params (M)	FLOPS (T)
Ours-C	36.10/0.9801	35.38/0.9700	37.12/0.946	34.81/0.908	1.94	40.10/0.9915	36.11/0.9809	30.81/0.9405	25.69/0.8710	4.3	0.27
Ours-E	36.20/0.9802	35.42/0.9699	37.17/0.946	34.99/0.910	1.96	40.15/0.9912	36.18/0.9809	30.89/0.9416	25.81/0.8725	6.7	0.29
Ours-B	36.52/0.9810	35.47/0.9703	37.33/0.947	35.14/0.911	1.90	40.20/0.9909	36.30/0.9810	31.00/0.9417	25.83/0.8722	16.2	1.07

Dataset

To comprehensively evaluate the proposed model's performance under various conditions and resolutions, experiments were conducted using multiple datasets.

Dataset Structure

/data/datasets/
  ├── middlebury
  ├── snufilm
  ├── ucf101
  ├── vimeo_triplet
  ├── Xiph

Dataset Preparation

The following datasets were used:

Vimeo90K dataset:
Consists of frame triplets with a resolution of 448×256. The test set includes 3,782 triplets.
UCF101 dataset:
The test set contains 379 frame triplets selected from DVF, with a resolution of 256×256.
Xiph dataset:
The original images were downsampled to 2K resolution as "Xiph-2K" and cropped centrally to form "Xiph-4K" for testing.
Middlebury OTHER dataset:
The OTHER set, with a resolution of approximately 640×480, was used for testing.
SNU-FILM dataset:
This dataset consists of 1,240 frame triplets with a resolution of approximately 1280×720. It is categorized into four difficulty levels—Easy, Medium, Hard, and Extreme—based on motion magnitude, enabling detailed performance comparisons.

Benchmarks

Run the benchmark using the following command:

Make bench_Ours-C
Make bench_Ours-E
Make bench_Ours-B

This project is distributed under the Apache 2.0 license. It incorporates concepts and code from RIFE, EMA-VFI, and VFIMamba, and users are advised to adhere to the licensing terms of these respective projects.

We extend our gratitude to the authors of these works for their exceptional contributions.