CineScale: Free Lunch in High-Resolution Cinematic Visual Generation

This repository contains the CineScale models presented in the paper CineScale: Free Lunch in High-Resolution Cinematic Visual Generation.

CineScale proposes a novel inference paradigm to enable higher-resolution visual generation. It broadens the scope by enabling high-resolution I2V (Image-to-Video) and V2V (Video-to-Video) synthesis, built atop state-of-the-art open-source video generation frameworks, significantly improving upon existing methods which are prone to repetitive patterns in high-resolution outputs.

Project Page: https://eyeline-labs.github.io/CineScale/ Code & Detailed Usage: https://github.com/Eyeline-Labs/CineScale

Models

CineScale provides a family of models, including Text-to-Video (T2V) and Image-to-Video (I2V) variants, capable of generating videos up to 4K resolution.

Model	Tuning Resolution	Checkpoint	Description
CineScale-1.3B-T2V	1088x1920	Hugging Face	Supports 3K (1632x2880) inference on A100 x 1
CineScale-14B-T2V	1088x1920	Hugging Face	Supports 4K (2176x3840) inference on A100 x 8
CineScale-14B-I2V	1088x1920	Hugging Face	Supports 4K (2176x3840) inference on A100 x 8

Quick Start

To get started, you will need to set up the environment and download the model checkpoints as described in the GitHub repository.

Inference examples for various resolutions and tasks are provided in the GitHub repository's command-line scripts. For instance, to run 2K-resolution text-to-video inference:

# Example for 2K-Resolution Text-to-Video (Base Model Wan2.1-1.3B)
# Single GPU
CUDA_VISIBLE_DEVICES=0 python cinescale_t2v1.3b_single.py
# Multiple GPUs
torchrun --standalone --nproc_per_node=8 cinescale_t2v1.3b.py

Refer to the GitHub repository for more detailed instructions and examples for 3K and 4K video generation.

Citation

If you find our work useful, please consider citing our paper:

@article{qiu2025cinescale,
  title={CineScale: Free Lunch in High-Resolution Cinematic Visual Generation},
  author={Haonan Qiu and Ning Yu and Ziqi Huang and Paul Debevec and Ziwei Liu},
  journal={arXiv preprint arXiv:2508.15774},
  year={2025}
}