CineScale: Free Lunch in High-Resolution Cinematic Visual Generation

CineScale is an extended work of FreeScale for higher-resolution visual generation, unlocking the 4k video generation!

     

Haonan Qiu, Ning Yu*, Ziqi Huang, Paul Debevec, and Ziwei Liu*
(* Corresponding Author)

From Nanyang Technological University and Netflix Eyeline Studios.

.
(Click to enjoy 4k demo 📹)

⚙️ Setup

Install Environment via Anaconda

git clone https://github.com/Eyeline-Labs/CineScale.git
cd CineScale

conda create -n cinescale python=3.10
conda activate cinescale
pip install -e .
pip install xfuser>=0.4.3
pip install flash-attn==2.7.4.post1 --no-build-isolation

🧰 Models

Model	Tuning Resolution	Checkpoint	Description
CineScale-1.3B-T2V (Text2Video)	1088x1920	Hugging Face	Support 3k(1632x2880) inference on A100 x 1
CineScale-14B-T2V (Text2Video)	1088x1920	Hugging Face	Support 4k(2176x3840) inference on A100 x 8
CineScale-14B-I2V (Text2Video)	1088x1920	Hugging Face	Support 4k(2176x3840) inference on A100 x 8

💫 Inference with Command

Model Preparation

Download the checkpoint from Hugging Face and put it the folder models.

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

3K-Resolution Text-to-Video (Base Model Wan2.1-1.3B)

torchrun --standalone --nproc_per_node=8 cinescale_t2v1.3b_pro.py

4K-Resolution Text-to-Video (Base Model Wan2.1-14B)

torchrun --standalone --nproc_per_node=8 cinescale_t2v14b_pro.py

4K-Resolution Image-to-Video (Base Model Wan2.1-14B)

# May set attention_coef to 1.5 for better results (line 123, diffsynth/distributed/xdit_context_parallel.py)

torchrun --standalone --nproc_per_node=8 cinescale_i2v14b.py

🤗 Acknowledgements

This codebase is built on top of the open-source implementation of Wan2.1 based on DiffSynth-Studio repo.

😉 Citation

@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}
}