Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation

Foresight proposes adaptive coarse grained reuse framework for accelerating text-to-video generation models while maintaining video quality.

This repository contains the source code implementation of Foresight.

This source code is available under the Apache 2.0 License.

Baseline	Static Reuse	Foresight (Adaptive Reuse)

⚙️ Environment Setup

conda Environment

You can create a new conda environment using script.

conda env create -n foresight-env python=3.10 -y
conda activate foresight-env

pip install -e .

💻 System Requirements

Right now, Foresight has been tested on a 1xA100 node for Open-Sora, Latte and CogVideoX models on single GPU.

We welcome contributions to evaluate Foresight across different models.

🏁 Using Foresight

Foresight Configuration

Foresight requires configuring below parameters to control the warmup phase and reuse phase reuse.

Parameters

• warmup: No of denoising steps used during warmup phase.

  • Type: Integer

• recalculate: Mandatory computation interval.

  • Format: Integer

• threshold: Scaling factor for threshold.

  • Type: Float

Example Configuration

warmup: 5recalculate: 2threshold: 0.5

Example Runs

cd examples/open_sora
python sample.py

cd examples/latte
python sample.py

cd examples/cogvideox
python sample.py

Thank You

Foresight has been implemented on top of VideoSys, an easy and efficient system for video generation.

📝 Citation

@article{foresight,
title={Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation},
author={Adnan, Muhammad and Kurella, Nithesh and Arunkumar, Akhil and Nair, Prashant},
year={2025},
booktitle = {Proceedings of the 39th International Conference on Neural Information Processing Systems},
location = {San Diego, CA, USA}
}