FreeInit : Bridging Initialization Gap in Video Diffusion Models

This repository contains the implementation of the following paper:

FreeInit: Bridging Initialization Gap in Video Diffusion Models
Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, Ziwei Liu

From MMLab@NTU affiliated with S-Lab, Nanyang Technological University

📖 Overview

We propose FreeInit, a concise yet effective method to improve temporal consistency of videos generated by diffusion models. FreeInit requires no additional training and introduces no learnable parameters, and can be easily incorporated into arbitrary video diffusion models at inference time.

🔥 Updates

• [12/2023] Arxiv paper available.
• [12/2023] Code released.
• [12/2023] Project page and video available.

📃 Usage

In this repository, we use AnimateDiff as an example to demonstrate how to integrate FreeInit into current text-to-video inference pipelines.

In pipeline_animation.py, we define a class AnimationFreeInitPipeline inherited from AnimationPipeline, showing how to modify the original pipeline.

In freeinit_utils.py, we provide frequency filtering code for Noise Reinitialization.

An example inference script is provided at animate_with_freeinit.py.

Please refer to the above scripts as a reference when intergrating FreeInit into other video diffusion models.

🔨 Quick Start

1. Clone Repo

git clone https://github.com/TianxingWu/FreeInit.git
cd FreeInit
cd example/AnimateDiff

2. Prepare Environment

conda env create -f environment.yaml
conda activate animatediff

3. Download Checkpoints

Please refer to the official repo of AnimateDiff. The setup guide is listed here.

4. Inference with FreeInit

After downloading the base model, motion module and personalize T2I checkpoints, run the following command to generate animations with FreeInit. The generation results is then saved to outputs folder.

python -m scripts.animate_with_freeinit \
    --config "configs/prompts/freeinit_examples/RealisticVision_v2.yaml" \
    --num_iters 5 \
    --save_intermediate \
    --use_fp16 \

where num_iters is the number of freeinit iterations. We recommend to use 3-5 iterations for a balance between the quality and efficiency. For faster inference, the argument use_fast_sampling can be enabled to use the Coarse-to-Fine Sampling strategy, which may lead to inferior results.

You can change the text prompts in the config file. To tune the frequency filter parameters for better results, please change the filter_params settings in the config file. The 'butterworth' filter with n=4, d_s=d_t=0.25 is set as default. For base models with larger temporal inconsistencies, please consider using the 'guassian' filter.

More .yaml files with different motion module / personalize T2I settings are also provided for testing.

🖼️ Generation Results

Please refer to our project page for more visual comparisons.

🖋️ Citation

If you find our repo useful for your research, please consider citing our paper:

@article{wu2023freeinit,
     title={FreeInit: Bridging Initialization Gap in Video Diffusion Models},
     author={Wu, Tianxing and Si, Chenyang and Jiang, Yuming and Huang, Ziqi and Liu, Ziwei},
     journal={arXiv preprint arXiv:2312.07537},
     year={2023}

🤍 Acknowledgement

This project is distributed under the S-Lab License. See LICENSE for more information.

The example code is built upon AnimateDiff. Thanks the team for their impressive work!