📖TL;DR: Anchor Forcing enables prompt switches to introduce new subjects and actions while preserving context, motion quality, and temporal coherence; prior methods often degrade over time and miss newly specified interactions.
📢 News
[2026-03-18] 🎉 We have officially released the code for public use!
✅ ToDo List for Any-to-Bokeh Release
Release the code
Release the inference pipeline
Release the training files
Release the model weights
:wrench: Installation
We tested this repo on the following setup:
Nvidia GPU with at least 40 GB memory (A100 tested).
Linux operating system.
64 GB RAM.
Other hardware setup could also work but hasn't been tested.
Environment
Create a conda environment and install dependencies:
This repository only provides the training code for step 2. We default to following the training method of LongLive's step 1. Therefore, you can directly train step 2 using LongLive's checkpoints.
📜 Acknowledgement
This codebase builds on LongLive. Thanks for open-sourcing! Besides, we acknowledge following great open-sourcing projects:
MemFlow: We followed its interactive video benchmark.
Self-Forcing: We followed its vbench prompt and checkpoints.
🌏 Citation
bibtex
1@article{yang2026anchor,
2 title={Anchor Forcing: Anchor Memory and Tri-Region RoPE for Interactive Streaming Video Diffusion},
3 author={Yang, Yang and Zhang, Tianyi and Huang, Wei and Chen, Jinwei and Wu, Boxi and He, Xiaofei and Cai, Deng and Li, Bo and Jiang, Peng-Tao},
4 journal={arXiv preprint arXiv:2603.13405},
5 year={2026}
6}
📧 Contact
If you have any questions and improvement suggestions, please email Yang Yang (yangyang98@zju.edu.cn), or open an issue.
