SlotMem: Character-Addressable Internal Memory for Narrative Long Video Generation

Official checkpoints for SlotMem, a character-addressable internal memory framework for multi-character narrative long video generation. SlotMem represents recurring characters with compact role-wise memory slots, updates them during autoregressive generation, and injects them only into localized character-relevant tokens.

• Repository: YilaiLiu-HKU/SlotMem
• Paper: SlotMem: Character-Addressable Internal Memory for Narrative Long Video Generation

Checkpoints

This repository provides the SlotMem LoRA checkpoints for the two-stage model:

ckpt/
  stage1/
    stage1_low.pt
    stage1_high.pt
  stage2/
    stage2_low.pt
    stage2_high.pt

Due to limited training data scaling, we strongly recommend training SlotMem checkpoints on your own target data before running final inference.

Download all checkpoints:

huggingface-cli download YilaiLiu-HKU/SlotMem --local-dir . --include "ckpt/*"

Usage

Download the SlotMem checkpoints into the code repository:

huggingface-cli download YilaiLiu-HKU/SlotMem --local-dir . --include "ckpt/*"

For training, data curation, benchmarking, and sample data format, please refer to the GitHub repository.

Citation

@article{liu2026slotmem,
  title={SlotMem: Character-Addressable Internal Memory for Narrative Long Video Generation},
  author={Liu, Yilai and others},
  journal={arXiv preprint arXiv:2607.15772},
  year={2026}
}