MemCam is a memory-augmented framework for scene-consistent interactive video generation, built on Wan2.1-T2V-1.3B. It treats previously generated frames as external memory and dynamically retrieves the most viewpoint-relevant frames via co-visibility estimation, enabling faithful scene reconstruction even after 360° camera rotations.

Files

File	Description
`dit_step20000.ckpt`	MemCam DiT checkpoint (trained 20k steps)

Usage

bash

git clone https://github.com/newhorizon2005/MemCam.git
cd MemCam

# Download this model
huggingface-cli download newhorizon2005/MemCam dit_step20000.ckpt --local-dir models/MemCam

# Run inference
python inference_memcam.py

Citation

bibtex

@inproceedings{gao2026memcam,
  title     = {MemCam: Memory-Augmented Camera Control for Consistent Video Generation},
  author    = {Gao, Xinhang and Guan, Junlin and Luo, Shuhan and Li, Wenzhuo and Tan, Guanghuan and Wang, Jiacheng},
  booktitle = {International Joint Conference on Neural Networks (IJCNN)},
  year      = {2026}
}