LOCI
LOCI is a hybrid spatial-memory video world model built on Wan2.2-TI2V-5B. When a camera revisits a previously observed region, it reproduces what was there before: in half of the transformer blocks, attention keeps a key-value cache of past observations; in the other half, attention is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with a bounded bank of retained observations, it streams long videos at constant memory.
- Paper: LOCI: Spatial Linear Memory for Streaming World Models
- Code: https://github.com/xiaji2021/LOCI
Models
This repository holds two models with the same architecture, each in its own subfolder:
| subfolder | sampler | steps |
|---|---|---|
LOCI/ | flow_euler | 50 |
LOCI-4step/ | few_step (distilled) | 4 |
Each subfolder contains config.json, the shard index, 3 bf16 shards (about 11 GB, transformer only) and SHA256SUMS. The sampler and the number of steps are read from config.json; --steps N overrides the number of steps.
Download
Download one model:
hf download sum0214/LOCI --include "LOCI-4step/*" --local-dir ./weights # 4-step model
hf download sum0214/LOCI --include "LOCI/*" --local-dir ./weights # 50-step model
The VAE, text encoder and tokenizer come from the Wan2.2 base release (Wan-AI/Wan2.2-TI2V-5B-Diffusers); the code loads them from the Hub.
Run
Install the code from https://github.com/xiaji2021/LOCI, then:
python examples/make_trajectory.py --preset look_around --chunks 16 --output traj.json
python scripts/generate.py --weights ./weights/LOCI-4step --wan Wan-AI/Wan2.2-TI2V-5B-Diffusers \
--image first_frame.png --prompt "A sunlit stone temple courtyard with red lanterns." \
--trajectory traj.json --height 512 --width 768 --output out.mp4
Recommended use
- Full history (default) is recommended for clips up to about one minute.
--history sparseselects the paper's bounded-memory setting: first frame + a view bank + the most recent frames, at constant memory.- Resolution 480x864 or 512x768, one GPU. An 80 GB-class GPU (H100 / H200) is recommended.
Speed
Measured on one NVIDIA H200 with FlashAttention-3: LOCI (50 steps), 480x864, bfloat16, a recorded 300 s camera trajectory. Time is seconds per second of generated video (one chunk = 20 frames = 1.25 s at 16 fps). Memory is the peak GPU memory allocated during generation (text encoding and VAE decoding not included).
| history | s per video second | peak GPU memory | measured over |
|---|---|---|---|
full (dense, default) | 7.0 | 30.0 GiB | the first 40 s (time and memory grow with length) |
bounded (sparse) | 5.0 | 16.2 GiB | 300 s (constant time per chunk and constant memory) |
Citation
@article{xia2026loci,
title={LOCI: Spatial Linear Memory for Streaming World Models},
author={Xia, Ji and Liao, Tingting and Liang, Xuezhi and Li, Hao and Liu, Guangyi},
journal={arXiv preprint arXiv:2609.40222},
year={2026}
}
Acknowledgements
LOCI builds on Wan2.2. We thank the authors of ARL² (Li, Shah and Shang, Attend Locally, Remember Linearly): LOCI adopts their hybrid layer layout and follows their sampling procedure for the recurrent memory.
License
Apache-2.0.