LOCI · Spatial Linear Memory for Streaming World Models

Look away.Come back.It's still there.

A streaming world model that remembers the places it has seen, so every revisit shows the same world.

LOCI: Spatial Linear Memory for Streaming World Models

Ji Xia¹ · Tingting Liao¹ · Xuezhi Liang¹ · Hao Li^2,3 · Guangyi Liu^1,†

¹ Institute of Foundation Models, Mohamed bin Zayed University of Artificial Intelligence
² Mohamed bin Zayed University of Artificial Intelligence ³ Pinscreen
† Corresponding author

Paper ↗arXiv ↗Code ↗Dataset ↗Checkpoints ↗BibTeX

Abstract

LOCI is a hybrid spatial memory for streaming video world models. Half of the transformer blocks keep a key–value cache of past observations; the other half pair current-chunk attention with a recurrent linear-attention memory whose reads and writes are conditioned on camera geometry. On MIND and held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models, and with a bounded bank of retained observations it streams long videos at constant memory.

+0.89dB: PSNR over a same-recipe full-softmax model under an identical bounded KV budget (MIND, all 50 segments; better in 44 of 50)
+1.82dB: PSNR above the strongest streaming world model we run on MIND (HY-WorldPlay)
~30%: lower peak memory than full softmax at equal length, full history; reachable length 157 s → 225 s
300s: streamed at a constant 23.6 GiB with bounded sparse access

Teaser: returning to a previously seen place. LOCI restores the tree and the building; full softmax replaces them. — **Returning to a previously seen place.** Each model continues the recorded video (top: full history; bottom: held-out Tokyo map, same bounded KV budget). Full softmax, trained with the same recipe, replaces the tree with a facade and loses the building; LOCI restores both (PSNR against ground truth; insets magnify the boxed region). Band: a fixed-size recurrent state shapes the queries into the retained KV.

Scroll to move the camera

Start here.

Then. Now.

Same camera pose, a minute apart. On the left, the first time the model drew the view. On the right, the view after a full turn and a walk. Drag to compare.

First seen Revisited

Method

Two traces of the past. Working together.

A camera-conditioned recurrent state summarizes the whole history at fixed size, while retained key–value (KV) features preserve observation-level evidence; the recurrent readout then guides attention toward the relevant retained views.

Key idea of LOCI — **Key idea.** Recurrent memory integrates historical context, while retained KV preserves observation-level detail.

Conceptual illustration of LOCI — **Overview.** At a revisit, the recurrent readout shapes the queries that attend to co-visible retained observations.

Hybrid layoutFifteen blocks add recurrent Kimi Delta Attention (KDA); fifteen keep softmax attention over retained keys and values.
Projective camera encodingPRoPE conditions the recurrent memory on each token's camera ray, so reads depend on viewpoint.
Chunk-synchronous memoryWrites use the delta rule; retention is applied once per chunk, so forgetting follows elapsed video time.
Dense or bounded historyBounded mode attends to at most 34 latent frames, while the recurrent state carries the whole history.

What it keeps.

A fixed-size recurrent state takes in every frame; a bounded set of retained frames keeps sharp detail of the views that cover the scene best.

Sink frameThe first view. Always kept.
1
AnchorsEarly views, pinned and never dropped.
0
Coverage bankOlder views chosen to cover the scene, up to 20.
0
Recent windowThe last 8 frames, for smooth motion.
0
Recurrent stateKDA layers fold every frame into a fixed-size state, read through camera geometry.
0

0 s

Results

Revisits, measured. Against the recording.

Evaluated zero-shot on the MIND memory test over the entire prediction segment; full softmax is the same backbone and recipe with softmax attention in all 30 blocks. Protocol details are in the paper.

MIND memory test

Method	Params	Trained on MIND	n	MSE ↓	PSNR ↑	SSIM ↑	LPIPS ↓
Reference (reported in the GIM-World paper)
SSM	–	✓	–	0.0796	11.96	0.395	0.744
FramePack	–	✓	–	0.0764	12.04	0.376	0.706
Context-as-Memory	–	✓	–	0.0706	12.50	0.392	0.695
GIM-World	1.3B	✓	–	0.0614	13.40	0.414	0.630
Run by us (entire prediction segment)
HY-WorldPlay	8B	×	50	0.0710	12.54	0.398	0.715
Matrix-Game 3.0	5B	×	49	0.0866	11.69	0.373	0.666
AlayaWorld	15B	×	50	0.0908	11.33	0.358	0.683
Alaya-EVOKE	14B	×	50	0.0884	11.14	0.347	0.710
LingBot-World	28B	×	50	0.0955	11.33	0.331	0.703
CaR^‡	5B	×	36	0.0538	13.85	0.458	0.622
LOCI (ours)	5B	×	50	0.0455	14.36	0.464	0.643
LOCI (bounded sparse)	5B	×	50	0.0622	13.02	0.409	0.679

LOCI is best among the models we ran; reference rows are as reported in the GIM-World paper. ^‡CaR: 36 shortest segments only.

Qualitative revisit comparison on held-out recorded trajectories — **Qualitative revisit comparison.** On return, LOCI restores the previously observed scene, whereas the other models replace or distort it.

Hybrid vs. full softmax, same backbone, data and recipe

Data mixture	History	n	PSNR ↑			LPIPS ↓
Data mixture	History	n	Softmax	LOCI	Δ [95% CI]	Softmax	LOCI	Δ [95% CI]
Uniform	bounded	50	12.17	12.85	+0.67 [+0.42, +0.93]	0.700	0.680	−0.020 [−0.028, −0.011]
UE-weighted	bounded	50	12.13	13.02	+0.89 [+0.64, +1.15]	0.704	0.679	−0.024 [−0.034, −0.015]
UE-weighted	full	32	13.90	14.25	+0.35 [+0.08, +0.62]	0.669	0.641	−0.029 [−0.040, −0.018]

Paired differences (LOCI − full softmax) with 95% bootstrap intervals.

**Long-range recall under the same bounded KV budget.** LOCI restores places first seen over 60 s earlier; full softmax replaces them.

**Localized consistency at short-horizon revisits.** Local DINOv2 error is significantly lower for LOCI.

Peak GPU memory versus generated length — **Cost versus generated length** on a 300 s trajectory (single H200): peak GPU memory.

Long-horizon generation and cost

Replacing half of the historical-KV layers with recurrent state extends the reachable length from 157 s to 225 s. Under bounded sparse access, LOCI runs the full 300 s at a constant 23.6 GiB.

Any world. Same memory.

Each world starts from a single image: a photograph, an engine render, or a generated painting. Open one to watch it explored along several camera paths.

Data

Training data

Unreal Engine scenes and CARLA towns that we rendered, plus real walking videos from Sekai. No MIND clips are used.

Held-out recorded trajectories

Held-out 40 s walk-through clips that repeatedly return to earlier viewpoints.

DatasetHugging Face · sum0214/LOCI-revisit-data

Live. On one GPU.

A 4-step distilled student streams the world as you steer it. This clip is a recording of the interactive demo: the keys lit below are the ones that were held.