WorldMemBench: A Comprehensive Benchmark for
Memory in Video World Models

Nan Zhou^1,3,*, Jianxiong Gao^2,3,*, Qiuyu Wang³, Wen Wang¹, Jiahao Wang^4,3,
Yufeng Yuan^4,3, Shuaiting Li^1,3, Hao Ouyang³, Yujun Shen³, Weiming Lu¹,
Jun Xiao¹, Yueting Zhuang¹, YongLiang Shen^1,†
1 Zhejiang University   ² Fudan University   ³ Ant Group   ⁴ Nanjing University

Overview

WorldMemBench evaluates whether video world models preserve people, objects, and environments across camera departures and returns, and whether dynamic events remain consistent over time. It also measures general video quality and camera following, with 12 metrics in total.

This release provides 6,705 distinct generated videos from nine models, together with model-independent source cases, recorded generation inputs, evaluation annotations, and prepared depth/camera-pose artifacts.

Evaluation	Manifest	Entries	Metrics
General Video Quality	general_video_quality/manifest.jsonl	5,400	imaging_quality, aesthetic_quality
Human Memory	human/manifest.jsonl	1,620	human_identity, human_appearance
Object Memory	object/manifest.jsonl	1,890	object_appearance, object_geometry
Environment Memory	environment/manifest.jsonl	1,890	environment_visible_memory, environment_depth_fidelity, environment_scene_reconstruction
Dynamic Memory	dynamic_memory/manifest.jsonl	540	dynamic_memory
Camera Rotation	camera_rotation/manifest.jsonl	2,295	camera_rotation
Camera Translation	camera_translation/manifest.jsonl	765	camera_translation

Leaderboard

Quantitative comparison of nine video world models on WorldMemBench, from the paper's main results table. ↑ / ↓ indicate higher / lower is better. Bold and underlined values denote the best and second-best results for each metric, respectively.

Metric	HY-WorldPlay	Matrix-Game-3	Matrix-Game-3.5	SANA-WM	YUME	LingBot-World	LingBot-World-v2	Evoke	DreamX-World
Video Quality
Imaging ↑	0.659	0.680	0.603	0.542	0.713	0.671	0.683	0.670	0.686
Aesthetic ↑	0.527	0.510	0.429	0.434	0.533	0.545	0.534	0.533	0.537
Static Memory
Human ID ↑	0.828	0.483	0.019	0.134	0.298	0.393	0.534	0.550	0.286
Human App. ↑	0.830	0.502	0.043	0.251	0.551	0.661	0.645	0.762	0.556
Object App. ↑	0.879	0.612	0.471	0.328	0.531	0.448	0.564	0.710	0.569
Object Geometry ↑	0.761	0.518	0.457	0.314	0.449	0.418	0.506	0.651	0.491
Env. Visible ↑	0.947	0.782	0.620	0.537	0.682	0.728	0.694	0.762	0.610
Env. Depth ↑	0.905	0.611	0.645	0.668	0.676	0.680	0.517	0.730	0.419
Env. Recon. ↑	0.619	0.423	0.390	0.367	0.347	0.364	0.253	0.400	0.105
Dynamic Memory
Overall ↑	0.294	0.046	0.240	0.415	0.775	0.808	0.696	0.706	0.563
Return Prog. ↑	0.292	0.050	0.183	0.342	0.708	0.775	0.667	0.667	0.558
Persist./Reset ↑	0.308	0.050	0.233	0.458	0.817	0.833	0.733	0.717	0.600
Final State ↑	0.267	0.042	0.200	0.333	0.717	0.750	0.592	0.658	0.492
Causal/Phys. ↑	0.308	0.042	0.342	0.525	0.858	0.875	0.792	0.783	0.600
Camera Follow
Trans. RPE ↓	0.609	0.424	1.018	0.771	0.853	0.575	0.565	0.838	1.139
Rot. RPE ↓	0.417	0.554	0.837	0.464	0.529	0.637	0.439	0.573	0.928

App., Env., Recon., Trans., Rot., Prog., Persist., and Phys. denote Appearance, Environment, Reconstruction, Translation, Rotation, Progress, Persistence, and Physics, respectively. The four Dynamic Memory breakdown rows accompany its Overall score. See the interactive project leaderboard for sorting and comparisons.

Download

From the WorldMemBench code directory:

hf download rescuerz/WorldMemBench --repo-type dataset --local-dir ./data

Evaluation weights are distributed separately in WorldMemBench-weights.

For a smaller first run, use WorldMemBench-examples, which covers all twelve metrics with five shared videos.

Structure

data/
├── testsets/                    # model-independent source cases
│   ├── human/
│   ├── object/
│   ├── environment/
│   ├── dynamic_memory/
│   └── camera_following/
├── human/
├── object/
├── environment/
├── dynamic_memory/
├── camera_rotation/
├── camera_translation/
├── general_video_quality/
│   └── manifest.jsonl           # references existing videos
└── viewer/                     # browsing tables, not evaluator inputs
    ├── gallery.parquet
    └── manifests/

Generated outputs follow <category>/<model>/<trajectory>/<case>/. Each evaluation category has a manifest.jsonl with paths relative to that manifest.

• Source cases: testsets/ provides source first_frame.jpg, prompts, and category-specific annotations. Use these model-independent inputs when evaluating a new model.
• Generation records: video.mp4 is the generated output. Case-level generation_meta.json records the original generation settings and camera/control input paths. Model-specific input files are retained where applicable.
• Evaluation inputs: annotations, source Object masks, Dynamic checklists, and trajectory.json supply the metric-specific protocol. da3/ and vipe/ contain prepared depth and pose predictions. Camera Rotation manifests reference videos stored under Object or Environment.

Source first_frame.jpg images are distinct from generated video frames and may have a different resolution. The model-specific inference implementations are not included; the recorded inputs document the released runs but do not alone guarantee bitwise regeneration.

Evaluation

worldmembench eval ./data/human/manifest.jsonl \
    --metrics human_identity human_appearance \
    --models lingbot-v2 --trajectories lrlr_sym_0p75dpf \
    --cases case_000 --device cuda:0

Evaluate General Video Quality on its full 5,400-video subset:

worldmembench eval ./data/general_video_quality/manifest.jsonl \
    --metrics imaging_quality aesthetic_quality

Use the other manifest paths and metric names from the overview table for the remaining evaluations. Human/Object manifests already contain the intended revisit counts. Prepared DA3/ViPE artifacts are reused; Dynamic Memory requires either a local VLM server or an external API, as described in the Dynamic Memory guide.

Use --gpus 0,1,2,3 for multi-GPU evaluation instead of --device cuda:0. Results are saved under eval_results/<timestamp>/, including per-video metrics/<metric>/cases.jsonl and aggregate summary.json. Use --output to choose another result directory.