rescuerz/WorldMemBench-examples

WorldMemBench Examples

A small, ready-to-evaluate example package for WorldMemBench, containing five LingBot-v2 videos and twelve metric manifests. Use it to try the evaluation pipeline; it is not the full benchmark and its scores are not the paper's aggregate results.

The Dataset Viewer opens the independent gallery subset, showing source first-frame images, model names, categories, trajectories, case IDs, video URLs, and item IDs. Its six entries reference five distinct videos because Camera Rotation reuses the Environment video. Images are embedded original source images, not frames extracted from generated videos; the gallery also includes source images from the corresponding WorldMemBench test cases where they were not needed in the original examples package.

The twelve metric subsets remain available through Parquet copies in viewer/manifests/, each with a test split. Original evaluation manifests remain unchanged. Supporting JSON files under data/ are not combined into the viewer's tables.

Download

Set up the scoring environment using the WorldMemBench installation instructions. Run the following commands from the WorldMemBench code directory:

conda activate worldmembench
hf download rescuerz/WorldMemBench-examples --repo-type dataset --local-dir ./examples
hf download rescuerz/WorldMemBench-weights --local-dir ./weights
python scripts/download_weights.py --verify

The examples are a dataset repository; the weights are a model repository. They are downloaded separately. The full weights package includes Qwen3.8-27B for local Dynamic Memory evaluation. To try only Imaging Quality without downloading all models, replace the weight-download and verification commands with:

python scripts/download_weights.py --metrics imaging_quality

Quick Start

Evaluate one metric:

bash scripts/eval_demo.sh ./examples imaging_quality

Replace imaging_quality with any metric listed below. Additional evaluation options can be appended, for example:

bash scripts/eval_demo.sh ./examples human_identity --device cuda:0

To evaluate all twelve metrics, first configure the Dynamic Memory VLM, using either a local server or an external API. With the default local server running:

bash scripts/eval_demo.sh ./examples all

For an external API, append the --vlm-backend api, --vlm-base-url, and --vlm-model options described in the code README, and set the API key environment variable. The script forwards these options to the evaluator.

The runner uses the code project's weights/ directory by default. To use existing weights elsewhere, append --weights-dir /path/to/weights. Results are saved to:

examples/eval_results/<metric>/<run>/

Each run includes command.txt, manifest.jsonl, and metrics/<metric>/cases.jsonl and summary.json. Evaluation results are generated locally and are not included in this dataset.

Included metrics

Each metric has an entry point at manifests/<metric>.jsonl:

Evaluation	Metrics	Videos per metric
General Video Quality	imaging_quality, aesthetic_quality	2
Human Memory	human_identity, human_appearance	1
Object Memory	object_appearance, object_geometry	1
Environment Memory	environment_visible_memory, environment_depth_fidelity, environment_scene_reconstruction	1
Dynamic Memory	dynamic_memory	1
Camera Following	camera_rotation, camera_translation	1

General Video Quality uses the Human and Environment videos. Camera Rotation also references the Environment video rather than storing another copy.

Data layout

examples/
├── README.md
├── data/
│   ├── human/
│   ├── object/
│   ├── environment/
│   ├── dynamic_memory/
│   ├── camera_rotation/
│   └── camera_translation/
├── manifests/
│   ├── imaging_quality.jsonl
│   ├── human_identity.jsonl
│   └── ...
└── viewer/                  # browsing only, not evaluator inputs
    ├── gallery.parquet
    └── manifests/

Use the supplied manifests: they contain relative video paths, prepared Human target matches, Object source-image/mask references, Dynamic checklists, and Environment/Camera artifact references. The DA3 and ViPE predictions are already included and reused for scoring; their preparation environments are needed only if you regenerate those predictions. Keep data/ and manifests/ together so the relative paths remain valid.

You can also evaluate a manifest directly:

worldmembench eval ./examples/manifests/object_geometry.jsonl \
    --metrics object_geometry

Unlike the demo script, a direct worldmembench eval command saves to eval_results/<timestamp>/ unless you specify --output.

For the full released model outputs, see WorldMemBench. For evaluation model licenses and usage restrictions, see WorldMemBench-weights.