SceneMemBench

A Benchmark for Scene Memory in Interactive Video World Models.

600 controlled trajectories across 33 Unreal Engine scenes, containing 280,973 valid frames. The four trajectory types contain 100, 100, 200 and 200 cases respectively.

Download

from huggingface_hub import snapshot_download
snapshot_download("scenemembench/SceneMemBench", repo_type="dataset",
                  allow_patterns=["core/**", "README.md"], local_dir="SceneMemBench")

core/index.json lists all 600 cases. Each case contains context.mp4, eval.mp4, manifest.json and pose/poses.json. Frame ranges are zero-based and half-open. Camera positions use Unreal Engine centimeters; primary_pose stores [x, y, z, pitch, yaw, roll], with angles in degrees. Query videos are ground-truth evaluation references. Straight trajectories stop after the turnaround at B; no return translation is included.

Optional original PNG frames are stored in rgb/train-*.tar, with paths continued/<case_path>/rgb/<frame_id>.png. Extract them into core/. rgb-index.json lists planned shards; rgb-checksums.json records shards uploaded and verified so far. Download only shards present in the latter. An optional optional-full/ package contains full videos ending at the same valid endpoint; it is unnecessary for evaluation.

To prepare Context images for model inference without downloading original PNGs, run scenemembench prepare --case <case_path>. If you extracted original PNG shards, use those images directly instead.

Run evaluation

Install the SceneMemBench code, then run:

scenemembench evaluate --task novel --case SceneMemBench/core/continued/<case_path> \
  --prediction continuation.mp4 --output metrics.json

Use --task seen for Reversed Rotation and Full Rotation, and --task novel for Forward and Look Back and New-Viewpoint Revisit. See the code repository for model setup and metric dependencies.