VMem-Bench

Text gold labels and SUT prompts for VMem-Bench: a benchmark of long-range visual memory in causal long video generation. This Hugging Face repo has no source videos.

Track	Question	What the system under test sees	What this repo ships
A	Can memory retrieve the right visual identity over a movie timeline?	Per-chunk prompt text + a real video segment (stand-in for generated pixels)	Frozen text gold: entity_registry, chunk_annotations, chunk_index
B	Does visual memory survive all the way into generated pixels?	A chronological prompt stream only	50 long-horizon stories: gt, gt_source, sut_prompts (zh/en)

• Scoring / protocol: github.com/Suchenl/VMem-Bench
• Method: github.com/Suchenl/MemStrata

The Dataset Viewer reads movie/story catalogs (index/*.jsonl). Nested gold JSON stays under trackA/ and trackB/; the dropdown is subset (trackA / trackB), not a train/test split.

Download

hf download Suchenl/VMem-Bench --repo-type dataset --local-dir ./VMem-Bench-data

from huggingface_hub import snapshot_download
snapshot_download("Suchenl/VMem-Bench", repo_type="dataset", local_dir="./VMem-Bench-data")

from datasets import load_dataset
track_a = load_dataset("Suchenl/VMem-Bench", "trackA")  # 92 movies
track_b = load_dataset("Suchenl/VMem-Bench", "trackB")  # 50 stories

Point the code repo at the download (docs/DATA.md in VMem-Bench). Clone next to MemStrata so adapters can see ../MemStrata/src.

Contents

Path	Contents
trackA/BlenderOpenMovies/<movie_id>/gold/	23 Blender Open Movies titles
trackA/LSMDC/<movie_id>/gold/	69 LSMDC titles
trackB/en/ and trackB/zh/	gt/, gt_source/, sut_prompts/ for 50 stories (~6.2k segments, ~8.6 h scripted)
trackB/freeze_manifest.json	Frozen sample hashes (0001_lighthouse_keeper is frozen)
index/trackA_movies.jsonl	Viewer catalog: source, movie id, gold paths
index/trackB_stories.jsonl	Viewer catalog: story id, segment count, en/zh paths

File schemas: trackA/README.md, trackB/README.md.

Track B stories are long-horizon cases, not single-prompt clips. Segment duration is 5 s; per-story length is about 5.7–16.4 minutes. SUT prompts must not contain entity IDs, probe labels, or scoring fields. GT is scoring-only.

Source videos (not in this repo)

• Blender Open Movies: official project pages; BBB via scripts/prepare_blender.sh in the code repo.
• LSMDC: apply at the LSMDC download page, stitch to LSMDC_Videos_Stitched/<movie_id>.mp4, cite Rohrbach et al., IJCV 2017.

Full layout: VMem-Bench docs/DATA.md.

License

JSON files in this dataset: CC BY 4.0 (LICENSE) — reuse is allowed, including commercially, with attribution. LSMDC and Blender videos keep their upstream terms.

Citation

@article{chen2026memstrata,
  title={Stratifying and Benchmarking Long-Range Memory for Causal Long Video Generation},
  author={Chen, Yuzhuo and Shi, Huafeng and Wang, Xinyu and Wang, Yucheng and Hong, Haoqin and Zhang, Guoxin and Ma, Zehua},
  year={2026}
}

When using LSMDC titles, also cite Rohrbach et al., IJCV 2017.