liyzzz/World-from-Motion-DyCheck-Masks

World from Motion: DyCheck evaluation masks

Dynamic-region masks for the held-out test frames of the seven DyCheck iPhone scenes. They are used only to evaluate the Table 1 reproduction of World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026). Using these masks lets you skip SAM 3 (a gated model) when reproducing the evaluation.

Project page · Paper · Code · Model

Which masks these are

dycheck-evaluation-masks-sam3-20261007 was regenerated with SAM 3 on 2026-10-07, during an independent reproduction that used only public inputs. These are not the masks behind the paper's Table 1 numbers. Those historical masks are not part of this release. Scores computed with this set count as a new reproduction, not a re-scoring of the paper's runs.

How they were made:

• Model: facebook/sam3 at revision 3c879f39826c281e95690f02c7821c4de09afae7 (sam3.pt, SHA-256 9999e2341ceef5e136daa386eecb55cb414446a00ac2b55eb2dfd2f7c3cf8c9e).
• Input: held-out test RGB from the DyCheck iPhone export. Each mask is the union of the prompts below.
• paper-windmill frames were passed to SAM 3 as a lossless FFV1 video; all other scenes used the default transport.

Scene	Prompts	Frames
apple	apple, arm	532
block	gray block, block	667
paper-windmill	pinwheel	419
space-out	person, human	681
spin	person, human	722
teddy	teddy, hand	688
wheel	steering wheel, arm	219

How the masks are used

During evaluation the masks exclude moving content when the held-out test cameras are aligned, and they split the Valid metrics into dynamic and static regions. They are never used for training, reconstruction or video generation.

Layout

The masks ship as a single uncompressed ZIP, so they download as one file rather than thousands:

dycheck-evaluation-masks-sam3-20261007.zip
  dycheck-evaluation-masks-sam3-20261007/
    manifest.json
    ms/<scene>/test_dynamic/<camera>_<frame>.png    # MoSca runs
    wt/<scene>/test_dynamic/<camera>_<frame>.png    # WorldTree runs

The archive is 17,644,275 bytes with SHA-256 a106a67d856acc241109333b2839706b6099fd7466287c4c89ec8488fc8e1e6e. Its entries have fixed timestamps, so the code repository's packager rebuilds it byte for byte. There are 14 sets (two methods × seven scenes): 7,856 PNGs totalling 11,264,264 bytes. File names match the DyCheck held-out frames. Each mask is an 8-bit single-channel PNG at the frame resolution (360 × 480, width × height): 255 marks dynamic content and 0 marks static background. The MoSca and WorldTree sets were produced separately and are byte-identical for every scene. Both are kept so that each method's run reads its own set.

manifest.json has SHA-256 daa2e01082bb660718347daeafd893c5ff780bb04dd0dd90d1577883c12eab53. For every mask it records the file size, the SHA-256 of its bytes and of its decoded pixels, its image size and mode, and the decoded-pixel identity of the matching held-out RGB frame. The loader therefore rejects a mask paired with the wrong frame. The manifest contains no paths and no RGB data.

Usage

With the code repository, download the masks and run a scene with them in place of SAM 3:

python reproduce/from_raw/download.py --root "$WFM_INPUTS" --group masks
python reproduce/from_raw/run.py --inputs "$WFM_INPUTS" --output "$RUN_DIR" \
  --method ms --scene wheel --evaluation-masks released

The downloader pins this repository to an exact commit and checks the archive, the manifest SHA-256 and every mask before use. The runner checks each mask against your downloaded held-out RGB before any GPU stage starts. See REPRODUCTION.md.

To fetch the archive on its own:

hf download liyzzz/World-from-Motion-DyCheck-Masks \
  dycheck-evaluation-masks-sam3-20261007.zip --repo-type dataset --local-dir .

License and attribution

The masks are released under Apache-2.0. They are derived from the DyCheck iPhone dataset (Gao et al., Monocular Dynamic View Synthesis: A Reality Check, NeurIPS 2022) and were produced with SAM 3 from Meta. Follow the DyCheck terms for the underlying frames, and acknowledge SAM 3 when publishing results that use these masks.

Citation

@inproceedings{zhu2026worldfrommotion,
  title = {World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video},
  author = {Liyuan Zhu and Shengyu Huang and Amrita Mazumdar and Tianye Li and Zan Gojcic and Gordon Wetzstein and Iro Armeni and Shalini De Mello and Alex Trevithick},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2026}
}

@inproceedings{gao2022dycheck,
  title = {Monocular Dynamic View Synthesis: A Reality Check},
  author = {Hang Gao and Ruilong Li and Shubham Tulsiani and Bryan Russell and Angjoo Kanazawa},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2022}
}