112 zero-byte first-frame images in dataset/all_pack/assets/Sim/Diff/first frame/ (breaks I2V/TI2V inference)

[xyz918]

Summary

In the released dataset EmbodiedCity/iWorld-Bench-Dataset, 112 first-frame images under
dataset/all_pack/assets/Sim/Diff/first frame/ are 0-byte (empty) files. They are listed in
metadata.csv as the first_frame_path for 112 Sim/Diff samples, so any image-to-video
(I2V/TI2V) pipeline that conditions on the first frame fails on these samples.

Scope (verified)

• The directory dataset/all_pack/assets/Sim/Diff/first frame/ contains 1200 files, of which 112 are 0 bytes.
• All other 24 .../first frame/ directories (TartanAir-V2, RealEstate-10K, DL3DV-10K, NCLT, 7-Scenes,
  SpatialVid, Princeton365, Nuscene, KITTI, Waymo, TUM-RGB-D, TartanGround, …) contain 0 zero-byte files.
• Only the Sim/Diff first frames are affected.

Examples (3 of 112)

EMCITYENV0_x6350.728_y-3636.378_z-17.000_3_6_4.jpg   (0 bytes)
EMCITYENV0_x6350.728_y-3636.378_z-17.000_4_7_5.jpg   (0 bytes)
EMCITYENV0_x6383.114_y-4189.384_z-17.000_1_8_0.jpg   (0 bytes)

(Full list of 112 filenames can be attached.)

How to reproduce

from huggingface_hub import HfApi
api = HfApi()
sibs = api.repo_info("EmbodiedCity/iWorld-Bench-Dataset", repo_type="dataset",
                     files_metadata=True).siblings
zero = [s.rfilename for s in sibs
        if "Sim/Diff/first frame" in s.rfilename and (s.size or 0) == 0]
print(len(zero))   # -> 112

Impact

These 112 samples cannot be run through any first-frame-conditioned (I2V/TI2V) generator; the request
has no usable conditioning image. In our run this surfaced as 112 hard failures
(task='ti2v' requires an image or first_frame), i.e. 112/4300 evaluation samples were lost.

Notes / questions

• The real Sim data is shipped as split archives under sim/archives_split/dataset_env*/ (videos + cameras),
  but there is no documented script that extracts/populates these Sim/Diff first-frame placeholders, and the
  relevant archive (dataset_env0) is ~285 GB, so regenerating just 112 frames locally is impractical.
• Could you please either (a) re-upload the 112 missing first-frame JPGs to
  dataset/all_pack/assets/Sim/Diff/first frame/, or (b) document the intended procedure to generate them
  (e.g. frame-0 extraction from the corresponding source_video_filename in metadata.csv)?

Thanks for releasing iWorld-Bench!

[F-Jian-Jie]

Thank you very much for your detailed feedback and for identifying this issue. We apologize for the inconvenience. This appears to be an oversight during our dataset packaging and organization process. We will check the affected Sim/Diff first-frame files and update the dataset as soon as possible in the coming days to resolve the issue.

Thanks again for your careful inspection and for using iWorld-Bench.

[F-Jian-Jie]

Thank you for your question and for your patience.

We have carefully checked the reported issue and updated the affected Sim/Diff first-frame files. The corrected first-frame images have now been uploaded to:

dataset/all_pack/assets/Sim/Diff/first frame

Please re-download or sync the files in this directory and verify again. If you encounter any further missing files, zero-byte files, or other dataset-related issues, please feel free to contact us.

Thank you again for your careful inspection and for bringing this issue to our attention. We also appreciate your interest in and use of iWorld-Bench.