OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation
Paper • 2407.02371 • Published • 52
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
passed: bool
samples: int64
frame_buckets: struct<25: int64, 41: int64, 33: int64, 49: int64, 57: int64, 65: int64, 73: int64, 81: int64>
child 0, 25: int64
child 1, 41: int64
child 2, 33: int64
child 3, 49: int64
child 4, 57: int64
child 5, 65: int64
child 6, 73: int64
child 7, 81: int64
batch_shapes: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
ddp_batch_alignment_verified: bool
actual_custom_trainer_verified: bool
mask_shape: list<item: int64>
child 0, item: int64
caption_count: int64
tar_sha256: string
batch_shape: list<item: int64>
child 0, item: int64
repo_revision: string
to
{'passed': Value('bool'), 'samples': Value('int64'), 'batch_shape': List(Value('int64')), 'caption_count': Value('int64'), 'mask_shape': List(Value('int64')), 'tar_sha256': Value('string'), 'repo_revision': Value('string'), 'actual_custom_trainer_verified': Value('bool')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
passed: bool
samples: int64
frame_buckets: struct<25: int64, 41: int64, 33: int64, 49: int64, 57: int64, 65: int64, 73: int64, 81: int64>
child 0, 25: int64
child 1, 41: int64
child 2, 33: int64
child 3, 49: int64
child 4, 57: int64
child 5, 65: int64
child 6, 73: int64
child 7, 81: int64
batch_shapes: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
ddp_batch_alignment_verified: bool
actual_custom_trainer_verified: bool
mask_shape: list<item: int64>
child 0, item: int64
caption_count: int64
tar_sha256: string
batch_shape: list<item: int64>
child 0, item: int64
repo_revision: string
to
{'passed': Value('bool'), 'samples': Value('int64'), 'batch_shape': List(Value('int64')), 'caption_count': Value('int64'), 'mask_shape': List(Value('int64')), 'tar_sha256': Value('string'), 'repo_revision': Value('string'), 'actual_custom_trainer_verified': Value('bool')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Sana Video 2 Latents
LTX-2.3 VAE latents of curated OpenVid-1M clips, with their original captions. License: CC BY 4.0 (see Source and license).
Fixed 81-frame release status: complete. Uploaded unique samples: 70,918. Uploaded TAR bytes: 59,374,970,880. The first release targets 70,918 duration-based candidates; actual decoded timestamp coverage is checked during encoding. The supplemental duration-bucket release is described below and in full_manifest.json.
Encoding contract
- 81 frames sampled at 16 fps, relative timestamps 0 through 5 seconds.
- 480 x 832 pixels (height x width), bilinear resize-to-cover and center crop, no antialias; input pixels normalized to [-1, 1].
- Pinned
diffusers/LTX-2.3-DiffusersVAE, BF16, full-video posterior mode, no spatial or temporal tiling. The checkpoint config setsencoder_causal=true. - Normalization:
(z - latents_mean) * scaling_factor / latents_std, using the VAE's own buffers. Per-channel arithmetic is performed in the VAE dtype. - Each NPZ contains one
zarray: float16, shape (128, 11, 15, 26) (C,T,H,W). - Captions are preserved as text. Gemma embeddings are not included.
- All frozen VAE settings, weight checksums and reference code hashes are in
encoding_config.json. Runtime versions and validation results are published with the pilot once it passes validation.
Download and extract
Download data/81x480x832/*.tar. The original 81-frame release commit
(f8dfbe11055a9e41b2232bcfe0218fae126c8878, pinned by instructions/download_latents.py)
is kept on the branch release-81.
Verify each archive against manifest.json, then extract the archives into one
directory. Archives contain:
LTX2VAE_diffusers_81x480x832/video/<key>.npz
metadata/81x480x832/<part>.jsonl
The pilot archive overlaps production samples. The supplied dataset adapter deduplicates sample keys when both metadata sets are present.
Training without the source MP4 archives
tools/latent_dataset.py reads NPZs and independent JSONL metadata. It returns
the eight-field sample contract of the pinned public Sana video trainer and
preserves all 128 latent channels. Original MP4 archives are not needed by this
adapter. To integrate, import it and call register_with_sana() before the
trainer calls build_dataset, then use:
data:
type: SanaLatentManifestDataset
data_dir: {video: /download/metadata/81x480x832}
vae_cache_dir: /download/LTX2VAE_diffusers_81x480x832
load_vae_feat: true
load_text_feat: false
num_frames: 81
image_size: 480
aspect_ratio_type: ASPECT_RATIO_VIDEO_480
caption_proportion: {prompt: 1}
model:
multi_scale: false
text_encoder:
model_max_length: 300
The custom 100M training branch has not yet been supplied for an actual trainer forward/backward check. The exported cache follows its supplied NPZ contract; the portable adapter and a DataLoader batch are checked after a fresh Hub download.
Provenance
Source identifiers, source file hashes, curation config hashes, selected source timestamps and encoding signatures are retained in each sample's metadata.
Source and license
The clips come from OpenVid-1M, which is released under
CC BY 4.0. This repository contains LTX-2.3 VAE latents of a curated subset of those clips and their unchanged
OpenVid-1M captions, and it is released under the same license, CC BY 4.0. Changes made: clip selection
(curation), temporal sampling at 16 fps, resize-to-cover and center crop to 480x832, and VAE encoding; the exact
settings are in encoding_config.json and multi_duration_encoding_config.json. If you use this data, please
cite OpenVid-1M:
@article{nan2024openvid,
title={OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation},
author={Nan, Kepan and Xie, Rui and Zhou, Penghao and Fan, Tiehan and Yang, Zhenheng and Chen, Zhijie and Li, Xiang and Yang, Jian and Tai, Ying},
journal={arXiv preprint arXiv:2407.02371},
year={2024}
}
Full collection and duration buckets
Full collection status: complete. Uploaded unique clips: 125,223 / 125,223.
The original 70,918-sample release is preserved. The remaining 54,305 accepted
clips are encoded using the largest supported 8k+1 frame count available
from real source timestamps at 16 fps, capped at 81 frames. The time dimension
of each normalized NPZ is (frames-1)/8+1; spatial shape remains 15x26.
full_manifest.json lists the 70 original production archives plus completed
supplemental archives in data/multi_duration/. Pilot archives are excluded.
After extraction, use tools/multi_duration_dataset.py with the parent of all
bucket cache directories as vae_cache_dir, and use
DistributedDurationBatchSampler so each batch has one duration.
The sampler coordinates duration batches across DDP ranks; do not shard its
DataLoader again. Call set_epoch(epoch) on the duration batch sampler.
The custom 100M trainer still needs its own forward/backward integration test.
The fixed-duration English guide and companion files are in instructions/.
- Downloads last month
- 46