File size: 4,789 Bytes
6dacb65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5021165
6dacb65
 
 
5021165
6dacb65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
---
pretty_name: ReMind Dynamic-Memory Training Data
license: other
task_categories:
  - text-to-video
  - image-to-video
language:
  - en
tags:
  - video-generation
  - world-model
  - camera-control
  - temporal-dynamics
---

# ReMind Dynamic-Memory Training Data

Dataset release for **Teaching Video Generators to Remember: Eliciting Dynamic
Memory for Out-of-Sight State Evolution**.

This is a manifest-first release of generated and simulated video data used for
ReMind training. It contains three source families:

1. physics-filtered videos generated with Helios;
2. pair-fixed InSpatio camera-augmentation videos; and
3. Kubric camera-augmentation renders.

All released manifests use the production `caption_v2` captions. These captions
describe scene content, physical dynamics, and temporal state changes without
encoding camera motion in text.

## Release manifests

| Component | Released rows | Hugging Face data | Production `caption_v2` manifest |
|---|---:|---|---|
| Helios generated, physics-filtered | 1,027,329 | `data/helios/shard-*.tar` | `manifests/train_manifest_helios.jsonl.gz` |
| InSpatio camera augmentation, pair-fixed | 265,945 | `data/inspatio/shard-*.tar` | `manifests/train_manifest_inspatio_pairfix.jsonl.gz` |
| Kubric camera augmentation | 126,057 | `data/kubric_camera_aug/shard-*.tar` | `manifests/train_manifest_kubric_camera_aug.jsonl.gz` |

The three manifests contain 1,419,331 released rows in total. The manifests are
authoritative for the released subset.

The Helios, InSpatio, and Kubric camera-augmentation files use
WebDataset-style tar shards. Each sample has a zero-padded key and includes an
`.mp4` video plus a `.json` annotation. Camera-conditioned samples additionally
include `.camera.npz`.

Depth, segmentation, and other non-release sidecars are excluded. In
particular, the InSpatio release contains the generated RGB video and paired
camera trajectory, but not depth videos or raw pipeline metadata.

## Critical InSpatio pairing note

Only use
`manifests/train_manifest_inspatio_pairfix.jsonl.gz`. It corrects the
historical 8×32 shard-export transpose and keeps each packed video paired with
its own identity, caption, and DA3 camera trajectory. Do not substitute an
InSpatio manifest dated before 2026-07-16.

The pair-fixed manifest contains:

| Trajectory family | Rows |
|---|---:|
| pan | 132,581 |
| loop | 93,658 |
| walk | 39,706 |

Each released InSpatio sample is self-contained: its `.mp4` and `.camera.npz`
members refer to the same rendered clip.

## Manifest schema

Every JSONL row includes at least:

- `video`: a `tar+hf://datasets/...!member` reference into this repository;
- `caption`: the production motion- and physics-focused `caption_v2`;
- `source`: source-family identifier;
- `id`: stable sample identity; and
- `caption_v2_meta`: caption audit and filtering metadata.

Depending on the source, rows may also contain `fps`, `num_frames`,
`trajectory_profile`, `camera_npz`, `packed_tar_case_id`, or other camera and
archive metadata. Consumers should preserve unknown fields rather than
projecting rows onto a smaller schema.

Representative `caption_v2_meta` fields include:

- `temporal_dynamics`;
- `physics_plausibility` for generated Helios data; and
- `n_moving_objects` for simulated data.

## Filtering policy

- Helios rows with low physics plausibility, static content, or unsuitable
  machine-cyclic motion were removed.
- InSpatio captions are inherited from their source clips after source mapping;
  camera geometry remains in separate pose fields.
- True-static simulation examples are removed, while camera-tour data is
  exempt from static-video filtering.
- Camera-motion instructions are scrubbed from captions because camera control
  is represented by camera metadata, not text.

## Intended use

The dataset is intended for non-commercial research on video generation,
world models, temporal dynamics, camera control, occlusion recovery, and
out-of-sight state evolution. Captions and metadata are automatically produced
and may contain errors. Users should audit samples for their application and
follow the terms of every underlying media source.

## License

See `LICENSE`. ReMind-authored `caption_v2` annotations, filtering/audit
metadata, and manifest arrangement are available under CC BY-NC 4.0.
Underlying videos and third-party metadata are not relicensed and remain under
their respective source terms.

## Citation

```bibtex
@article{xu2026teaching,
  title={Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution},
  author={Xu, Tianshuo and Xie, Yichen and Meng, Depu and Peng, Chensheng and Herau, Quentin and Jiang, Bo and Hu, Yihan and Zhan, Wei},
  journal={arXiv preprint arXiv:2605.25333},
  year={2026}
}
```