Cseti/LTX2.3-22B_IC-LoRA-CrossView-Prompt-Dataset

CrossView Prompt Dataset

The training dataset behind the CrossView Prompt IC-LoRA for LTX-Video 2.3 — a "virtual second camera" adapter that re-renders a scene from a new viewpoint described by a short prompt.

Each sample is a pair of static-camera clips of the same scene (a reference view and a target view) plus a camera-delta caption describing how the target camera differs from the reference.

Contents

clips/<scene>/<cam>.mp4   # 504 unique clips, native 1280x1280, 81 frames, 15 fps
train.jsonl               # 294 ordered pairs
val.jsonl                 # 21 held-out pairs (scene-disjoint from train)
captions_all_63.txt       # the full 63-phrase caption vocabulary
LICENSE / NOTICE          # Apache-2.0 + source attribution

Each row in train.jsonl / val.jsonl:

{
  "video": "clips/scene679/cam08.mp4",           // target view (to generate)
  "reference_video": "clips/scene679/cam02.mp4",  // reference view (input)
  "caption": "crossview. new camera angle: far to the left, higher, closer.",
  "azimuth_bin": "far to the left", "elevation_bin": "higher", "distance_bin": "closer",
  "azimuth_deg": 47.9, "elevation_deg": 9.1, "distance_ratio": 0.52,
  "scene": "scene679"
}

*_deg / distance_ratio are the measured continuous relative-pose values; the *_bin fields and the caption are their discretization.

Caption vocabulary

Fixed template — crossview. new camera angle: {azimuth}, {elevation}, {distance}.

Axis	Phrases
azimuth	same angle · slightly to the left/right · to the left/right · far to the left/right
elevation	lower · same height · higher
distance	closer · same distance · further

63 combinations total (see captions_all_63.txt). The pairs are balanced across the 7 azimuth bins (42 train pairs each). Azimuth spans a frontal sector, max ~60° — the source cameras sit on a hemisphere in front of the subject.

How it was built

Derived from a 150-scene subset of SynCamVideo (10 static cameras per scene on a hemisphere around the subject). For every ordered camera pair we parsed the UE extrinsics, computed the relative pose in a subject-centred frame (azimuth / elevation / distance-ratio), and wrote the binned caption. Pairs were curated to a flat azimuth-bin balance with a scene-disjoint train/val split.

License & attribution

Apache License 2.0. This is a curated, re-captioned subset of SynCamVideo (KlingTeam, SynCamMaster paper arXiv:2412.07760, Apache-2.0) — see NOTICE. The original video frames are unmodified.