license: other
task_categories:
- text-to-video
- audio-to-audio
language:
- en
tags:
- spatial-audio
- audio-visual
- stereo
- video-editing
- ltx
size_categories:
- 100K<n<1M
SpatialAV2AV — Spatial Audio-Video Editing Dataset
Source→edit pairs for training spatial (binaural/stereo) audio-video editing with LTX-2.
Each pair renders the same clip under a different camera trajectory; the model learns
source (pre-edit space) + trajectory instruction → edit (target space) for both video and
the stereo sound field.
Contents
Each sample is a triple sharing one basename <video_id>+<traj>:
| Path | What |
|---|---|
final_json/<clip>.json |
metadata: video_id, traj_name, width, height, fps, and relative paths to source/edit |
final_edit/<clip>.mp4 |
target video — carries embedded 2-channel / 44.1 kHz stereo audio |
final_source/<clip>.mp4 |
condition video (same clip, source camera) — also embedded stereo audio |
Audio is inside the MP4s (AAC, real stereo). No separate
.wavfiles are needed — the trainer reads the audio track directly from each MP4.
all.list lists the 116,147-pair training split (relative json paths). The final_source/
final_edit folders may contain a larger pool; all.list is the authoritative training set.
Camera trajectories: push_in, pull_out, pan_left, pan_right, rotate_left,
rotate_right, fixed_left, fixed_right, fixed_rot_left, fixed_rot_right.
Sizes
- Training split: 116,147 pairs, ~147 GB (edit ~0.79 MB + source ~0.50 MB per pair).
- Resolutions: short side ~480; assorted aspect ratios.
fps= 25.
Usage
from huggingface_hub import snapshot_download
root = snapshot_download("BingoG/LTX", repo_type="dataset")
# point the trainer's data.json_list at f"{root}/all.list"
# json paths inside are relative to `root`, so they resolve after download.
The training loader normalizes each final_json entry to (edit=target video+audio, source=condition video+audio) and derives the caption from traj_name.