---
license: cc-by-sa-4.0
task_categories:
  - image-to-video
language:
  - en
tags:
  - driving
  - street-view
  - world-model
  - view-interpolation
  - kuala-lumpur
size_categories:
  - 10K<n<100K
---

# kl-interp-v1 — densified Kuala Lumpur street-view clips

18,870 five-second driving clips (1280×704, 24 fps, 73 or 121 frames) generated
from sparse street-view image sequences of Kuala Lumpur (Mapillary and KartaView,
CC BY-SA 4.0) with a view-interpolation video model. Each clip is a 5 s window
of a real capture sequence: the model was given 2–9 real keyframes with their
GPS positions and synthesised the frames in between. Keyframes are real; the
frames between them are generated.

Built for world-model research: cross-session revisits of the same place, GPS
per clip, and per-frame camera motion.

## Layout

```
clips/kl-s{0..3}-NNN.tar      tar shards, ~300 clips each, members <name>.mp4
poses/kl-poses.tar            <name>.npy  (T,7) float32 per-frame relative camera
                              action [dx,dy,dz,droll,dpitch,dyaw,flag] in CARLA
                              convention (x forward, y right, z up), estimated
                              with VGGT-Omega and scaled by the GPS arc length
manifest_s{0..3}.jsonl        one row per clip: name, clip_id, frames, arc_m,
                              caption ("window" = per-window caption used,
                              "clip" = whole-sequence fallback), guidance
qc_scores.jsonl               cheap motion/sharpness QC per clip (weak tag; see
                              miniworld/scripts/benchmarking/qc_clips.py)
arcs.json                     GPS arc length (m) per clip
captions/                     Qwen-generated short/long captions per sequence
                              window (kl_windows_s*.jsonl) and per sequence
```

`name` is `<source>-<sequence>_w<start>`; `clip_id` is `<source>:<sequence>`.
The original keyframes and GPS come from the `kl_real_v1` corpus.

## Generation

Wan2.2-TI2V-5B with a rank-128 LoRA trained with the freeze-frame keyframe
recipe of the Seoul World Model (arXiv:2603.15583) on CARLA v2 synthetic video,
OpenDV and Kuala Lumpur sequences; 35 sampling steps, flow shift 7, guidance 3.
Code: https://github.com/kamwoh/miniworld (`scripts/generation/interpolate_batch.py`).

## Caveats

- Generated frames can morph objects on close-range lateral pans; `qc_scores`
  is a weak filter. Treat non-keyframe content as synthetic.
- Poses are monocular estimates; scale drift is possible on long arcs.

## License and attribution

Source imagery: Mapillary and KartaView contributors, CC BY-SA 4.0. This
derivative is released under the same license.