---
license: apache-2.0
task_categories:
  - image-to-video
language:
  - en
tags:
  - driving
  - world-model
  - view-interpolation
  - drivingdojo
size_categories:
  - 10K<n<100K
---

# dojo-interp-v1 — DrivingDojo densified to 24 fps

10,068 five-second driving clips (1280×704, 24 fps, 73 or 121 frames) generated
from DrivingDojo (Apache-2.0) 5 fps sequences with a view-interpolation video
model. Every 5th generated frame is a real DrivingDojo frame; the frames in
between are synthesised. Uniform keyframe spacing, guidance 2.

## Layout

```
clips/dojo-NNN.tar              tar shards, ~300 clips each, members <name>.mp4
poses/dojo-poses.tar            <name>.npy  (T,7) float32 per-frame relative camera
                                action [dx,dy,dz,droll,dpitch,dyaw,flag], CARLA
                                convention, estimated with VGGT-Omega on the
                                generated clip (unscaled: monocular)
ref_actions/dojo-ref-actions.tar <clip_id>.npy  actions estimated on the original
                                5 fps DrivingDojo frames (reference for the clip)
manifest_s{0..3}.jsonl          name, clip_id, frames, arc_m (null), caption, guidance
qc_scores.jsonl                 cheap motion/sharpness QC per clip (weak tag)
```

`name` is `<clip_id>_w<k>`; `clip_id` is the DrivingDojo sequence id.

## 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); 35 sampling steps, flow
shift 7. Code: https://github.com/kamwoh/miniworld
(`scripts/generation/interpolate_batch.py`).

## Caveats

About 20% of DrivingDojo clips are near-stationary (stopped at lights); the QC
`static` flag marks them. Generated frames can morph on close-range pans.

## License and attribution

Source frames: DrivingDojo (Apache-2.0). This derivative keeps Apache-2.0.