README.md

---
license: gpl-3.0
pretty_name: Civitai motion scripts
tags:
- funscript
- motion
- civitai
configs:
- config_name: default
  data_files:
  - split: train
    path: data/catalog.jsonl
- config_name: videos
  data_files:
  - split: train
    path: data/videos.jsonl
---

# Civitai motion scripts

541 Civitai video IDs and 470 saved Main variants. All available
axes are exported. The funscripts and dataset metadata are licensed under
GNU GPL version 3 (see LICENSE). This license does not cover the linked videos.

Review labels follow matching approved folder exports. The publisher must have reviewed those scripts.

## Look up a video

Compute the lowercase hexadecimal SHA-256 of UTF-8 `civitai:<decimal video ID>`,
without whitespace or leading zeroes. Request `index/<first two hash characters>/<hash>.json`.
The index lists script variants and a preferred variant (approved first, then
highest quality; ties use the variant hash). Check `review_status` before using a
draft. Quality is the author's 0–5 rating; 0 means unrated. Hashes provide stable
names, not anonymity: numeric Civitai IDs are present in the index and catalog.

Each variant and catalog row includes `tags` (normalized lowercase labels) and
`tag_sources` (civitai, local, manual, curve). The curve source can label a clean
automatic draft; it does not approve a script. Automatic tags are suggestions, not verified
identities or motion descriptions. User removals are excluded from export. The
per-video index also includes tags from local copies without a usable script.

`data/videos.jsonl` is a separate video catalog: one row per Civitai ID, including
clips with no funscript. Rows contain `video_key`, `civitai_id`, `index_path`,
`tags`, `tag_sources`, `categories`, `category_paths`, `audio_sync` and `has_script`.
Tags and Audio sync do not require a script or imply approval. Join this catalog
to script rows by `video_key` or `civitai_id`. Folder categories remain separate
from tags. Metadata is combined across local copies of the same Civitai ID.
The video catalog is also available as the Hugging Face `videos` configuration.

Video rows, indexes and script variants also include `creator_username`, `post_id`
and `civitai_metadata`. Unknown creator/post values are null. Source details can
include `created_at`, `width`, `height`, `base_model`, `model_version_ids`,
`content_rating`, `stats`, `site`, `fetched_at`, `video_url`, `post_url` and
`creator_url`. These are cached public API observations, not inferred identities;
counts and links may change. Missing metadata can be recovered for old downloads
through the Civitai browser's Library metadata controls before the next upload.

When `has_script` is false, the index has `variants: []` and
`preferred_variant: null`. There are no synthetic script rows or approval labels.
`data/catalog.jsonl` remains one row per actual script variant for compatibility.
New video labels also refresh retained older script records, without changing
their curves, hashes, quality ratings or approval status.

Each variant and catalog row includes `audio_sync`, the author's checkbox label
for audio synchronization. The video catalog exports this label even without a
script. It is independent of quality and review status;
`false` (or a missing field in older exports) means unmarked, not a failed sync test.
Identical script copies combine this label: it is `true` if any copy is marked.

`intensity` is a 1–5 motion intensity rating: 1 very gentle, 2 gentle,
3 moderate, 4 strong, 5 very strong. Zero (or a missing field in older exports)
means unrated. It is independent of quality, audio synchronization and validation;
it is not a device speed or physical measurement. Identical script copies use the
highest assigned intensity; unrated copies do not erase a rating. Changing
intensity does not change the script or variant hash. `intensity_mode` is `manual`
for an author's rating or `auto` for a curve estimate from Main L0. Automatic
values are recalculated from the exported curve; they are not a human review.
Older exports without this field contain manual ratings. Ties between identical
copies prefer the manual rating's mode.

Download the chosen axis at `variant.scripts.L0.path` (or another axis), and verify
its `sha256`. Pin all downloads to the same Hugging Face commit to keep a consistent
snapshot. The catalog is JSON Lines, one row per variant. `manifest.json` documents
the schema, counts and SHA-256 of every managed file.

## VR variants

`has_vr` and `vr_variants` identify saved full stereo conversions. Each VR variant
has the `vr` tag, its own ID, dimensions and frame rate, `layout: full_sbs`,
`projection: flat`, and `eye_order: left_right`. `video_owner` and `script_owner`
refer to the original video key. The variant uses that original's scripts at the
same timestamps; no duplicate or retimed script is needed. Portrait and landscape
aspect ratios are preserved. Original 2D clips do not gain the VR tag.

Each VR variant also includes its recorded `settings`:

```json
{"model":"VDA_L","resolution":392,"divergence":3.5,"convergence":0.5}
```

`divergence` is **3D strength** (0–5); `convergence` is **screen depth** (0–1).
`resolution` is the depth inference detail (280, 392 or 518), not the video's
output dimensions. These values describe that specific conversion, independently
of the latest controls in the workspace. A player can filter or select with
`vr.settings.divergence === 3.5`; several matching variants can coexist.
Missing fields in older exports mean unknown; do not assume default strength 2.
The same settings appear in public/per-category `vr_variants` and encrypted media
index entries with `kind: "vr"`. Use the current decrypted index metadata when
reusing an older encrypted video bundle; its embedded metadata may predate these
fields. Adding settings refreshes metadata while reusing unchanged video bytes.

VR video bytes remain in encrypted media bundles. Category encryption policies
also govern their public metadata. Benchmarks, archived variants, excluded takes,
stale conversions and missing saved files are omitted. Disabling a variant changes
the next current catalog; previously uploaded history remains available.

## Categories

Each catalog row and variant includes `categories` (final folder names, such as
`dance`) and `category_paths` (folders relative to the library, such as
`September_2026/dance`). Each video's index includes the union across its variants.
Identical scripts found in several folders retain all their categories without
duplicating the scripts or changing their hashes. These are the publisher's folder
labels, not Civitai tags or validation status.

Files directly in the library root have empty lists. Temporary downloads use their
chosen destination category, if any; internal review directories are never labels.
Older snapshots without these fields can be treated as having empty lists.

## Download a script

```python
import hashlib, json
from huggingface_hub import HfApi, hf_hub_download

repo = "YOUR_ACCOUNT/YOUR_DATASET"
video_id = "142699180"
revision = HfApi().repo_info(repo, repo_type="dataset").sha
key = hashlib.sha256(f"civitai:{video_id}".encode()).hexdigest()
def get(path):
    return hf_hub_download(repo, path, repo_type="dataset", revision=revision)
with open(get(f"index/{key[:2]}/{key}.json")) as file:
    index = json.load(file)
if not index["variants"]:
    raise SystemExit("Video metadata is available; no funscript has been published.")
variant = next(v for v in index["variants"] if v["variant_id"] == index["preferred_variant"])
axis = variant["scripts"]["L0"]
with open(get(axis["path"]), "rb") as file:
    data = file.read()
assert hashlib.sha256(data).hexdigest() == axis["sha256"]
script = json.loads(data)
```

## Timing and review

Actions use integer milliseconds from the local source video's beginning and
positions 0–100. Use the same untrimmed clip and check `duration_ms` before playback.
The Civitai ID identifies an individual media item, not the bytes of a particular
download or edit. `post_id` groups media items uploaded together in a gallery.
Transcodes, cuts and speed changes can affect synchronization. Multiple script
variants for the same video stay separate; axes are never combined across variants.

Only completed saved Main results produce script rows. Identified clips still
contribute video metadata while processing, without Main, or with outdated results.
Clips without a Civitai ID cannot be addressed in this dataset. Skip for now (formerly
Ignore) is a local review choice and does not withdraw saved scripts. Uploads
keep the last published result for a video until a fresh replacement is ready.
Unsaved browser edits and historical source tracks are not included.
Drafts can contain tracking errors or incomplete
coverage and are not automatically approved. An approved result edited afterward
is labelled as a draft until approved again when folder approval labels are enabled.

`foley` is null for original sound, or describes an explicitly selected generated
soundtrack: its dataset-relative `path`, SHA-256, duration, sample rate, prompt
and generation settings. Download that FLAC and play it against the original
Civitai video at the same timestamps; fall back to the original if unavailable.
It does not modify `audio_sync`, script approval, identity or timing.
`foley_variants` lists separately classified takes with audio descriptors, IDs,
`category_paths` and `categories_updated_at`. They share the same video and scripts.
`original_category_paths` keeps physical folders and default-soundtrack labels;
`label_category_paths` identifies Category manager labels, with `labels_updated_at`
so removed labels stay removed on updates. `category_paths` includes all
classifications. Players can select a matching Foley by category,
while `foley` remains the independent default. Empty variant category lists mark
removed assignments and contain no audio path. Unclassified, unselected takes
and short instructions stay local.

Category names and relative category paths are public. No source videos/audio,
thumbnails, original filenames, absolute local paths, private notes,
model caches or API keys are included. Obtain videos separately through Civitai
with the access required by that service. `start_ms` and `end_ms` describe each
axis's action extent, not a claim that every scene has been reviewed.