maxwellinked/time-lapse-artifacts

Time-Lapse Artifacts

What this archive is

Time-Lapse Artifacts contains 801 indexed video files documenting one artist's traditional drawing practice from September 17, 2024 through September 17, 2026, with standardized acquisition since July 13, 2025. The recordings began as personal practice documentation and a durable record of manual work. They were not created as a prospective scientific protocol; their organization as a dataset and any research or model-analysis uses are retrospective.

Video is the primary published evidence. The archive follows the same creator working in physical media—including pencil, ballpoint pen, fountain pen, brush, graphite, and ink—on fixed-format paper and illustration board. Metadata identifies and organizes the recordings without turning undocumented details into facts.

Unit of record: each active metadata row identifies one published MP4 video object. A row does not necessarily represent one drawing, one uninterrupted session, or a complete start-to-finish process. Use the persistent record_id to identify a record; filenames are descriptive and may be corrected. Fragment-sequence, timing, and physical-reference relationships live in separate revision-pinned sidecars joined by record_id. If a relationship or attribute is not documented, it remains unknown.

In this dataset, provenance means the documented relationship between a physical drawing and its time-lapse record, not a history of ownership or sale. The collection is a single-creator longitudinal archive: it can support within-practice study, but it does not represent artists, materials, studios, or drawing practices broadly.

Status and coverage

This card describes the current active data state validated at immutable content revision 21648a6a76b9174d32a4a3314cbe490219fc2151 and the separate immutable tagged and DOI releases. Release-only documentation commits do not change those frozen counts.

Snapshot	Standard	Pre-Standard	Indexed video files	Meaning
Current active data (content pinned at 21648a…)	673	128	801	Indexed media through 2026-09-17; 2,130,736,360,966 bytes (approximately 2.13 TB), 1.94 TiB, and 59.552 encoded playback hours
Current release v0.6.0	629	128	757	Immutable tagged snapshot validated at 0f3e0559…; indexed media through 2026-09-16, 2,108,662,791,273 bytes (approximately 2.11 TB), 1.92 TiB, and 58.371 encoded playback hours
DOI release v0.5.1	581	128	709	Immutable, citable historical snapshot registered as 10.57967/hf/10449

The Hub exposes two configurations, both using the split name archive:

• canonical/archive is the default. It loads root metadata.csv and the direct files in Standard_Time_Lapses/. Core technical metadata is complete for each indexed Standard file.
• pre_standard/archive loads Pre_Standard_Time_Lapses/metadata.csv and the direct files in that folder. These records cover the earlier public period; nine dates and some historical descriptive or provenance fields remain unknown.

Both configurations use the same declared 29-field schema. Nullable fields remain null when evidence is unavailable; null must not be read as zero, false, or proof that a property was absent. No official train, validation, or test partition has been defined.

A separate 2012–2016 archival holding contains 2,381 locally held real-time recordings totaling 9,060,310,121,597 bytes (about 9.06 TB). Its file-level holdings and SHA-256 manifest is public, but the media is not hosted and is excluded from the active totals above. The date range is a historical scope label; it does not assert a capture year for every listed file.

Browse and load

Browse the Standard recordings through the searchable public browser with muted motion previews. The browser is a derived presentation layer and may lag the dataset. Hugging Face remains authoritative for media, record identities, and machine-readable metadata.

For lightweight media inspection, the separate derived-access pilot provides six silent proxies, 21 normalized encoded-position frames per record, and revision-pinned derivation manifests. It is an engineering fixture, not a representative sample; this repository remains authoritative.

For reproducible index inspection, pin the exact revision, stream the index, and disable video decoding:

python -m pip install 'datasets==4.8.5'

from datasets import Video, load_dataset

records = load_dataset(
    'maxwellinked/time-lapse-artifacts',
    'canonical',
    split='archive',
    streaming=True,
    revision='21648a6a76b9174d32a4a3314cbe490219fc2151',
)
metadata = records.cast_column('video', Video(decode=False))
first = next(iter(metadata))
print(first['record_id'], first['video']['path'])

This reads rows without intentionally transferring video payloads. The CSV path column is named file_name; Hugging Face resolves it to the typed video feature for a declared configuration. Use pre_standard as the configuration name for the earlier tier. Decoded access requires torchcodec and a compatible video stack; individual files may be several gigabytes.

To verify one Standard record without downloading its video, install huggingface_hub and run:

python tools/download_verified.py tla-a66bdad6c903b89f \
  --revision 'v0.6.0' \
  --check-only

The helper resolves the tag to an immutable commit and checks the indexed size and Hub identity. Remove --check-only and supply --output-dir only when a large download is intentional.

Interpret records

Canonical fields and sidecars answer different questions. Keep these distinctions when filtering, citing, or deriving annotations:

• Drawing/group finish versus recording time: all members of one documented standard-2026-08-27 grouping share the finish date and time of the complete grouping or physical work. That value is not the individual end time of each fragment.
• File order versus relationship scope: RecNN records file order within a documented group; it does not itself establish session boundaries or completeness. Compare the prospective 2026-08-27 packet with the separate June 18–30 historical relationship.
• Per-file timing evidence is separate. A device filename may document a recording-start clock; an MP4 creation_time tag may document a creation/finalization boundary; a calibrated value is an estimate. These sources are not interchangeable, and none by itself establishes drawing-active time or frame-level capture timestamps.
• duration_seconds and encoded_frame_rate describe encoded playback. They are not wall-clock duration, acceleration, or physical-motion calibration.
• acquisition_protocol_version marks an acquisition workflow. The earlier Standard workflow is standard-2025-07-13; reviewed prospective fragment groups use standard-2026-08-27. Membership comes from metadata and pinned provenance packets, not filename dates alone. This is not a skill-change label.
• fragment_set_label is a descriptive sidecar grouping label, not a global work identifier. Group completeness and sequence order are asserted only where the pinned relationship layer documents them.
• Multiple tool or medium values use semicolons, such as Pencil;FountainPen and Graphite;Ink. The two lists are independent; positional pairing is not guaranteed. A pipe is not a valid separator.
• orientation, width_px, and height_px describe the stored encoded file. They do not prove capture orientation, crop history, or camera geometry.

The Standard protocol records a nominal 0.2-second capture interval, but that is a protocol-level default rather than a confirmed per-record measurement. Do not infer uncertain chronology, material, ownership, endpoint state, or timestamp semantics from appearance or filename similarity alone. Full field definitions and filename rules are in SCHEMA.md.

Evidence layers and limits

The active CSV indexes carry stable identity, collection tier, documented session descriptors, encoded properties, repository identity, and provenance fields where supported. Separate evidence layers preserve claims that should not be flattened into the main 29-field table:

• The provenance index routes to dated fragment-sequence manifests. At current main, eleven reviewed packet layers spanning ten calendar dates identify 310 canonical fragments. They preserve grouping, order, completeness, and any noncanonical derivative relationship without adding derivative media to the active index.
• Recording Timing v1 is an exploratory, revision-pinned layer for 186 fragments in six earlier dated manifests. It separates 22 direct filename-to-MP4-boundary observations from 164 profile-matched estimates produced by a versioned in-sample model. The DOI snapshot fixes this published v1 artifact as historical evidence; it does not convert estimates into canonical timing truth or cover later groups.
• Physical_References/endpoint_references.csv links documented Standard records to owner-confirmed photographs through record_id and a pinned source revision. The photos are not bundled. See its schema, manifest, and capture and ingest protocol.

The videos have no frame-level bounding boxes, segmentation masks, pose, hand, pen-tip, stroke, aesthetic, or style labels. The archive also lacks a universal capture interval, original per-frame timestamps, camera calibration, and pixel-to-physical mapping. Researchers may derive annotations, but should bind them to source record_id values and an exact repository revision.

Use is best limited to archival and drawing-process research and carefully framed within-creator or representation studies. Temporal and physical-motion analyses need additional calibration; population, clinical, biometric, identity, authorship, and leakage-safe benchmark claims are not validated.

Important biases remain: one creator and one evolving practice, different acquisition and documentation conditions across tiers, possible visual shortcuts from the workspace, and storage costs that make exhaustive inspection expensive. Consult Archive_Specifications.md for acquisition details and known gaps.

Integrity, rights, and license

SHA256SUMS contains ordinary byte-level SHA-256 values for published MP4 and ZIP payloads. It is a portable fixity manifest, not encryption. The metadata field hub_xet_hash is a Hugging Face Xet file ID, not the SHA-256 of downloaded bytes; preserve both without conflating them. The checksum manifest intentionally excludes documentation, scripts, CSV indexes, and itself.

For reproducible use:

1. Pin a tag or full 40-character repository commit.
2. Record the selected record_id values.
3. Preserve the corresponding SHA256SUMS entries and hub_xet_hash values.
4. State whether audio was retained.
5. Document decoding, rotation, cropping, sampling, and proxy generation.

The creator confirms ownership of the published video files and depicted creator-owned drawings. Material that does not pass ownership and provenance review is not included in the active public index. This does not independently grant rights in incidental third-party content that may appear on screens, in references, reflections, or captured audio.

Most active indexed videos at this revision contain an audio stream (485 of 801, or 60.5%). has_audio reports stream presence, not whether the stream contains meaningful signal. Before redistribution or model release, review audio, visible screens, reference material, reflections, tattoos, and other identifying or third-party content. Derived proxy releases should remove audio by default while preserving the archival source objects.

The dataset is released under CC BY-NC 4.0. The license applies only to material for which the licensor holds the necessary rights. Users must review incidental third-party content and ensure that downstream use complies with the license notice. Commercial use outside those terms is not permitted.

Validation procedures and current checks are documented in VALIDATION.md. Run the repository validators before publishing metadata, sidecar, checksum, or release changes.

Citation and documentation

The current v0.6.0 release freezes a snapshot of 757 indexed video files validated at 0f3e0559ee3d6f5342b7043dd81c9b07ec123be1. It is a separate immutable release and is not the object registered under the existing DOI.

The DOI-backed v0.5.1 release remains registered as 10.57967/hf/10449 and remains fixed at 709 indexed video files. Cite the exact tag-resolved revision actually used. The BibTeX entry below intentionally describes the DOI-backed v0.5.1 snapshot, not v0.6.0.

@dataset{maxwell_time_lapse_artifacts_2026_v0_5_1,
  author    = {Maxwell, Tim},
  title     = {Time-Lapse Artifacts},
  year      = {2026},
  version   = {0bbf65f},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10449},
  url       = {https://huggingface.co/datasets/maxwellinked/time-lapse-artifacts/tree/v0.5.1},
  note      = {Tagged maintenance release; resolve the tag to its 40-character Hub commit for exact reproducibility}
}

Core documentation:

• metadata.csv and Pre_Standard_Time_Lapses/metadata.csv — the active machine-readable indexes.
• SCHEMA.md and Pre_Standard_Time_Lapses/SCHEMA.md — field, filename, and tier contracts.
• Standard_Time_Lapses/Provenance/ and Physical_References/ — relationship evidence kept outside the active indexes.
• Archive_Specifications.md and Earlier_Archive_2012_2016/README.md — acquisition scope, known gaps, and the offline holding.
• VALIDATION.md, SHA256SUMS, and RELEASE_MANIFEST_V0.6.0.json — validation, fixity, and the current release receipt. The historical RELEASE_MANIFEST_V0.5.1.json remains the DOI release receipt.
• CHANGELOG.md — release history and post-release changes.