maxwellinked/time-lapse-artifacts

Time-Lapse Artifacts

873 indexed video files document one artist's traditional drawing practice. The recorded finish dates span September 17, 2024 through September 20, 2026; nine Pre-Standard dates remain unknown. Standardized acquisition began July 13,

1. 1.The current indexes contain 2,196,054,134,482 indexed video bytes (approximately 2.20 TB).

The recordings began as personal practice documentation and a durable record of manual work. The archive was initially organized as a dataset retrospectively; the current acquisition workflow is maintained prospectively. Recording was not initiated as a prospective scientific study. Each downstream study must state its own design and evidence requirements.

Video is the primary published evidence. The archive follows one creator working with physical drawing tools and materials on paper and illustration board. Metadata organizes the evidence and preserves unknowns. Here, provenance means documented origins and relationships among files, source recordings, derivatives, and physical drawings, with the evidence basis stated for each relationship. Physical-drawing identity is recorded only where confirmed.

Choose a collection

Both use the archive split and the same 35-field schema. Historical descriptive fields may be unknown even when technical metadata is available. Blank/null values are not zero or evidence of absence. No official train, validation, or test partition is defined.

The Hub aggregate of 1,618 rows counts Standard through both the combined view and its two subdivisions below. There are 873 unique indexed video files. For archive-wide selection, use canonical + pre_standard, or the two Standard views + pre_standard. Do not combine all four configurations as independent records.

The separate 2012–2016 archival holding has a public file manifest, but its media is held locally and excluded from these totals. Its historical scope label does not date every listed file.

<!-- BEGIN generated workflow-subsets -->

Select a Standard recording workflow

These views partition canonical/archive. Their totals must not be added to Standard again. Both retain the archive split.

Membership follows acquisition_protocol_version and its documented relationship packets. Unknown starts and unresolved work IDs remain included; filename dates and RecNN suffixes alone do not assign a workflow.

For metadata-only selection, use the generated earlier Standard CSV or individual-recording CSV. They preserve every source field and repository-relative media path. The loading example below pins a revision containing both views. Older releases retain their original configurations. See the selection and maintenance guide to save selected IDs or regenerate these views. <!-- END generated workflow-subsets -->

Browse and load

Browse the Standard files with muted motion previews. The browser may lag the dataset; Hugging Face is the authoritative source for published media, identities, and metadata. The separate derived-access pilot offers lightweight proxies and sampled frames from six files. It is an engineering fixture, not a representative sample.

Install the tested loading version:

python -m pip install 'datasets==4.8.5'

This pinned revision contains all four configurations. Choose canonical, pre_standard, standard_earlier, or standard_individual_recordings:

from datasets import Video, load_dataset

revision = '01615dbd09db3eee725872ad2952c62b6a236c21'
records = load_dataset(
    'maxwellinked/time-lapse-artifacts',
    'standard_individual_recordings',
    split='archive',
    revision=revision,
    streaming=True,
)
records = records.cast_column('video', Video(decode=False))
first = next(iter(records))
print(first['record_id'], first['video']['path'])

This reads the index without intentionally transferring video payloads. The CSV file_name column resolves to the typed video feature. Decoded access requires torchcodec and a compatible video stack; individual files may be several gigabytes. To inspect a newer state, resolve its revision once and retain that commit and the selected IDs using the selection guide.

Recording and drawing identities

A `record_id` identifies one published MP4 file. Historical files may combine multiple source recordings; the Standard individual-recording workflow preserves one unchanged source recording per file. Neither a metadata row nor a file boundary establishes a drawing, session, or complete process.

A `work_id` identifies a creator-confirmed drawing/composition. Currently, 745 files link to 380 work IDs; the other 128 files have unresolved drawing identity. Blank does not mean a different drawing. The column is generated from the work registry. Membership does not establish gapless capture, and different IDs are not independently verified negative pairs for a work-disjoint benchmark.

IDs apply across this archive's views. Keep the dataset name and selected repository revision alongside them. record_id survives descriptive filename corrections; work-ID correction and supersession follow the registry contract. RecNN specifies order within a documented group and does not itself establish session boundaries or completeness.

Reading dates and recording starts

Canonical filename times are documented finish labels. Under standard_individual_recordings, every member shares the complete drawing/group's finish label; it is not an individual recording's start or stop. Historical finish-label scope may be unspecified.

• —recording_start_local is a reviewed original device-filename timestamp, interpreted in the declared timezone. It is not an independently calibrated clock; blank means unavailable or outside reviewed coverage.
• —recording_start_basis distinguishes a filename-derived start, a reviewed unknown, and not_in_timing_snapshot for records outside registered coverage.
• —documented_finish_date and documented_finish_time retain the original date and time; finish_label_scope states their documented scope.

Use the current timing view for coverage and evidence. MP4 creation tags retain their creation/finalization meaning. Encoded playback duration and frame rate do not establish wall-clock duration, active drawing time, or capture intervals. No missing start is inferred from a shared finish label or playback duration.

Evidence and limits

The single-creator archive supports examination of this documented practice; generalization to other artists or settings requires additional evidence. Acquisition and documentation vary across workflows, and workspace appearance may provide visual shortcuts for models. Downstream task performance and population, clinical, biometric, authorship, or leakage-safe benchmark claims require their own validation.

The current source release contains no frame-level bounding boxes, segmentation masks, pose, hand, pen-tip, stroke, aesthetic, or style labels. Small annotation-feasibility studies may be developed as derived work; any resulting labels remain separate from the source archive.. The archive lacks universal per-record capture calibration, original per-frame timestamps, camera calibration, and pixel-to-physical mapping. The nominal 0.2-second protocol setting is not a confirmed per-record measurement. Temporal or physical-motion claims must be supported by the calibration they require.

The schema defines field meanings, semicolon-separated tool/medium sets, stored-video orientation, and missing-value rules. Tool and medium lists do not guarantee positional pairing. The provenance index locates sequence evidence and historical timing layers; the physical-reference sidecar records supported photograph relationships, with reference media availability stated separately. See acquisition specifications for defaults and known exceptions. Pin derived annotations to source IDs and an exact revision.

Integrity and rights

`SHA256SUMS` records ordinary byte-level SHA-256 for published MP4 and ZIP payloads. It excludes documentation, scripts, CSV indexes, and itself; those files are versioned by the repository commit. hub_xet_hash is a distinct Hub storage identity. A checksum establishes byte identity with a comparison file; provenance requires evidence about that file's origin.

For reproducible use, retain the full repository commit, selected record_id values, corresponding checksum entries, and Xet identities. Document decoding, rotation, cropping, sampling, proxy generation, and whether audio was retained. The verified downloader guide shows a metadata-only check before intentional media transfer.

Audio streams occur in 486 of 873, or 55.7% of indexed files. has_audio reports stream presence, not meaningful signal. Review incidental audio, visible references, and identifying or third-party content before redistribution or model release. Derived proxies should remove audio by default while preserving archival sources.

The creator confirms ownership of the published videos and depicted creator-owned drawings. This does not independently grant rights in incidental third-party content. The dataset is released under CC BY-NC 4.0, applying only to material for which the licensor holds the necessary rights. See the license notice; commercial use outside its terms is not permitted.

Citation and documentation

DOI 10.57967/hf/10529 identifies the September 20, 2026 snapshot (845 indexed video files; revision `f6d92df`). The latest tagged release remains v0.7.0 (831 files). The earlier DOI, 10.57967/hf/10449, identifies v0.5.1 (709 files). Cite the exact snapshot used. The release guide contains both DOI citations, snapshot comparisons, and validation-revision details.

Field contracts live in SCHEMA.md; maintenance instructions in the workflow guide and current timing guide. Validation receipts, the changelog, and the supporting-documentation index preserve the detailed history.