time-lapse-artifacts

677 time-lapse recordings of one artist's traditional drawing practice across nearly two years (September 2024–September 2, 2026), with standardized acquisition since July 13, 2025. Across 2,083,318,611,246 bytes (approximately 2.08 TB) of video, the archive follows the same artist working in physical media—including pencil, ballpoint pen, fountain pen, brush, graphite, and ink—on fixed-format paper and illustration board. The release combines 549 canonical Standard videos and 128 earlier Pre-Standard recordings. Every active indexed video has a persistent record ID, content hash, byte size, and encoded-stream metadata; historical session and provenance fields remain less complete in the Pre-Standard tier.

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. When documented, fragment-sequence and physical-reference relationships remain in separate revision-pinned sidecars joined by record_id; absent relationships remain unknown.

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

In this dataset, provenance refers to the documented relationship between a physical drawing and its time-lapse record, rather than to histories of ownership or sale.

An earlier holding associated with the 2012–2016 archival scope comprises 2,381 locally held real-time recordings totaling 9,060,310,121,597 bytes (approximately 9.06 TB). It consists primarily of uploaded 1080p drawing videos, with some livestream-derived material. Its file-level holdings and SHA-256 manifest is published in this repository; the media itself is not hosted and remains excluded from the current 677-record release totals.

Snapshot and Coverage

The table below describes the v0.4.0 release candidate. Batch-level history is maintained in CHANGELOG.md. The immutable v0.3.0 tag remains a 467-row historical release and does not describe this snapshot.

Collection tier	Published MP4s	Indexed rows	Date coverage	Metadata state
standard	549 direct files	549	2025-07-13–2026-09-02	Complete core technical metadata for every direct Standard file
pre_standard	128	128	2024-09-17–2025-07-09, plus 9 unknown dates	Complete core technical metadata; historical descriptive and provenance gaps remain
Earlier archive (metadata-only holding)	0 hosted videos	0 dataset rows; 2,381 fixity entries	Historical scope label 2012–2016	9,060,310,121,597 bytes held offline; file-level SHA-256 manifest published

The Hub configurations use the split name archive. No official train/validation/test partition has been defined.

Viewer and Metadata Discovery

The release exposes two explicit configurations so metadata-complete canonical records are not silently mixed with historically less complete records:

• canonical/archive is the default and loads root metadata.csv with the 549 direct files in Standard_Time_Lapses/
• pre_standard/archive loads Pre_Standard_Time_Lapses/metadata.csv with the 128 direct files in that folder

Each configuration has the same declared 29-column schema. Stored dimensions and byte sizes are nullable integers, has_audio is a nullable boolean, and encoded frame rate remains a float because historical encodings include fractional reported rates. Undocumented values load as null rather than as zero or false. Both indexes declare tier membership directly: the root index sets collection_tier = standard and is_canonical = true; the Pre-Standard CSV sets collection_tier = pre_standard and is_canonical = false.

Published-index completeness is defined by the 549 rows in canonical/archive plus the 128 rows in pre_standard/archive and their generated Parquet exports. Hugging Face's separate column Statistics job is an auxiliary, asynchronous surface; its availability is not used to claim or reject row-level completeness.

The configuration lists the two active metadata.csv files and only direct MP4 paths in Standard_Time_Lapses/ and Pre_Standard_Time_Lapses/. It does not use a recursive canonical glob, so the nested Standard_Time_Lapses/Provenance/ subtree is excluded.

Overview

time-lapse-artifacts documents analog drawing processes across different recording periods and acquisition conditions. Video is the primary published evidence. Metadata identifies, organizes, compares, and documents recordings without imposing aesthetic scores or stylistic labels.

The collection is a single-creator longitudinal archive. This makes it useful for studying within-practice change, but it does not represent artists, materials, studios, or drawing practices broadly.

Repository Structure

Standard_Time_Lapses/

The directory contains 549 direct MP4 files. metadata.csv is their complete machine-readable canonical index: every direct file has one validated row with core technical metadata. Direct placement alone is not sufficient for future additions; metadata and release validation are required.

Each canonical metadata row now includes:

• a persistent record_id
• collection_tier = standard
• is_canonical = true
• an acquisition_protocol_version: 363 earlier-workflow rows use standard-2025-07-13, while the 186 canonical fragments in the six dated relationship sidecars use standard-2026-08-27
• descriptive session fields
• encoded video properties
• repository content hash and size
• provenance fields where documented

Files use a structured filename convention described in SCHEMA.md. A filename is descriptive, but it is not the stable identity of a record; external references should use record_id.

Standard_Time_Lapses/Provenance/

This subtree contains supporting historical records and retained exceptions. It is not part of the canonical record set.

2026-07-31/ preserves the completed provenance release, including mappings between historical Series_9x12 filenames and canonical records. The former full duplicate video snapshot is not retained. Relationships are preserved through mappings, hashes, metadata, and repository history. See the provenance navigation index.

Pre_Standard_Time_Lapses/

These 128 recordings predate the standard acquisition protocol. They are published for archival continuity but are not interchangeable with canonical standard records. Each now has a stable record_id, exact Hub content hash and size, playback duration, stored dimensions and orientation, frame rate, codec, pixel format, container, and audio-stream presence. Historical dates, times, materials, and provenance remain incomplete where evidence is unavailable. Missing values must not be interpreted as zero, false, or evidence that a property was absent.

The folder-local metadata.csv is the active index for these 128 videos and is loaded by the pre_standard/archive configuration. It follows the same declared 29-field schema as the canonical index while leaving undocumented historical fields empty.

Material that does not meet the archive's ownership and provenance threshold is excluded from both the published media and the active metadata index.

Earlier_Archive_2012_2016/

An earlier holding of 2,381 files totaling 9,060,310,121,597 bytes is held outside the hosted-media release. It consists primarily of uploaded 1080p drawing videos, with some livestream-derived material. The repository publishes a file-level holdings and SHA-256 manifest, not the media itself. These files remain excluded from active dataset rows and published media totals. References to 2012–2016 describe the historical scope label and do not assert a capture year for every listed file.

Metadata and Provenance

The metadata distinguishes documented facts, established historical relationships, and unresolved information. Evidence may include:

• exact file/blob identity
• original filenames and repository paths
• file sizes and hashes
• duration comparison
• visual continuity across recordings
• documented physical references
• repository commit history

Uncertain relationships are not promoted to verified relationships solely because dates or filenames look similar.

The active CSV files and live Viewer use semicolons for multiple tool and medium values, for example Pencil;FountainPen and Graphite;Ink. A pipe (|) is not a valid separator in v0.2. The two fields are independent observed-value lists; equal list lengths and positional pairing are not guaranteed. See SCHEMA.md.

The Standard protocol documents a nominal 0.2-second capture interval. This is a protocol-level default, not a confirmed per-record measurement, and it does not establish a calibrated real-time basis. encoded_frame_rate remains encoded playback rate, not original drawing-time sampling, and must not be used to infer real drawing time unless separately documented.

Physical-reference relationship layer

Physical_References/endpoint_references.csv links 309 Standard video records to owner-confirmed photographs of their physical drawings through persistent record_id values and the pinned source revision. It is a separate derived relationship table and does not alter the 29-field canonical video index.

Version 0.1 records 52 confirmed endpoint photographs, one longitudinal continuation, 232 owner-confirmed record references whose exact endpoint state is not separately asserted, and 24 two-sided references. The photographs are not bundled in this release. Their reviewed filenames and available hashes preserve relationship evidence without implying that the media are publicly downloadable. See the sidecar schema and manifest.

Acquisition and Stored-Video Orientation

The earlier Standard workflow established July 13, 2025 is identified as standard-2025-07-13. The prospective canonical fragment-sequence workflow established August 27, 2026 is identified as standard-2026-08-27; this is a sidecar-defined membership boundary, not a blanket finish-date cutover. Framelapse source fragments remain unchanged acquisition evidence. For reviewed fragment ingestions, a drawing may be represented canonically either by one selected chronological assembly or by a complete, verified RecNN sequence of unchanged source fragments. Grouping, order, completeness, and noncanonical derivative coverage for those ingestions remain in separate dated relationship manifests; the active 29-field metadata schema is unchanged.

All RecNN members in one prospective grouping share the documented finish date and finish time of the complete grouping/physical work. The timestamp is not the individual end time of each fragment; RecNN records chronological position within the group. The manifests do not backfill every historical RecNN record. The 19 historical RecNN rows outside the six sidecars retain standard-2025-07-13; their grouping or completeness claims remain limited to evidence explicitly recorded in notes or other documentation. A documented 4K/30-fps InShot processing profile remains a candidate and does not retroactively redefine existing records. See the capture and ingest protocol.

The dated relationship sidecars preserve each ingestion as a separate pinned observation layer. The 2026-08-27 sidecar has 27 rows: 22 canonical fragments across five complete sequences and five noncanonical InShot derivatives. The 2026-08-29 sidecar has 19 canonical fragments across four complete sequences and no derivative rows. The 2026-08-30 sidecar has 39 canonical fragments across five complete sequences and no derivative rows. The 2026-08-31 sidecar has 32 canonical fragments across four complete sequences and no derivative rows. The 2026-09-01 sidecar has 13 canonical fragments across two complete sequences and no derivative rows. The 2026-09-02 sidecar has 61 canonical fragments across five complete sequences and no derivative rows. In all six, fragment_set_label is a descriptive grouping label, not a global work_id. Together they identify exactly 186 canonical fragments for standard-2026-08-27. The nested derivative payloads are not included in the 677 active rows or the active byte and duration totals.

The protocol's capture orientation and a file's stored pixel orientation are different concepts. The orientation, width_px, and height_px metadata describe the published encoded file. Historical canonicalization may include a rotation, crop, or transcode; those transformations are documented only where evidence currently exists and must not be inferred from dimensions alone.

The July 13, 2025 protocol boundary is explicitly encoded in canonical metadata. It supports era-aware comparison of repeated physical-media practice, but it is an acquisition change, not a skill-change label. Differences across that boundary may reflect capture conditions as well as practice development.

Annotation State

Session-level descriptive and provenance metadata are curated by the creator from the physical practice, filenames, file properties, and retained historical evidence. This is the scope of annotations_creators: expert-generated; it does not mean that frame-level task labels exist.

The videos themselves are unannotated. Processing history varies and is not uniformly documented. The release does not provide:

• frame-level bounding boxes or segmentation masks
• hand, pen-tip, pose, or stroke labels
• aesthetic or quality scores
• imposed stylistic classifications
• calibrated physical trajectories
• original capture timestamps for individual time-lapse frames

Researchers may derive annotations, but derived labels should identify the source record_id and pinned repository revision.

Intended and Unsupported Uses

Reasonable uses in the current release

• archival and drawing-process research
• qualitative within-creator analysis of practice and skill development across repeated physical-media sessions
• self-supervised video representation experiments
• frame-change, scene, and workspace analysis
• development of annotation and tracking methods

Uses requiring additional validation or calibration

• hand or pen-tip tracking
• optical-flow comparison across acquisition eras
• temporal behavior analysis
• real-world trajectory, velocity, acceleration, or motor-control measurement

The release does not provide confirmed per-record capture intervals, acceleration factors, source-frame timestamps, camera calibration, or pixel-to-physical coordinate mapping. The nominal Standard interval alone is not calibrated real-time evidence.

Not validated

• population-level claims about artists or drawing behavior
• clinical, biometric, authorship, or identity inference
• a leakage-safe training or evaluation benchmark
• commercial use outside the CC BY-NC 4.0 terms

Limitations and Sources of Bias

• The archive documents one creator and one evolving practice.
• Current public video coverage begins in 2024, not 2012.
• The standard and pre-standard tiers have substantially different historical descriptive/provenance completeness and acquisition conditions.
• The workspace can contain hands, tools, phones, tablets, reference material, and other objects that may create visual shortcuts or occlusion.
• Stored resolution and frame rate do not establish physical or temporal calibration.
• Large files make exhaustive inspection and replication expensive.
• No official leakage-safe split exists.

Rights, Privacy, and Audio

The creator confirms ownership of the published video files and depicted creator-owned drawings. Material that cannot pass the ownership and provenance review is not part of the active public index and is not distributed by this release. This statement does not independently grant rights in incidental third-party material that may appear on screens, in references, or in captured audio.

Most active indexed videos at this revision contain an audio stream (485 of 677, or 71.6%). has_audio records stream presence only; it does not establish that the stream contains meaningful signal. Users are responsible for reviewing audio, visible screens, references, reflections, tattoos, and other potentially identifying or third-party content before redistribution or model release. A future proxy release should remove audio by default while preserving the archival video objects.

Access and Reproducibility

The active 677-video release contains 2,083,318,611,246 bytes of video (approximately 2.08 TB, or 1.89 TiB). Canonical files are often several gigabytes each, so users should select records through metadata.csv before downloading video.

Bandwidth-safe Quick Start

Install Hugging Face Datasets, stream the archive, and disable decoding while selecting records:

from datasets import Video, load_dataset

dataset = load_dataset(
    "maxwellinked/time-lapse-artifacts",
    "canonical",
    split="archive",
    streaming=True,
)
metadata = dataset.cast_column("video", Video(decode=False))
first = next(iter(metadata))
print(first["video"]["path"], first["tool"], first["medium"])

The CSV contract names the repository path column file_name. When a declared configuration is loaded, Hugging Face resolves that reserved metadata-file column against the included MP4 paths and exposes the typed feature as video. That is why the example uses first["video"] rather than first["file_name"].

This exposes the complete metadata schema without transferring video bytes. Default decoded row access additionally requires torchcodec and a compatible video-decoding stack. Use streaming=True unless a full archive download is intentional.

Use "pre_standard" as the configuration name to stream the earlier tier.

Verify the Published Payloads

SHA256SUMS contains the ordinary byte-level SHA-256 of every published MP4 and ZIP payload. It is a portable fixity manifest, independent of Hugging Face Xet storage, and is not encryption. The manifest intentionally excludes documentation, scripts, CSV indexes, and itself.

After downloading the complete archive, hash all payload bytes and compare them with the manifest:

python tools/validate_sha256sums.py SHA256SUMS --verify-root .

To validate complete manifest coverage against a pinned Hub inventory without downloading roughly 2 TB of media:

python tools/validate_sha256sums.py SHA256SUMS \
  --repo-json repo_api_blobs.json \
  --revision "<40-hex-commit>"

The inventory must come from that exact immutable revision with blob metadata enabled. This check compares every manifest digest with the LFS-compatible content SHA-256 and requires an exact path match for every .mp4 and .zip.

Download and Verify One Record

Install huggingface_hub, then use the included helper with a stable record_id:

python -m pip install huggingface_hub

The helper resolves the requested revision to an immutable commit, checks the indexed byte size and Xet identity against the pinned Hub headers, and reports the ordinary LFS-compatible SHA-256 separately.

Check a record without downloading its video:

python tools/download_verified.py tla-a66bdad6c903b89f --check-only

Download that record and verify its local byte size and SHA-256:

python tools/download_verified.py tla-a66bdad6c903b89f \
  --revision "<commit-or-tag>" \
  --output-dir downloads

hub_xet_hash is an Xet file ID, not the SHA-256 of the downloaded bytes. The helper verifies both identities without conflating them. Video downloads may be many gigabytes; use --check-only first.

For reproducible work:

1. Pin a repository commit or release tag.
2. Record the selected record_id values.
3. Preserve the selected entries from SHA256SUMS and the published hub_xet_hash values; they are distinct identities.
4. State whether audio was retained.
5. Document all rotation, crop, sampling, decoding, and proxy-generation steps.

Run the included validator before publishing a metadata change:

python tools/validate_metadata.py metadata.csv \
  --pre-standard-metadata Pre_Standard_Time_Lapses/metadata.csv \
  --fragment-relationships-root Standard_Time_Lapses/Provenance

tools/enrich_prestandard.py reproduces the Pre-Standard technical fields from a pinned Hub inventory and MP4 container headers. It uses HTTP range requests through FFmpeg; it does not intentionally download complete videos. tools/capture_hub_inventory.py records the corresponding Xet file IDs from resolve headers without following redirects or transferring video payloads.

The v0.2 index can be regenerated from the pinned 335-row source index plus the three audited additions:

python tools/repair_metadata.py source_metadata.csv metadata.csv \
  --additions tools/metadata_additions_v0.2.csv

Versioning

The v0.4.0 release candidate establishes the current 677-record snapshot after the reviewed post-v0.3 ingests, provenance corrections, and focused Standard protocol-version boundary. Record identity is carried by record_id; a filename may be corrected without creating a new conceptual record. Content replacement must retain provenance and should receive a new content hash.

See CHANGELOG.md for release-level changes.

Documentation

• metadata.csv — canonical machine-readable index
• Pre_Standard_Time_Lapses/metadata.csv — active pre-standard public index
• SHA256SUMS — ordinary SHA-256 for every published MP4 and ZIP payload
• LICENSE — CC BY-NC 4.0 notice and rights-scope qualification
• SCHEMA.md — filename and metadata contracts
• Archive_Specifications.md — acquisition specifications and known gaps
• VALIDATION.md — validation scope and current results
• VIEWER_MIGRATION.md — current Viewer configuration and verification invariants
• RELEASE_MANIFEST_V0.4.0.json — current v0.4.0 release-candidate identity, coverage, and critical-file receipt
• RELEASE_MANIFEST.json — historical v0.2.0 release-candidate manifest; it does not describe the current snapshot or provide video-payload checksums
• tools/ — inventory capture, verified download, checksum, enrichment, and validation helpers
• Standard_Time_Lapses/Provenance/ — historical evidence and dated fragment-sequence relationships
• Physical_References/ — physical-reference relationship schema, manifest, and sidecar

Citation

Until a DOI-backed release is available, cite the dataset and the exact v0.4.0 tag or repository revision used:

@dataset{maxwellinked_time_lapse_artifacts_2026,
  author    = {maxwellinked},
  title     = {time-lapse-artifacts},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/maxwellinked/time-lapse-artifacts},
  note      = {v0.4.0 release candidate; include the exact tag or repository revision used}
}

License

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 conduct their own review of incidental third-party content and downstream use.