Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
2.6
24.4

Time-Lapse Artifacts: Derived Access Layer

This repository contains access derivatives and bounded inspection evidence for maxwellinked/time-lapse-artifacts. The source dataset remains authoritative for source facts and evidence records. Existing artifacts/ packs derive from originals published in that source dataset. The separately scoped Drawing_Process/ pack publishes excerpts whose original-byte identities are catalogued there; its original media payloads have not yet been uploaded. Its manifest records that availability explicitly.

Related interfaces

The browser is a static presentation layer and may lag the current Hugging Face dataset. Hugging Face remains authoritative for media, record identities, and machine-readable metadata.

This construction release is deliberately narrower than a benchmark. Its first purpose is to test whether a deterministic pipeline can make the visual record inspectable without collapsing encoded process into a static endpoint or claiming that encoded playback is real drawing time.

Authority boundary

The repository keeps five kinds of information separate:

  1. Source identity and curator-verified facts
  2. Explicitly unresolved states
  3. Deterministic derivatives and measurements
  4. Human interpretations
  5. Machine interpretations

Only the first three are in scope for the initial pilot. Model captions, embeddings, style labels, quality judgments, and benchmark claims are out of scope.

Initial outputs

For each processed record_id, the pipeline creates:

  • a silent H.264 access proxy fitted within 1280 Γ— 1280 pixels
  • 21 JPEG frames sampled at normalized encoded positions from 0.00 to 1.00
  • exact SHA-256 hashes and deterministic 64-bit difference hashes for frames
  • adjacent-frame Hamming distances, reported as measurements rather than interpreted as meaningful or meaningless change
  • a JSON derivation manifest pinned to the source repository revision, published Xet identity, Hub LFS SHA-256, and independently computed download SHA-256

Normalized encoded position is not real drawing time. The final sampled frame is an encoded endpoint, not a claim that the complete process was captured.

The 21 frames are a fixed access profile within each processed video, not 21 independent records or a statistical sample size. Increasing frame density would improve temporal resolution, not record-level coverage.

Pipeline validation set

data/pipeline_validation_set.csv is an engineering test fixture, not a representative sample. Records are selected by declared rules designed to expose pipeline failures: duration and file-size extremes, median cases, tier, codec, orientation, audio-stream presence, documented shortened derivatives, and filename-level multi-recording cases.

The six low-volume Wave 1 records exercise short-duration, shortened, Pre-Standard, HEVC, landscape, and no-audio conditions before costly stress tests. Large and multi-recording cases remain in Wave 2 until Wave 1 establishes actual runtime and cost.

Wave 1 is published under artifacts/pilot-v1. It contains 6 silent proxies, 126 progress frames, 6 record manifests, summary.csv, and a run manifest. See PILOT_REPORT.md for the verified result and the failures that changed the pipeline.

Audio

source_has_audio records stream presence only. Proxies remove audio uniformly. This does not classify a source stream as music, speech, silence, or meaningful signal.

Work identity and completeness

record_id identifies a video object. work_id is reserved for a curator-verified physical drawing and remains null when that relationship has not been established. Rec01, shared dates, and similar filenames are not treated as proof of shared work identity or sequence.

Visible start state, visible end state, capture extent, and source-master status are separate fields. A completed-looking or signed endpoint never implies that the complete process was captured.

See DERIVATION_CONTRACT.md for the normative contract.

Reproduction

Generate the validation manifest from a pinned metadata snapshot:

python tools/select_validation_set.py \
  --standard-metadata metadata.csv \
  --prestandard-metadata Pre_Standard_Time_Lapses/metadata.csv \
  --source-revision 46e0be60f606635ee2b627be911bc98ab46565d3 \
  --output data/pipeline_validation_set.csv \
  --relationship-output data/work_relationship_review.csv

Run the tools from a complete, revision-pinned checkout. Generator 1.1 requires tools/generate_derivatives.py, its sibling tools/run_metrics.py, and schemas/derivation-manifest.schema.json; a single downloaded script is not a standalone Jobs entry point. For Jobs, provide the complete checkout in the container or mounted input. PEP 723 declares package dependencies, not these repository files.

Process Wave 1 locally (or from that checkout inside a Job):

python tools/generate_derivatives.py \
  --validation-repo maxwellinked/time-lapse-artifacts-derived \
  --validation-revision main \
  --pilot-wave 1 \
  --output-dir derived-output

The script records the complete FFmpeg build string. When a runtime lacks FFmpeg and FFprobe, its pinned PEP 723 dependency installs static binaries without elevated privileges.

Supplying --output-repo maxwellinked/time-lapse-artifacts-derived persists the results to the Hub. Each completed record is checkpointed before the next begins; run-level summary files are committed only after every selected record succeeds. --resume reuses a local checkpoint only after schema, source/documentary/build identity, canonical artifact paths, file sizes and every artifact SHA-256 pass. The local media are not decoded again during this checkpoint check; the independent validator recomputes media properties and frame hashes. A reused record still reaches publication, so a failed upload can be retried without encoding again. Resume requires the local output directory; it does not restore checkpoints from the Hub onto a new ephemeral filesystem.

Local completion, verified local bytes and confirmed publication are distinct. Publication checks record manifests and media identities at immutable commits and writes local publication_receipts/ and run_publication_receipt.json. Writes use an expected parent commit and stop on concurrent changes. Existing artifact identities or a different record selection cannot silently replace a published prefix. Choose a fresh --output-prefix for a new batch; retain artifacts/pilot-v1 as the historical six-record pack.

The separate run_metrics.json reports stage times, cache state, resolved dependencies and sampled resource use. Its scope explicitly distinguishes record processing/checkpoint measurements from final run-metadata publication; these measurements are not deterministic record-manifest contents. Source downloads remain in the selected cache (--cache-dir can isolate it); this tool does not evict source files or modify shared caches.

An explicit batch can be checked before upload with the independent validator:

python tools/validate_pilot.py --snapshot-root . \
  --prefix artifacts/expansion-v1-2026-11-10 --pilot-wave 2 \
  --record-id tla-0897038ffd1f516b --record-id tla-5fe34f800eee6bb4 \
  --local-only

For publication verification, omit --local-only and supply --revision with the immutable 40-character derived-repository commit. Online validation checks the exact published fixture, schema, summary and manifest bytes, source identity, and media hashes. The original six-record Wave 1 invocation remains the default.

Hub metadata requests and source downloads use four bounded attempts with exponential backoff; partial standard-HTTP downloads can continue on retry. Without --output-repo, no upload occurs.

Recording inspection pilot β€” October 25, 2026

A separate native-frame access pack contains 81 decoded PNGs from nine source recordings across three inspection cases. It supports the recording inspection view. The earlier six-record pilot is unchanged. The new pack is not a representative sample, benchmark, or source-media replacement; source revision, hashes and encoded timestamps are retained in its manifest.

Bounded expansion β€” November 10, 2026

A separate two-record engineering pack is available under artifacts/expansion-v1-2026-11-10. It adds two silent proxies, 42 uniform JPEG progress frames and source-bound manifests from the original Wave 2 fixture. Verified recovery, explicit batch validation and separate operational measurements are documented in the pack. The historical six-record pilot and nine-record inspection pack retain their original identities. The existing Dataset Viewer configuration continues to select only the historical pilot.

Native access companion β€” November 10, 2026

A separate native access companion adds 42 uncropped 2160 Γ— 3840 decoded RGB PNGs for the two bounded-expansion recordings, at 21 uniform first-to-last packet-PTS positions per recording. Each frame retains an exact selected PTS/time base, rational target, source binding and SHA-256; the existing proxies and JPEG profiles remain linked. Native packet-span targets and earlier JPEG duration targets may select different frames. The run report records warm-cache resource measurements and verification limits. Local generation, independently verified artifacts and confirmed publication remain distinct. The earlier packs and Dataset Viewer configuration retain their original identities.

Drawing process introduction β€” November 13, 2026

One drawing, plane by plane presents a six-photograph sequence and bounded real-time/time-lapse details alongside the artist's method account. The authoritative method and source catalog records original filenames, byte sizes, independently computed SHA-256 identities, qualified selectors and explicitly unresolved canonical identities. Original media payloads for this pack are not yet published; publishing these excerpts does not add the catalogued files to the canonical recording indexes.

This pack follows the separate Drawing Process Evidence Contract. Artist statements and visual observations retain their distinct evidence status. The earlier artifacts/ packs, their source requirements, and the default Dataset Viewer configuration remain unchanged. Media attribution is Tim Maxwell / maxwellinked under CC BY-NC 4.0, within the existing rights scope.

Downloads last month
403