changwangss/opens2v-calibration

OpenS2V calibration manifest

This directory contains a small normalized calibration subset sourced from BestWishYsh/OpenS2V-5M. It is intended to prototype AutoRound diffusion T2V and I2V calibration without requiring the full 11 TB upstream dataset.

Format

output/opens2v_calibration.tsv uses AutoRound's normalized schema:

• id: stable upstream sample identifier
• caption: metadata.face_cap_qwen, falling back to metadata.cap[0]
• image: relative path to a reference frame extracted from the source video

T2V uses id and caption; I2V additionally uses image. output/prompts.jsonl exposes the same records with an explicit prompt field for direct inspection and reuse outside AutoRound. output/provenance.json records source paths, scores, frame/crop information, selection rules, and license metadata.

Selection

The manifest contains 128 samples selected deterministically with seed 42 across all 13 OpenS2V parts. It follows OpenS2V's published high-quality duration, resolution, and aesthetic rules, adds a conservative motion-score guard, and uses at most one clip per source video. A greedy stratified selection balances source parts and caption-derived duration, motion, aesthetic, indoor/outdoor, single/multiple-subject, static/dynamic-camera, and interaction dimensions. These caption-derived labels are sampling heuristics, not OpenS2V annotations. Reference images are cropped from the midpoint of face_cut to avoid cut-boundary and fade frames.

The candidate pool uses complete videos found in the first 128 MiB of the first archive shard for each of the 13 parts. The result has broader coverage than a single-part subset, but it is not a uniform random sample of all 5M records.

Rebuild

Run build_manifest.py with a directory containing OpenS2V total_part*.json metadata and a folder containing extracted source videos:

python build_manifest.py \
  --metadata /path/to/Jsons \
  --videos /path/to/extracted/videos \
  --output ./output \
  --nsamples 128 \
  --seed 42

The upstream dataset card declares CC-BY-4.0. Preserve attribution and review the upstream terms before redistributing the derived manifest and frames.