DefectLens-H3 / README.md
XuanhaoChang's picture
Add defect ground-truth annotations and schema
05fb7de verified
|
Raw History Blame Contribute Delete
4.86 kB
metadata
language:
  - en
  - zh
pretty_name: DefectLens-H3
size_categories:
  - n<1K
tags:
  - video
  - audio
  - multimodal
  - video-generation
  - minimax-h3
  - defectlens
configs:
  - config_name: default
    default: true
    data_files:
      - split: test
        path: data/**
    drop_labels: true

DefectLens-H3

MiniMax H3 generated audio-video clips with the actual generation prompts, ordered reference images, original prompts, and generation parameters.

This release contains 88 clips: 69 reference-to-video-and-audio (Ref2VA) and 19 text-to-video-and-audio (T2VA) samples. The MP4 files retain their generated audio tracks. There are 242 ordered reference inputs, stored as 227 distinct image files after content deduplication.

Contents

  • data/videos/sample_XXX.mp4: generated clips, preserving original bytes.
  • data/metadata.jsonl: one record per clip, indexed by the original sample_id.
  • references/: the actual input images, addressed by SHA-256.
  • generation/requests/: complete portable generation requests.
  • generation/config.json: generator, adapter, scheduler and runtime settings.
  • verification/: media checks and clip checksums.
  • checksums.sha256: checksums of all release files except itself.

file_name in data/metadata.jsonl is relative to data/. reference_images, generation_request, and verification are relative to the repository root. Reference image order is significant: index 0 corresponds to <Picture 1>, index 1 to <Picture 2>, and so on. T2VA records have empty reference lists. Numeric sample IDs retain gaps from the generation manifest.

Loading

The test split denotes evaluation material; no training split is implied. The repository uses the Hugging Face VideoFolder layout with JSONL metadata: https://huggingface.co/docs/hub/datasets-video

from datasets import load_dataset, Video

ds = load_dataset("XuanhaoChang/DefectLens-H3", split="test")
ds = ds.cast_column("video", Video(decode=False))
print(ds[0]["sample_id"], ds[0]["prompt"])

To work directly with the original media and ordered reference files:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download("XuanhaoChang/DefectLens-H3", repo_type="dataset"))
rows = [json.loads(line) for line in (root / "data/metadata.jsonl").read_text().splitlines()]
item = rows[0]
video = root / "data" / item["file_name"]
references = [root / path for path in item["reference_images"]]
prompt = item["prompt"]

Prompts and task adaptation

prompt is the exact prompt used to generate the corresponding clip; original_prompt is retained as task provenance. 31 requests were adapted to the available input conditions. For these records, the original prompt may contain requirements that were changed or omitted; evaluate the generated clip against prompt. The release includes only the reference media actually used. It does not reconstruct missing original inputs or establish identical task pairing with other generators.

Generation

  • Base: MiniMax H3, using the existing local ModelScope distribution.
  • Adapter: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
  • Checkpoint: minimax_h3_turbo_v4_step600_ema.safetensors, strength 1.0, bypass.
  • Six model evaluations; simple sigma lookup, video shift 12 / audio shift 3.
  • Seed, native canvas dimensions, frame count, duration and workflow are stored per sample.
  • Native Diffusers runtime with eight-GPU Ulysses context parallelism and FSDP2.
  • Generated streams: 24 FPS video and 32 kHz stereo audio.

Temporary inference tensors, model weights and API credentials are not part of the dataset. Media verification records concern file integrity and stream structure; the owner has reported a spot check of the generated clips.

Ground truth

annotations/gt.jsonl contains defect annotations for 88 videos, with 259 defect events. Each record includes sample_id, the video path, the video SHA-256, and an issues list. Records without labeled defects have an empty list. Join annotations with data/metadata.jsonl using sample_id.

Issues include a description, coarse and fine defect types, a temporal interval, and available keyframe and bounding-box references. Box coordinates use normalized [x1, y1, x2, y2] format, and each box is associated with its keyframe timestamp. Field definitions are in annotations/schema.json.

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("XuanhaoChang/DefectLens-H3", "annotations/gt.jsonl", repo_type="dataset")
gt = [json.loads(line) for line in open(path, encoding="utf-8")]
print(gt[0]["sample_id"], gt[0]["issues"])

License

A dataset license has not been specified by the owner in this release. Model and adapter licensing is documented by their respective distributors.