README.md

---
license: other
license_name: research-only
license_link: LICENSE
task_categories:
- video-to-video
- audio-to-audio
- text-to-video
language:
- en
- zh
tags:
- audio-visual
- video-editing
- instruction-following
- multimodal-reasoning
- benchmark
pretty_name: ReasonAVEdit-Bench
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: metadata.csv
---

# ReasonAVEdit-Bench

A 1,200-sample benchmark for **instruction-guided joint audio-video editing**, built to measure
not only the final edit but the **multimodal reasoning** behind it: *which region* to change,
*which sound source* to change, and *when* the change happens.

Every sample ships the source audio-video, a natural-language instruction, and two decodable
reasoning ground truths:

* **single-frame visual reasoning** - the edit-start frame with the target region flattened to
  gray 127, plus a binary mask of that region;
* **full-timeline acoustic reasoning** - a waveform the same length as the clip that is silent
  outside the edit window and, inside it, contains only the source to be replaced.

No edited target video is provided. This is a benchmark: the input is a real clip plus an
instruction, and the edited result is what a method under test is expected to produce.

## Contents

| Path | Count | Description |
|---|---|---|
| `videos/` | 1200 | Source clip, 25 fps, `frames % 8 == 1`, sides multiple of 32, H.264 + 16 kHz AAC |
| `audios/` | 1200 | Source audio, 16 kHz mono WAV |
| `masks/` | 1200 | Full-timeline binary mask of the target region |
| `reason_visual/` | 1200 | Single-frame visual reasoning GT (target region = gray 127) |
| `reason_visual_mask/` | 1200 | Binary region at the anchor frame, for Visual IoU |
| `reason_audio/` | 1200 | Full-timeline acoustic reasoning GT, 16 kHz mono WAV |
| `ref_images/` | 150 | Reference image (reference-guided and instance-addition tracks) |
| `annotations/` | 1200 | Per-sample JSON: instruction, windows, geometry, measured evidence |
| `metadata.csv` | 1 | One row per sample |
| `review.html` | 1 | 80-window visual inspection page (serve over HTTP) |
| `train_exclude.list`, `train_exclude_videoids.txt` | 2 | Leakage-control lists for the source corpora |

Total 3.3 GB. All media are real files; nothing is a symlink.

### Media is shipped as tar archives

On the Hub the seven media folders live under `archives/` as one tar each, because 7,150 loose
LFS objects trip the Hub's request rate limit. Unpack them next to `metadata.csv` and the paths
in the CSV resolve as written:

```bash
huggingface-cli download bigfacing/ReasonAVEdit-Bench --repo-type dataset --local-dir ReasonAVEdit-Bench
cd ReasonAVEdit-Bench && for f in archives/*.tar; do tar -xf "$f"; done
```

`annotations/`, `metadata.csv` and `README.md` are stored loose so they stay browsable on the
Hub. `review.html` is an 80-window inspection page that expects the extracted layout, so open it
after unpacking and serve the folder over HTTP (`python3 -m http.server`).

## Tracks

Five mutually exclusive categories; a source video appears in exactly one sample.

| Category | Sub-track | N | What is hard |
|---|---|---|---|
| **Where-only** 250 | `W1_same_class_pair` | 100 | Two instances of the same class, comparable size - only position disambiguates |
| | `W2_spatial_reference` | 80 | Instruction refers by left/right/larger/smaller |
| | `W3_small_or_occluded` | 70 | Target covers < 2% of the frame |
| **Which-only** 300 | `H1_overlapping_voices` | 120 | A second voice overlaps the target speaker |
| | `H2_speech_vs_ambient` | 100 | Speech competes with ambient sound at a comparable level |
| | `H3_offscreen_source` | 80 | The source to edit is never visible; visual GT is a null anchor |
| **When-only** 300 | `T1_word_phrase_swap` | 200 | Replace 1-4 words inside a sentence |
| | `T2_sentence_rewrite` | 100 | Rewrite part of a sentence |
| **Combined** 250 | `C1_overlap_local_speech` | 100 | Pick the right voice *and* the right time span |
| | `C2_ambient_local_speech` | 50 | Same, against ambient interference |
| | `C3_animal_local` | 40 | Intermittent non-speech source |
| | `C4_spatial_add` | 50 | Add an instance at a required position |
| | `C5_complex_remove` | 10 | Remove an instance from a busy scene |
| **Reference** 100 | `R1_person_identity` | 50 | Swap identity to a reference person |
| | `R2_animal_object_look` | 30 | Swap an animal/object to a reference appearance |
| | `R3_ref_guided_add` | 20 | Add the reference subject to an empty scene |

Balanced across languages (zh 605 / en 595) and orientation (landscape 952 / portrait 248).

## Suggested metrics

Final edit quality applies to all samples (FVD, TV-A, TC, SSIM, FAD, TA-A, LPAPS, AV-A, PEAVS,
Sync-C/D, PQ, BG, DeSync). The reasoning GT enables per-category diagnostics:

* Where-only: Visual IoU, BG
* Which-only: Which-Acc, Acoustic Time-Frequency IoU, LPAPS
* When-only: Acoustic Temporal IoU, Video Temporal IoU, DeSync, A/V support offset
* Combined: all three groups

The 1,100 pure-instruction samples are intended for the main ranking; the 100 reference-guided
samples are reported separately, since they take an extra conditioning input.

## How the annotations were derived

Selection is grounded in measurements, not caption guesswork:

* **Instance count and geometry** come from connected components of the SAM2 binary mask, so
  "the dog on the left" is a checkable claim and the visual GT keeps only the named instance.
* **Competing sources** come from the source/target/residual decomposition shipped with the
  underlying corpora: level ratio and temporal overlap are measured directly.
* **Voice vs non-voice** is decided by Whisper's `<|nospeech|>` posterior. DSP heuristics
  (voiced-band energy times syllabic modulation) confuse animal calls with speech at a 30-43%
  false-positive rate, which is exactly the distinction the Which-only track rests on.

Off-screen sources are annotated on the residual track rather than the mask: an object grounded
by SAM2 is by construction visible, so a genuinely unseen source lives in "source audio minus
visible target". For those samples the acoustic edit target is the residual and the visual
reasoning GT is a null anchor.

## Loading

```python
import csv, os
root = "."  # or the snapshot path from huggingface_hub.snapshot_download
rows = list(csv.DictReader(open(os.path.join(root, "metadata.csv"))))
r = rows[0]
print(r["sub_track"], r["instruction"])
video   = os.path.join(root, r["video"])              # model input
audio   = os.path.join(root, r["audio"])              # model input
rv_gt   = os.path.join(root, r["reason_visual"])      # single-frame visual reasoning GT
rv_mask = os.path.join(root, r["reason_visual_mask"]) # binary region for IoU
ra_gt   = os.path.join(root, r["reason_audio"])       # full-timeline acoustic reasoning GT
```

`annotations/{id}.json` additionally carries the edit window in seconds and frames, the anchor
frame index, per-instance geometry, the selected instance index, and the measurements that
placed the sample in its sub-track.

## Notes and limitations

* Instructions for the SAM-derived tracks are templated from measured evidence (grounded object
  phrase, component geometry, a donor drawn from a different category, a donor line taken from
  real in-domain ASR). They are grammatical and specific but not human-written.
* `C4`/`R3` are instance-addition cases: nothing in the source needs replacing, so the acoustic
  reasoning GT is a silent null anchor and the visual GT is the placement box.
* Clip length follows the ladder 121/105/97/89/81/73/65 frames at 25 fps, picked as the longest
  rung a source supports; 990 of 1200 samples are the full 121 frames.
* Source material is drawn from internal talking-head, film/drama and sounding-object corpora.
  Redistribution beyond research use may be restricted by the terms of those sources.