Datasets:

bigfacing
/

ReasonAVEdit-Bench-Landscape

---
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 (Landscape)
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: metadata.csv
---
# ReasonAVEdit-Bench — Landscape Edition

**2,313 samples, every one of them landscape**, for instruction-guided joint audio-video
editing. Two halves:

| Half | N | What it is |
|---|---|---|
| Newly sampled | 1,200 | 15 reasoning sub-tracks, selected from measured evidence |
| Legacy tracks | 1,113 | Every landscape sample from the hand-curated AniAVEditBench tracks |

The benchmark measures not just the final edit but the **multimodal reasoning** behind it:
*which region* to change, *which sound source* to change, and *when* the change happens. No
edited target video is shipped — the input is a real clip plus an instruction, and the edited
result is what a method under test produces.

## Per-sample contents

| Path | Description |
|---|---|
| `videos/{id}.mp4` | Source clip. 25 fps, `frames % 8 == 1`, sides multiple of 32, H.264 + 16 kHz AAC |
| `audios/{id}.wav` | Source audio, 16 kHz mono |
| `masks/{id}.mp4` | Full-timeline binary mask of the target region |
| `reason_visual/{id}.png` | **Single-frame visual reasoning GT** — edit-start frame, target region flattened to gray 127 |
| `reason_visual_mask/{id}.png` | Binary region at the anchor frame, for Visual IoU |
| `reason_audio/{id}.wav` | **Full-timeline acoustic reasoning GT** — silent outside the edit window, inside it only the source to replace |
| `ref_images/{id}.png` | Reference image, where the track uses one |
| `annotations/{id}.json` | Instruction, edit window, anchor frame, instance geometry, measured evidence |
| `raw/{id}/` | **Every original ingredient**: source clip, isolated target/residual audio, mask, reference image, upstream meta, plus `origin.json` recording each file's source path and the time window taken |

`raw/` is what makes the bench adjustable: windows, anchor frames and reasoning GT can all be
re-derived without going back to the upstream corpora.

### Media ships as tar archives

`archives/` holds one tar per media folder (`raw/` split across four shards), because ~14k loose
LFS objects trip the Hub's request rate limit. Unpack next to `metadata.csv` and every path in
the CSV resolves as written:

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

`annotations/`, `metadata.csv` and `review.html` are stored loose. `review.html` is an 80-window
inspection page covering all 26 sub-tracks; open it after unpacking, served over HTTP.

## Sub-tracks

**Newly sampled (1,200).** A source video appears in exactly one sample.

| Category | Sub-track | N | What is hard |
|---|---|---|---|
| Where-only 260 | `W1_same_class_pair` | 110 | Two same-class instances of comparable size — only position disambiguates |
| | `W2_spatial_reference` | 80 | Referred to by left/right/larger/smaller |
| | `W3_small_or_occluded` | 70 | Target covers < 2% of the frame |
| Which-only 310 | `H1_overlapping_voices` | 130 | 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 310 | `T1_word_phrase_swap` | 210 | Replace 1–4 words inside a sentence |
| | `T2_sentence_rewrite` | 100 | Rewrite part of a sentence |
| Combined 250 | `C1_overlap_local_speech` | 100 | Right voice *and* 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 70 | `R1_person_identity` | 50 | Swap identity to a reference person |
| | `R3_ref_guided_add` | 20 | Add the reference subject to an empty scene |

**Legacy tracks (1,113).** Hand-curated sets carried over intact, with reasoning GT derived
where the ingredients exist.

| Sub-track | N | Origin |
|---|---|---|
| `L1_region_human_add` | 200 | Empty scene + reference person |
| `L2_region_human_swap` | 203 | Single-person region regeneration |
| `L3_region_human_multi` | 102 | **Two-person dialogue** — the only genuine visible-two-speaker material available |
| `L4_region_animal_object` | 173 | Animal/object appearance reference (ships nobg reference images) |
| `L5_speech_edit_wordlevel` | 100 | Word-level speech edit, **with word timestamps** |
| `L6_singing` | 50 | Singing |
| `L7_reason_prod` | 154 | prod instruction editing |
| `L8_reason_scratch` | 77 | Leakage-free instruction editing |
| `L9_reason_headshot` | 41 | Headshot reference identity swap |
| `L10_reason_xa` / `L11_reason_xb` | 7 / 6 | Source-vs-reference cross diagnostics |

The JAVEdit-derived `instruction` track is deliberately excluded.

## How the annotations were derived

Selection is grounded in measurements rather than caption guesswork:

* **Instance count and geometry** come from connected components of the SAM2 binary mask, so
  "the dog on the left" is checkable and the visual GT keeps only the named instance.
* **Competing sources** come from the source/target/residual decomposition shipped with the
  upstream corpora: level ratio and temporal overlap are measured directly.
* **Voice vs non-voice** is decided by Whisper's `<|nospeech|>` posterior. DSP heuristics
  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, not 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 GT is a null
anchor.

## Known gaps, stated plainly

* **285 samples have `visual_reasoning_kind = unavailable`.** The `reason*` legacy tracks ship no
  mask, so no single-frame region can be recovered. Their acoustic GT is present (windowed by the
  stored `edit_frac`). Nothing was fabricated to fill the gap.
* **`L1_region_human_add` and `L6_singing` carry a silent acoustic GT.** These are
  instance-addition cases: the source contains nothing to replace, so the acoustic reasoning GT
  is a null anchor, matching `C4`/`R3`.
* **224 legacy samples share media with another legacy sample** (singing over human_single,
  headshot/xa/xb over reason_scratch). That is the original bench design — same clip, different
  prompt. The newly sampled half has zero source overlap with the legacy half.
* Instructions for the SAM-derived tracks are templated from measured evidence, not human-written.
* Clip length follows 121/105/97/89/81/73/65 frames at 25 fps (legacy clips may go down to 49);
  1,524 of 2,313 are the full 121 frames.
* Source material comes from internal talking-head, film/drama and sounding-object corpora.
  Redistribution beyond research use may be restricted by those sources' terms.

## Leakage control

`train_exclude.list` (prod meta paths) and `train_exclude_videoids.txt` (2,087 source ids, the
stricter of the two) list everything the bench consumes. Remove them before training:

```bash
O=/path/to/ReasonAVEdit-instruction-data
comm -23 <(sort $O/reason_train_full_v2.list) <(sort train_exclude.list) > $O/reason_train_v3.list
```

	---
	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 (Landscape)
	size_categories:
	- 1K<n<10K
	configs:
	- config_name: default
	data_files:
	- split: test
	path: metadata.csv
	---

	# ReasonAVEdit-Bench — Landscape Edition

	2,313 samples, every one of them landscape, for instruction-guided joint audio-video
	editing. Two halves:

	\| Half \| N \| What it is \|
	\|---\|---\|---\|
	\| Newly sampled \| 1,200 \| 15 reasoning sub-tracks, selected from measured evidence \|
	\| Legacy tracks \| 1,113 \| Every landscape sample from the hand-curated AniAVEditBench tracks \|

	The benchmark measures not just the final edit but the multimodal reasoning behind it:
	which region to change, which sound source to change, and when the change happens. No
	edited target video is shipped — the input is a real clip plus an instruction, and the edited
	result is what a method under test produces.

	## Per-sample contents

	\| Path \| Description \|
	\|---\|---\|
	\| `videos/{id}.mp4` \| Source clip. 25 fps, `frames % 8 == 1`, sides multiple of 32, H.264 + 16 kHz AAC \|
	\| `audios/{id}.wav` \| Source audio, 16 kHz mono \|
	\| `masks/{id}.mp4` \| Full-timeline binary mask of the target region \|
	\| `reason_visual/{id}.png` \| Single-frame visual reasoning GT — edit-start frame, target region flattened to gray 127 \|
	\| `reason_visual_mask/{id}.png` \| Binary region at the anchor frame, for Visual IoU \|
	\| `reason_audio/{id}.wav` \| Full-timeline acoustic reasoning GT — silent outside the edit window, inside it only the source to replace \|
	\| `ref_images/{id}.png` \| Reference image, where the track uses one \|
	\| `annotations/{id}.json` \| Instruction, edit window, anchor frame, instance geometry, measured evidence \|
	\| `raw/{id}/` \| Every original ingredient: source clip, isolated target/residual audio, mask, reference image, upstream meta, plus `origin.json` recording each file's source path and the time window taken \|

	`raw/` is what makes the bench adjustable: windows, anchor frames and reasoning GT can all be
	re-derived without going back to the upstream corpora.

	### Media ships as tar archives

	`archives/` holds one tar per media folder (`raw/` split across four shards), because ~14k loose
	LFS objects trip the Hub's request rate limit. Unpack next to `metadata.csv` and every path in
	the CSV resolves as written:

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

	`annotations/`, `metadata.csv` and `review.html` are stored loose. `review.html` is an 80-window
	inspection page covering all 26 sub-tracks; open it after unpacking, served over HTTP.

	## Sub-tracks

	Newly sampled (1,200). A source video appears in exactly one sample.

	\| Category \| Sub-track \| N \| What is hard \|
	\|---\|---\|---\|---\|
	\| Where-only 260 \| `W1_same_class_pair` \| 110 \| Two same-class instances of comparable size — only position disambiguates \|
	\| \| `W2_spatial_reference` \| 80 \| Referred to by left/right/larger/smaller \|
	\| \| `W3_small_or_occluded` \| 70 \| Target covers < 2% of the frame \|
	\| Which-only 310 \| `H1_overlapping_voices` \| 130 \| 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 310 \| `T1_word_phrase_swap` \| 210 \| Replace 1–4 words inside a sentence \|
	\| \| `T2_sentence_rewrite` \| 100 \| Rewrite part of a sentence \|
	\| Combined 250 \| `C1_overlap_local_speech` \| 100 \| Right voice and 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 70 \| `R1_person_identity` \| 50 \| Swap identity to a reference person \|
	\| \| `R3_ref_guided_add` \| 20 \| Add the reference subject to an empty scene \|

	Legacy tracks (1,113). Hand-curated sets carried over intact, with reasoning GT derived
	where the ingredients exist.

	\| Sub-track \| N \| Origin \|
	\|---\|---\|---\|
	\| `L1_region_human_add` \| 200 \| Empty scene + reference person \|
	\| `L2_region_human_swap` \| 203 \| Single-person region regeneration \|
	\| `L3_region_human_multi` \| 102 \| Two-person dialogue — the only genuine visible-two-speaker material available \|
	\| `L4_region_animal_object` \| 173 \| Animal/object appearance reference (ships nobg reference images) \|
	\| `L5_speech_edit_wordlevel` \| 100 \| Word-level speech edit, with word timestamps \|
	\| `L6_singing` \| 50 \| Singing \|
	\| `L7_reason_prod` \| 154 \| prod instruction editing \|
	\| `L8_reason_scratch` \| 77 \| Leakage-free instruction editing \|
	\| `L9_reason_headshot` \| 41 \| Headshot reference identity swap \|
	\| `L10_reason_xa` / `L11_reason_xb` \| 7 / 6 \| Source-vs-reference cross diagnostics \|

	The JAVEdit-derived `instruction` track is deliberately excluded.

	## How the annotations were derived

	Selection is grounded in measurements rather than caption guesswork:

	* Instance count and geometry come from connected components of the SAM2 binary mask, so
	"the dog on the left" is checkable and the visual GT keeps only the named instance.
	* Competing sources come from the source/target/residual decomposition shipped with the
	upstream corpora: level ratio and temporal overlap are measured directly.
	* Voice vs non-voice is decided by Whisper's `<\|nospeech\|>` posterior. DSP heuristics
	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, not 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 GT is a null
	anchor.

	## Known gaps, stated plainly

	* 285 samples have `visual_reasoning_kind = unavailable`. The `reason*` legacy tracks ship no
	mask, so no single-frame region can be recovered. Their acoustic GT is present (windowed by the
	stored `edit_frac`). Nothing was fabricated to fill the gap.
	* `L1_region_human_add` and `L6_singing` carry a silent acoustic GT. These are
	instance-addition cases: the source contains nothing to replace, so the acoustic reasoning GT
	is a null anchor, matching `C4`/`R3`.
	* 224 legacy samples share media with another legacy sample (singing over human_single,
	headshot/xa/xb over reason_scratch). That is the original bench design — same clip, different
	prompt. The newly sampled half has zero source overlap with the legacy half.
	* Instructions for the SAM-derived tracks are templated from measured evidence, not human-written.
	* Clip length follows 121/105/97/89/81/73/65 frames at 25 fps (legacy clips may go down to 49);
	1,524 of 2,313 are the full 121 frames.
	* Source material comes from internal talking-head, film/drama and sounding-object corpora.
	Redistribution beyond research use may be restricted by those sources' terms.

	## Leakage control

	`train_exclude.list` (prod meta paths) and `train_exclude_videoids.txt` (2,087 source ids, the
	stricter of the two) list everything the bench consumes. Remove them before training:

	```bash
	O=/path/to/ReasonAVEdit-instruction-data
	comm -23 <(sort $O/reason_train_full_v2.list) <(sort train_exclude.list) > $O/reason_train_v3.list
	```