AviadDahan/SepGen-Dataset

SepGen training manifests

Text and metadata for the 4,843 training segments (and 7 validation segments) used to train SepGen. Each segment is a 4.52 s, two-source audio-video clip: a video, an audio mix, and the two source stems that sum to it. This repository holds, per segment, the captions the model was conditioned on, the per-source bounding boxes, and the provenance needed to rebuild the clip from its upstream datasets.

Paper · Project page · Code · Models

No media is included: no audio, video, latents or text embeddings. The precomputed LTX-2.5 latents and text features used for training will be released soon.

Contents

File	Rows	What
train.jsonl	4,843	one JSON object per training segment, sorted by segment
validation.jsonl	7	the segments used for validation during training (same schema; see notes)
build_manifest.py		the script that produced both files from the internal build

Composition

family	Segments	layout (segments)	Source media
celebvhq	1,658	split-screen (1,406), split-screen-augmented (252)	CelebV-HQ talking-head clips, two speakers side by side, overlapping or turn-taking speech
urmp	1,437	concert (135), concert-pair (897), concert-augmented (405)	URMP ensemble recordings and their separated instrument tracks
generated	749	generated (749)	free-placement composites, static or moving panels (see below)
vgg_composite	999	split-screen (499), off-screen (500)	VGGSound and MUSIC-21 clips
total	4,843

Layouts:

• split-screen: two panels side by side, one source per panel.
• concert: the original URMP duet video (players seated left to right), cropped to the stage band.
• concert-pair: two players of a larger URMP ensemble, each cropped out of the original frame and placed side by side; the mix is the sum of those two players' separated tracks only.
• *-augmented: a split-screen or concert segment with its two panels rescaled and moved; the audio is identical to the base segment (source.augmentation.base_segment).
• generated: two panels placed freely on a 768×512 canvas, either static or moving along a smooth path (source.motion). The 749 segments are 200 generated clip + CelebV-HQ speaker, 199 generated clip + URMP player, 150 generated clip + generated clip and 200 CelebV-HQ speaker + URMP player. The generated clips (177 distinct) were produced with LTX-2.3 22B (dev); their prompt and seed are in the manifest.
• off-screen: one source fills the frame, the other is audible only (visible: false).

All segments are 113 frames at 25 fps (4.52 s), 768×512. In every family the mix is the sum of the two stems.

Fields

Field	Description
segment	segment id
family	celebvhq, urmp, generated or vgg_composite
layout	see above
duration_s, fps, num_frames, width, height	4.52, 25, 113, 768, 512
captions	up to three caption registers: semantic, positional, short. Each holds scene (video-side scene caption), scene_audio (caption for the audio-mix), stem0 and stem1 (one caption per source). Only registers that were used in training for that segment are present.
stem_boxes	per-source boxes over the 15 latent frames: boxes[stem][latent_frame] = [x0, y0, x1, y1], normalized to [0, 1], 4 decimals; valid[stem]; method says how they were derived. Latent frame t covers video frames 8(t−1)+1…8t (frame 0 alone for t = 0).
source	per-source provenance and timing; schema depends on the family (below)
mix	how the mix was formed: always mix = stem0 + stem1; per-stem gains and levels where they were applied, the peak gain, and stem_order_swapped where the stem indices were exchanged (see below)

Caption registers

• semantic: a full scene description plus a descriptive caption per source ("A woman, speaking. She says, \"...\"", "A trumpet, playing.").
• positional: sources named by their side ("The instrument on the left, playing.").
• short: the semantic scene caption with a short noun phrase per source ("A woman.", "A trumpet.").

Availability over the training set: semantic 3,844 segments, positional 4,047, short 4,553. Every segment has at least one complete register. The VGGSound/MUSIC-21 composites have no semantic register; the off-screen ones have short only.

source by family

• CelebV-HQ (source.stems[k]): celebvhq_clip_id (CelebV-HQ clip id, <youtube_id>_<n>), youtube_id, chunk_index (index of the 121-frame, 25 fps window of that clip the speech was taken from), source_start_s/source_end_s (the speech span inside that window), in_segment_start_s/in_segment_end_s (where it was placed in this segment), transcript, side. source.speech is turn-taking or overlapping speech.

• URMP: piece (URMP folder name), video (Vid_*.mp4), piece_start_s (segment start within the piece), ensemble_parts, and per stem the instrument, URMP part number and separated-track file name track (AuSep_*.wav, as named in the URMP release), and side. concert-pair segments add player_crops_xywh_px (the two players' crops in the 1920×1080 URMP frame); concert segments add frame_crop_xywh_px.

• Generated composites: group, motion, and per stem kind (generated, celebvhq_speaker or urmp_player), side_at_start (canvas side at the first frame; panels may move, see stem_boxes), the panel description, and:

  • generated: generator, generation_prompt, seed, generation_steps, generation_resolution_wh, generation_frames, generation_fps, and window_start_s (start of the 4.52 s window inside the generated clip);
  • celebvhq_speaker / urmp_player: the CelebV-HQ or URMP provenance as above, plus origin_segment and origin_stem (the two-source segment the panel was cut from).

• VGGSound / MUSIC-21 (source.stems[k]): dataset, youtube_id, label, visible, side (split-screen only) and timing:

  • VGGSound: vggsound_start_s (the clip's start time in the VGGSound index), window_offset_s (start of the 4.52 s window inside the downloaded 10 s clip) and youtube_start_s_approx = vggsound_start_s + window_offset_s. For 135 of 1,571 VGGSound stems the downloaded file was longer than 10 s (local_download_duration_s), so the reference of the offset is unknown and youtube_start_s_approx is null.
  • MUSIC-21: music21_class, window_offset_s and youtube_start_s_approx = 20 s + offset (the window was cut from the section starting 20 s into the video).

  On 500 of the 999 composites (250 per layout) the two stem indices were exchanged after composition so that the stem index does not predict the side or the visible source; these have mix.stem_order_swapped: true. The picture is unchanged.

Validation

validation.jsonl lists the 7 segments evaluated during training. They are not in train.jsonl. Each carries only the register it was conditioned on at validation time; its scene and scene_audio are the same prompt. stem_boxes is null except for the split-screen composite.

Obtaining the source media

The manifests identify every source; the media itself must be obtained from the upstream datasets under their own licenses and terms of use.

Dataset	Official page	Used for
CelebV-HQ	https://celebv-hq.github.io/	talking-head clips (celebvhq, speaker panels in generated)
URMP	https://labsites.rochester.edu/air/projects/URMP.html	ensemble videos and separated instrument tracks (urmp, player panels in generated)
VGGSound	https://www.robots.ox.ac.uk/~vgg/data/vggsound/	in-the-wild sound sources (vgg_composite)
MUSIC-21	https://github.com/roudimit/MUSIC_dataset	solo instrument performances (vgg_composite)
LTX-2.3	https://github.com/Lightricks/LTX-2	generated clips (generated), reproducible from prompt and seed

CelebV-HQ, VGGSound and MUSIC-21 index YouTube videos; some may no longer be available.

Personal information

CelebV-HQ shows real people. Some of the machine-written captions used in training named or guessed the identity of the person shown, or the film or show a clip is from. In these manifests every such name and title has been replaced by a neutral description (e.g. "a middle-aged man"), in 734 of the 4,850 rows; these captions therefore differ slightly from the exact text the model was trained on. One spoken line that quotes a name reads "[name]". Transcripts are otherwise the words spoken in the clip.

Citation

@article{dahan2026sepgen,
  title={SepGen: Multi-Stem Audio-Video Separation and Generation in a Single Model},
  author={Dahan, Aviad and Khatib, Rajaei and Bitton, Yonatan and Szpektor, Idan and Wolf, Lior and Giryes, Raja},
  journal={arXiv preprint arXiv:2610.11361},
  year={2026}
}

Please also cite the upstream datasets you use:

• CelebV-HQ: Zhu et al., CelebV-HQ: A Large-Scale Video Facial Attributes Dataset, ECCV 2022.
• URMP: Li et al., Creating a Multitrack Classical Music Performance Dataset for Multimodal Music Analysis: Challenges, Insights, and Applications, IEEE Transactions on Multimedia, 2019.
• VGGSound: Chen et al., VGGSound: A Large-Scale Audio-Visual Dataset, ICASSP 2020.
• MUSIC-21: Zhao et al., The Sound of Motions, ICCV 2019 (MUSIC: Zhao et al., The Sound of Pixels, ECCV 2018).

Code: https://github.com/AviadDahan/SepGen