Ref10K

Ref10K is a set of about 10K reference-to-audio-video (Ref2VA) inputs for MiniMax-H3: reference images, videos and audio, each paired with a rewritten instruction and a six-section H3 prompt. It was built to distill a few-step student (Anyflow2) from the H3 Ref2VA teacher without paired target videos. Every record is an input only; there is no ground-truth output.

Split	Records	File
train	9,987	train.jsonl
validation	491	val.jsonl (split by source group and near-duplicate clusters, no overlap with train)
eval70	70	eval/val_eval70.jsonl (10 per task, drawn from validation)
focus12	12	eval/focus12.jsonl (hand-written: outfit / background / person swap, character drive)

All referenced media are included under media/ and raw/, at the relative paths given in each record.

Task mix (train)

7,000 records follow the task shares of GOKU-2M; 3,000 cover the other Ref2VA abilities.

Task	Records	Subtasks
video_edit	5,215	multi_task 945, add 880, swap_alter 795, subject_movement 710, camera 685, style_transfer 640, remove 560
image_video_mix	1,785	reference_add 905, reference_swap 880
single_image_ref	1,200	identity, subject, object, product, scene, style
multi_image_ref	787	two_subjects, three_way, subject_scene, subject_style, multi_view
video_ref	300	motion / camera / rhythm / style transfer
video_continuation	300	continue_from_end
audio_av_mix	400	music / voice-timbre / sfx reference, soundtrack reuse

• References (train): 7,856 video, 4,921 image and 295 audio clips; 1–3 references per record (at most 3 images, 2 videos, 4 references).
• Clips: every video or audio reference is trimmed to at most 107/24 s (≈4.46 s) via start_sec / end_sec.
• Target spec: each record's output field gives 480×864, 107 frames at 24 fps (≈4.46 s), 32 kHz stereo audio.
• Audio policy: generate 9,287, reference 490, reuse 210. 569 records contain spoken dialogue in 8 languages.
• Instruction language: English 5,994, Chinese 3,993.
• Source footage: all real (filmed or photographed). Synthetic/rendered sources were filtered out with a VLM check.

Record format

{
  "id": "train_image_video_mix_reference_swap_00001",
  "task": "image_video_mix", "subtask": "reference_swap",
  "instruction": "Rewritten user instruction (en/zh)",
  "seed_instruction": "Original instruction from the source dataset, when there is one",
  "reference_assets": [
    {"asset_id": "video_1", "type": "video", "path": "media/video/...mp4", "role": "edit_source",
     "start_sec": 0.0, "end_sec": 4.36, "use_audio": false, "sha256": "...", "media_id": "..."},
    {"asset_id": "image_1", "type": "image", "path": "raw/...jpg", "role": "replacement", "sha256": "...", "media_id": "..."}
  ],
  "reference_bindings": {"<Video 1>": "video_1", "<Picture 1>": "image_1"},
  "h3_prompt_sections": {"subject_definitions": "...", "summary": "...", "retention_analysis": "...",
                         "detailed_description": "...", "overall_soundscape": "...", "non_diegetic_music": "..."},
  "h3_prompt": "the six sections serialized in the official MiniMax-H3 format",
  "output": {"height": 480, "width": 864, "num_frames": 107, "fps": 24, "audio_sampling_rate": 32000, "audio_channels": 2},
  "audio_policy": "generate | reference | reuse | reuse_and_reference",
  "dialogue_language": null,
  "provenance": {"dataset": "...", "license": "...", "source_url": "...", "media_ids": ["..."]},
  "source_descriptions": {"video_1": {"summary": "...", "subjects": [...], "scene": "..."}},
  "quality": {"media_checked": true, "prompt_checked": true, "judge_checks": {...}}
}

• Labels: <Picture i>, <Video k> and <Audio j> are numbered per modality in the order of reference_assets, as in the native MiniMax-H3 Ref2VA pipeline. A video whose use_audio is true also receives an <Audio j> label, and that label comes first.
• Prompts: every label and every <Subject N> in the prompt resolves to a reference.
• prompts.jsonl: holds just the prompts for every split: seed_instruction, the rewritten instruction, h3_prompt and h3_prompt_sections.

How it was built

1. Media: sources were downloaded and decoded, trimmed to ≤4.4 s with audio and video kept in sync, then de-duplicated by sha256 and perceptual hash.
2. Description: each image, video and audio clip was described with Qwen3-Omni-30B-A3B-Instruct, which also checked whether it looked synthetic and flagged content issues.
3. Task assembly: records were split into train and validation first, by source group and near-duplicate cluster, then assembled to the quotas above. Dialogue, instruction language and audio policy were assigned in this step.
4. Prompt rewriting: prompts were rewritten with Qwen3-235B-A22B-Instruct-2507-FP8, which was given the official H3 prompt-writing guides (base and ref). Records were then checked by a schema validator covering labels, sections, limits and durations.
5. Consistency check: an LLM judge compared each prompt against the media descriptions, and rejected records were rewritten once with the judge's feedback.
   • 12,300 records were generated and 11,366 passed.
   • The final splits were drawn from the records that passed.

Sources and licenses

Media come from the datasets below. Each record's provenance.license and provenance.source_url give the per-item license and origin. The collection as a whole is released for non-commercial research under CC BY-NC 4.0, because several sources are NC. Individual items remain under their original terms, which take precedence.

Source	Files	License of the original items
GOKU-2M (source videos and reference crops)	8,189	CC BY-NC 4.0 (footage from Koala-36M)
Unsplash Lite	713	Unsplash Dataset terms
Amazon Berkeley Objects	557	CC BY 4.0
Open Images V7	395	CC BY 2.0 (per image)
VGGSound	228	YouTube-derived clips, copyright of the original uploaders
Wikimedia Commons	193	per file (CC BY / CC BY-SA / public domain)
WikiArt	193	per work, some still in copyright
Free Music Archive	175	per track Creative Commons (incl. NC / ND)
FSD50K	69	per clip (CC0 / CC BY / CC BY-NC)
Pexels	52	Pexels License
FLEURS	49	CC BY 4.0
LibriTTS-R	15	CC BY 4.0

If you are a rights holder and want an item removed, please open a discussion on this repository with the file path or media_id, and it will be taken down.

Intended use

Ref10K is meant for research on reference-conditioned audio-video generation and distillation, and for evaluating it. It contains no target videos; the outputs come from a teacher model at training time. Some videos and photos show real people.