Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
32
32
ode
stringlengths
51
51
prompt_source
stringclasses
2 values
audio
stringclasses
0 values
dabe180051842052332b8cd9b1a210c0
ode/da/dabe180051842052332b8cd9b1a210c0.safetensors
generic
null
0bea11e610684696aae114a23125aa58
ode/0b/0bea11e610684696aae114a23125aa58.safetensors
generic
null
a16a37baf9ab048efe46d161a641b674
ode/a1/a16a37baf9ab048efe46d161a641b674.safetensors
generic
null
196420b0567ec7c59a801a4432720e61
ode/19/196420b0567ec7c59a801a4432720e61.safetensors
generic
null
7eb693a700f14f8b3542bac6407a4b56
ode/7e/7eb693a700f14f8b3542bac6407a4b56.safetensors
caption
null
df4ef1c00563ee2da0563bcc65661203
ode/df/df4ef1c00563ee2da0563bcc65661203.safetensors
generic
null
75aa9d565100fd22495070c8d7af6f74
ode/75/75aa9d565100fd22495070c8d7af6f74.safetensors
generic
null
df5610572f18363a9272177f2b72c0c2
ode/df/df5610572f18363a9272177f2b72c0c2.safetensors
generic
null
e73278d299b94e2c26c162616fa463a4
ode/e7/e73278d299b94e2c26c162616fa463a4.safetensors
generic
null
22edefef9292184d55cd55e66cc7d10f
ode/22/22edefef9292184d55cd55e66cc7d10f.safetensors
generic
null
ab6a5d3cd5b71dd7e149946cc890d68b
ode/ab/ab6a5d3cd5b71dd7e149946cc890d68b.safetensors
generic
null
473a74a1dd65d1c893205c6b7aa00faf
ode/47/473a74a1dd65d1c893205c6b7aa00faf.safetensors
generic
null
40a4fb7b8aadfaf2ae1b5b6ef7ed748b
ode/40/40a4fb7b8aadfaf2ae1b5b6ef7ed748b.safetensors
generic
null
2d607779c771be56e0f8a8f95d49a99e
ode/2d/2d607779c771be56e0f8a8f95d49a99e.safetensors
generic
null
d223ca533bf3667e389e3cfbe102952c
ode/d2/d223ca533bf3667e389e3cfbe102952c.safetensors
generic
null
2a28ea85ac69c4d9f26793c793aa5225
ode/2a/2a28ea85ac69c4d9f26793c793aa5225.safetensors
generic
null
e55a0c490cc538070a30c1b36e9e14a6
ode/e5/e55a0c490cc538070a30c1b36e9e14a6.safetensors
generic
null
4d5e72da951dd26967d692722be6c3f9
ode/4d/4d5e72da951dd26967d692722be6c3f9.safetensors
generic
null
9006e8369ded1282dd2eb32932af8d20
ode/90/9006e8369ded1282dd2eb32932af8d20.safetensors
caption
null
da62519ddce38cb073d3ed9565ef802e
ode/da/da62519ddce38cb073d3ed9565ef802e.safetensors
caption
null
7ae2d9cf0a689926b3085eef72d84fbf
ode/7a/7ae2d9cf0a689926b3085eef72d84fbf.safetensors
generic
null
5dcd669c9908f27cc7a3c5cba6db5f1d
ode/5d/5dcd669c9908f27cc7a3c5cba6db5f1d.safetensors
generic
null
bceb2d28753166355cba783cfbaa5e57
ode/bc/bceb2d28753166355cba783cfbaa5e57.safetensors
generic
null
8f6615b6e69485408561e6c0b631543c
ode/8f/8f6615b6e69485408561e6c0b631543c.safetensors
generic
null
f6f468cc875f4d3a5a1b5a0a288490e6
ode/f6/f6f468cc875f4d3a5a1b5a0a288490e6.safetensors
generic
null
24ae56a19ab32b69c4519d7caa10162d
ode/24/24ae56a19ab32b69c4519d7caa10162d.safetensors
generic
null
b769c43c1fe72304fbde7d9433d9206b
ode/b7/b769c43c1fe72304fbde7d9433d9206b.safetensors
generic
null
a98f20016d5fd018c1bb507556eeb26a
ode/a9/a98f20016d5fd018c1bb507556eeb26a.safetensors
generic
null
600f477ae79fe8a092efd5dd3e1262f5
ode/60/600f477ae79fe8a092efd5dd3e1262f5.safetensors
generic
null
b788ae4c5ed52f7844ec118d36a5dcd1
ode/b7/b788ae4c5ed52f7844ec118d36a5dcd1.safetensors
generic
null
5813d4191c3a99335893952f578b5988
ode/58/5813d4191c3a99335893952f578b5988.safetensors
generic
null
1ca1b3c989e89d73f4be81e7b6230481
ode/1c/1ca1b3c989e89d73f4be81e7b6230481.safetensors
generic
null
0a2a9dd963f3fcf5ad3a07db247a28f9
ode/0a/0a2a9dd963f3fcf5ad3a07db247a28f9.safetensors
caption
null
1ad763df94a65f7fd7995efc0c80c51e
ode/1a/1ad763df94a65f7fd7995efc0c80c51e.safetensors
generic
null
b96f52dcc497ff0ac36de59e51572a82
ode/b9/b96f52dcc497ff0ac36de59e51572a82.safetensors
generic
null
959a8aa13190049d36134069a5e225a0
ode/95/959a8aa13190049d36134069a5e225a0.safetensors
generic
null
dfdc3cf42b844d8d5b9dcd9db3f75ea7
ode/df/dfdc3cf42b844d8d5b9dcd9db3f75ea7.safetensors
generic
null
d9093cef310742f72afc0a4feb7e525e
ode/d9/d9093cef310742f72afc0a4feb7e525e.safetensors
caption
null
0a4d4c8310d225d826f54005c4b6633a
ode/0a/0a4d4c8310d225d826f54005c4b6633a.safetensors
generic
null
7da6c3cd88068b295f70af04345d1f67
ode/7d/7da6c3cd88068b295f70af04345d1f67.safetensors
generic
null
0dc77e7ea2ea504ee0a5dbfb14e260dc
ode/0d/0dc77e7ea2ea504ee0a5dbfb14e260dc.safetensors
caption
null
432eaff97adf68e09ed758ac6eb01dde
ode/43/432eaff97adf68e09ed758ac6eb01dde.safetensors
generic
null
8e677b25ac4761423d2071e0662fcd86
ode/8e/8e677b25ac4761423d2071e0662fcd86.safetensors
generic
null
38df425d3dd8edf8ce8f8cf495f89618
ode/38/38df425d3dd8edf8ce8f8cf495f89618.safetensors
generic
null
8eaaab9a3d04eab56070dcfe7ceb7744
ode/8e/8eaaab9a3d04eab56070dcfe7ceb7744.safetensors
generic
null
ecdaad2e62a4a281c7e54229d1253ea7
ode/ec/ecdaad2e62a4a281c7e54229d1253ea7.safetensors
generic
null
2bc7527edfc76b1f9c4162489be049f3
ode/2b/2bc7527edfc76b1f9c4162489be049f3.safetensors
generic
null
51dc60135c7be892a567d7df4c5ba668
ode/51/51dc60135c7be892a567d7df4c5ba668.safetensors
generic
null
463402e315ad637b3064949b56f86ad8
ode/46/463402e315ad637b3064949b56f86ad8.safetensors
generic
null
ecbb2d1b43cedf7a7b8fc62054198368
ode/ec/ecbb2d1b43cedf7a7b8fc62054198368.safetensors
generic
null
709cc111cf681669ee423dc7b3f66d74
ode/70/709cc111cf681669ee423dc7b3f66d74.safetensors
generic
null
59c7214b3c4dfcb28c93992f18d6d45e
ode/59/59c7214b3c4dfcb28c93992f18d6d45e.safetensors
caption
null
6dc1684fe3b15b8fae49b40e97276b0d
ode/6d/6dc1684fe3b15b8fae49b40e97276b0d.safetensors
generic
null
0d48d929d2c8df07c8180820ee282155
ode/0d/0d48d929d2c8df07c8180820ee282155.safetensors
generic
null
fe7c157abb5e33a55e10e16422565ddf
ode/fe/fe7c157abb5e33a55e10e16422565ddf.safetensors
caption
null
1c13e70fd77f86f6dfd23705d2224d91
ode/1c/1c13e70fd77f86f6dfd23705d2224d91.safetensors
generic
null
c1deff547ca883e3d47954634dd82a94
ode/c1/c1deff547ca883e3d47954634dd82a94.safetensors
generic
null
fc5302db595307542fa12a8a591f6053
ode/fc/fc5302db595307542fa12a8a591f6053.safetensors
generic
null
1ec3af653d15e311ff4519c596653c2c
ode/1e/1ec3af653d15e311ff4519c596653c2c.safetensors
generic
null
9cceb16400f11986415120fc0e12247f
ode/9c/9cceb16400f11986415120fc0e12247f.safetensors
caption
null
b546d444f84fa2042a728e7fad1a6f54
ode/b5/b546d444f84fa2042a728e7fad1a6f54.safetensors
generic
null
c78dafb4a558c9fca6d5c64889d0e72b
ode/c7/c78dafb4a558c9fca6d5c64889d0e72b.safetensors
generic
null
3f00f023ebbe0b1dc829af702b16ebac
ode/3f/3f00f023ebbe0b1dc829af702b16ebac.safetensors
generic
null
557ffdd87959af8c7cb3a588143aa81a
ode/55/557ffdd87959af8c7cb3a588143aa81a.safetensors
generic
null
fd2206a2748b06d916bedbd1764f7c56
ode/fd/fd2206a2748b06d916bedbd1764f7c56.safetensors
generic
null
d26a098b7bd91fdc6cf3527407f71179
ode/d2/d26a098b7bd91fdc6cf3527407f71179.safetensors
generic
null
2d9c24bce4db9ad257f334cb2a5b9806
ode/2d/2d9c24bce4db9ad257f334cb2a5b9806.safetensors
caption
null
fd8a4874025720668377d875055ae265
ode/fd/fd8a4874025720668377d875055ae265.safetensors
generic
null
4f12419e5dfab28d2c00d36814ca72e8
ode/4f/4f12419e5dfab28d2c00d36814ca72e8.safetensors
generic
null
a60655ad126a8140159b9ad1b61028a3
ode/a6/a60655ad126a8140159b9ad1b61028a3.safetensors
generic
null
82974aa95cdad098d223a1a7856401fa
ode/82/82974aa95cdad098d223a1a7856401fa.safetensors
generic
null
26969d8d7917db420a9b9d4ccd24840b
ode/26/26969d8d7917db420a9b9d4ccd24840b.safetensors
generic
null
4301b0e083ad11e44dfd5e63ede04192
ode/43/4301b0e083ad11e44dfd5e63ede04192.safetensors
caption
null
af30ad86e261e478cf02fd08e176743b
ode/af/af30ad86e261e478cf02fd08e176743b.safetensors
generic
null
681666b20a565815ffd8ff36d177a2a3
ode/68/681666b20a565815ffd8ff36d177a2a3.safetensors
generic
null
a0bafabf97d1a66ae8f14aae0b7f7b56
ode/a0/a0bafabf97d1a66ae8f14aae0b7f7b56.safetensors
generic
null
9bd6ac33c08a089a515d61043b1df3f4
ode/9b/9bd6ac33c08a089a515d61043b1df3f4.safetensors
generic
null
36236c247afdf4a0bfff1739f4afbe5a
ode/36/36236c247afdf4a0bfff1739f4afbe5a.safetensors
generic
null
97893208766f681a73719322dd3aa9bd
ode/97/97893208766f681a73719322dd3aa9bd.safetensors
caption
null
8134fa5bd0d70a4cb1ea29723d26e400
ode/81/8134fa5bd0d70a4cb1ea29723d26e400.safetensors
generic
null
fc08751603a9b5b2198125e795f64a29
ode/fc/fc08751603a9b5b2198125e795f64a29.safetensors
generic
null
27cc6eddaec93892e29fc9657a89ceb3
ode/27/27cc6eddaec93892e29fc9657a89ceb3.safetensors
generic
null
ca0542046b6fac67a297b3d26d78c749
ode/ca/ca0542046b6fac67a297b3d26d78c749.safetensors
generic
null
7d91ee4d737807262bc61164398abc46
ode/7d/7d91ee4d737807262bc61164398abc46.safetensors
generic
null
924cb290a509f90e199fe7d1557afaf1
ode/92/924cb290a509f90e199fe7d1557afaf1.safetensors
generic
null
1a461da9672204e31f3378b9074b3e46
ode/1a/1a461da9672204e31f3378b9074b3e46.safetensors
generic
null
14b5b32745549057932b6d6ac4fafe05
ode/14/14b5b32745549057932b6d6ac4fafe05.safetensors
caption
null
95783dde0775490435ff4ea202dd30cf
ode/95/95783dde0775490435ff4ea202dd30cf.safetensors
generic
null
ef4720ac1b0d229580bc7e637b2985ad
ode/ef/ef4720ac1b0d229580bc7e637b2985ad.safetensors
generic
null
ff750d65b3021ad833a2f06cd506e71e
ode/ff/ff750d65b3021ad833a2f06cd506e71e.safetensors
generic
null
f235bd3699c5cd8bad627d8fbef3b557
ode/f2/f235bd3699c5cd8bad627d8fbef3b557.safetensors
caption
null
17efc592ff3a16dada062ba868e75514
ode/17/17efc592ff3a16dada062ba868e75514.safetensors
generic
null
eee7889d30a88d470fb6493cc33c6572
ode/ee/eee7889d30a88d470fb6493cc33c6572.safetensors
generic
null
3df140dbc7c2df2e03193b7f517b2f01
ode/3d/3df140dbc7c2df2e03193b7f517b2f01.safetensors
caption
null
bd2791ba777c3c9330be88a6b801e8d2
ode/bd/bd2791ba777c3c9330be88a6b801e8d2.safetensors
generic
null
32826ff0e371830510abe1a3a4291434
ode/32/32826ff0e371830510abe1a3a4291434.safetensors
generic
null
517c6333017bab9cc798a772670449e0
ode/51/517c6333017bab9cc798a772670449e0.safetensors
generic
null
32cd4feaeed50440a47731484bdc47ad
ode/32/32cd4feaeed50440a47731484bdc47ad.safetensors
generic
null
caf53c209553436acc1c277537623e14
ode/ca/caf53c209553436acc1c277537623e14.safetensors
generic
null
e80fb4c329fe4c25c3226a4b8ded7dbe
ode/e8/e80fb4c329fe4c25c3226a4b8ded7dbe.safetensors
generic
null
End of preview. Expand in Data Studio

AURA ODE Distillation Dataset (1.3B, Mimi)

These are teacher ODE trajectories used for stage 1 (ODE distillation) of AURA-1.3B. They were generated by AURA-1.3B-Teacher-TalkVerse, a model built on Wan2.1-T2V-1.3B and conditioned on Mimi audio. Each sample is a speech clip together with the latents the teacher visited while denoising. The teacher was also conditioned on a reference image and a text prompt from Hallo3. Those are not redistributed here, but you can rebuild them from your own Hallo3 download (see Reference image and prompt).

Project page · Paper · Code

Partial release, more coming soon. For now, this repository only contains the ODE latents of the AURA-1.3B teacher (initial noise, intermediate states and final clean latent, as Wan2.1 VAE latents), not the videos themselves; video_lat decodes to the generated video with the Wan2.1 VAE. The rest of the dataset described in the paper, the 30K TalkVerse (AURA-5B) clips with their decoded videos, audio tracks and ODE latents, is coming soon.

Splits

Split Samples Size
validation 200 ~6.3 GB
train 30,000 ~980 GB

Layout

data/<split>/metadata.jsonl                one row per sample
data/<split>/ode/<xx>/<id>.safetensors     teacher ODE trajectory
data/<split>/audio/<xx>/<id>.wav           driving speech (validation only, see below)

<xx> is the first two characters of the sample id. Each row of metadata.jsonl looks like:

{"id": "038054f9…", "ode": "ode/03/038054f9….safetensors", "prompt_source": "caption", "audio": "audio/03/038054f9….wav"}

All paths are relative to the split folder. prompt_source says which text prompt the teacher used: "caption" for the Hallo3 caption of the clip, "generic" for the fixed prompt "a person is talking" (61% of train, 55 of the 200 validation samples).

ODE trajectory file

Each .safetensors file holds five Wan2.1 VAE latents of shape (16, 20, 64, 96) (channels, latent frames, height, width). That corresponds to 80 video frames.

Key dtype Content
noises bf16 Initial Gaussian noise
inter745, inter498, inter255 fp32 Intermediate latents on the teacher's sampling trajectory, at timesteps 745 / 498 / 255
video_lat fp32 Final clean latent produced by the teacher
import json
from huggingface_hub import snapshot_download
from safetensors.torch import load_file

root = snapshot_download("kyutai/AURA-Dataset", repo_type="dataset", allow_patterns="data/validation/*")
split = f"{root}/data/validation"
rows = [json.loads(l) for l in open(f"{split}/metadata.jsonl")]
traj = load_file(f"{split}/{rows[0]['ode']}")   # dict: noises, inter745, inter498, inter255, video_lat

Source data

  • Audio: in validation, synthetic speech generated by the authors that is unrelated to the reference video. It is included. In train, the audio is the Hallo3 video's own soundtrack, so it is not included (metadata.jsonl has no audio field there).
  • Reference images and prompts come from the Hallo3 training data (CC BY-NC-ND 4.0). They are not redistributed here.

Reference image and prompt (from Hallo3)

Every sample id is the name of a Hallo3 video. For each sample, the teacher was conditioned on:

  • reference image: frame 0 of videos/<id>.mp4 (720×1080, used as-is);
  • prompt: the text in caption/<id>.txt if prompt_source is "caption", otherwise "a person is talking";
  • audio (train only): the first 4 s (at most) of the audio track of videos/<id>.mp4.

All 30,200 ids (train + validation) are present in Hallo3. To rebuild them, download videos.tgz (44 GB) and caption.tgz (10 MB) from the Hallo3 repo, then run:

# pip install huggingface_hub av pillow
import json, tarfile
from pathlib import Path
import av
from huggingface_hub import hf_hub_download

hallo3 = Path("hallo3")
for f in ["videos.tgz", "caption.tgz"]:
    tarfile.open(hf_hub_download("fudan-generative-ai/hallo3_training_data", f, repo_type="dataset")).extractall(hallo3)

def reference(row):
    """(first frame as a PIL image, prompt) used to condition the teacher for this sample."""
    sample_id = row["id"]
    with av.open(str(hallo3 / "videos" / f"{sample_id}.mp4")) as c:
        frame = next(c.decode(video=0)).to_image()
    if row["prompt_source"] == "generic":
        prompt = "a person is talking"
    else:
        prompt = (hallo3 / "caption" / f"{sample_id}.txt").read_text().strip()
    return frame, prompt

rows = [json.loads(l) for l in open(f"{split}/metadata.jsonl")]   # `split` as in the snippet above
image, prompt = reference(rows[0])

Trajectories were generated with 50 ODE steps and a guidance scale of 4.5. No sample overlaps the HALLO3 test set used in the paper.

We checked that every "caption" prompt matches the Hallo3 caption file exactly, and that frame 0 decoded this way matches the images we used, up to JPEG re-compression (mean absolute pixel difference < 1/255).

Responsible use

The ODE latents decode to videos depicting real people. Use it only for research on audio-driven video generation and distillation. Do not use it to identify, impersonate or create content about the people depicted.

Contact

For questions, contact Nathaniel Cohen (nathaniel.cohen@kyutai.org) or Amélie Royer (amelie@kyutai.org).

Citation

@inproceedings{cohen2026aura,
  title     = {{AURA}: {AU}dio-d{R}iven streaming {A}vatar},
  author    = {Cohen, Nathaniel and Dufour, Nicolas and Royer, Am{\'e}lie
               and Newson, Alasdair and P{\'e}rez, Patrick},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2026},
}
Downloads last month
-

Models trained or fine-tuned on kyutai/AURA-Dataset

Collection including kyutai/AURA-Dataset

Paper for kyutai/AURA-Dataset