noakraicer/ID-LoRA-CelebVHQ

CelebV-HQ Preprocessed Dataset for ID-LoRA

Preprocessed training data for ID-LoRA: identity-driven audio-video personalization with In-Context LoRA (paper).

Overview

Property	Value
Source dataset	CelebV-HQ
Training pairs	2,963
Unique videos	1,959
Speakers	311
Resolution	512×512
Frame rate	25 fps
Frames per clip	121 (~4.84 s)

Quick Start

Browse and load data

load_dataset returns training pairs with playable video objects and metadata:

from datasets import load_dataset

ds = load_dataset("noakraicer/ID-LoRA-CelebVHQ")
print(ds["train"].column_names)  # ['target', 'reference', 'speaker_cluster_id', 'target_caption']
print(ds["train"][0]["target_caption"])

Note: Loading videos requires the decord package (pip install decord).

Download for training

from huggingface_hub import snapshot_download

# Full dataset (videos + precomputed latents)
snapshot_download("noakraicer/ID-LoRA-CelebVHQ", repo_type="dataset", local_dir="./data/celebvhq")

# Precomputed latents only (skip videos)
snapshot_download(
    "noakraicer/ID-LoRA-CelebVHQ",
    repo_type="dataset",
    allow_patterns=["precomputed/**", "train/metadata.jsonl"],
    local_dir="./data/celebvhq",
)

Extract precomputed latents

After downloading, extract the .tar.zst archives:

cd ./data/celebvhq/precomputed
for f in *.tar.zst; do
  tar --use-compress-program=unzstd -xf "$f"
done

This creates per-video .pt files in latents/, audio_latents/, audio_latents_clean/, and conditions/.

Dataset Structure

ID-LoRA-CelebVHQ/
├── train/
│   ├── metadata.jsonl                 # Pair metadata (2,963 rows)
│   └── {video_id}.mp4                 # 1,959 unique video clips
└── precomputed/
    ├── latents.tar.zst                # Video VAE latents
    ├── audio_latents.tar.zst          # Audio VAE latents (target)
    ├── audio_latents_clean.tar.zst    # Denoised audio VAE latents (reference)
    └── conditions.tar.zst             # Text + caption embeddings (Gemma 3)

Video filenames follow the pattern {video_id}.mp4. Latent filenames follow {video_id}.pt.

Metadata Columns

Each row in train/metadata.jsonl represents one training pair. When loaded via load_dataset, the *_file_name columns are resolved to video objects:

Column	Type	Description
target	Video	Target video clip
reference	Video	Reference video clip (same speaker, different clip)
speaker_cluster_id	string	Speaker identity cluster
target_caption	string	Structured caption with [VISUAL], [SPEECH], [SOUNDS], [TEXT] sections

Precomputed Latents

Ready-to-train representations stored as .tar.zst archives, each containing per-video .pt files.

Archive	Description
latents.tar.zst	Video VAE latents
audio_latents.tar.zst	Audio VAE latents — target audio (original recording, including environmental sounds)
audio_latents_clean.tar.zst	Audio VAE latents — reference audio (denoised speech, used for IC conditioning)
conditions.tar.zst	Text embeddings from Gemma 3, computed from the full structured caption

Audio Latent Types

Training uses two types of audio latents:

• audio_latents (target): Audio encoded from the original video. This is what the model learns to generate.
• audio_latents_clean (reference): Denoised audio with environmental sounds removed, used as in-context conditioning. Clean audio helps the model focus on speaker identity rather than background noise.

See the training code for how reference audio latents are mapped to training pairs.

Latent File Format

Each .pt file is a dict:

• Video latents: latents (bf16 [C, F, H, W]), num_frames, height, width, fps
• Audio latents: latents (fp32 [C, T, F]), num_time_steps, frequency_bins, duration
• Conditions: prompt_embeds (bf16 [seq_len, hidden]), prompt_attention_mask

Caption Format

Each caption follows a structured format with four tagged sections:

[VISUAL]: <scene description, people, actions, setting>
[SPEECH]: <word-for-word speech transcription>
[SOUNDS]: <speaker vocal style, environmental sounds>
[TEXT]: <on-screen text or "None">

Training

This dataset is designed for training ID-LoRA adapters built on LTX-2. Each pair consists of a target video and a reference video from the same speaker. The model jointly generates audio and video, transferring the speaker's voice identity from the reference while the visual appearance is controlled via first-frame conditioning and the text prompt.

See the paper and code for training details.

Citation

@misc{dahan2026idloraidentitydrivenaudiovideopersonalization,
  title     = {ID-LoRA: Identity-Driven Audio-Video Personalization
               with In-Context LoRA},
  author    = {Aviad Dahan and Moran Yanuka and Noa Kraicer and Lior Wolf and Raja Giryes},
  year      = {2026},
  eprint    = {2603.10256},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url       = {https://arxiv.org/abs/2603.10256}
}