Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

LTX-Video 2.5 Preprocessed Dataset

Preprocessed training data for LTX-Video 2.5 (joint audio + video), stored as PyTorch tensors. Every file is a torch.save'd dict and can be loaded with:

import torch
d = torch.load("latents/10s/clip000001.pt", map_location="cpu", weights_only=True)

Structure

Clips are split by duration bucket (5s, 10s) and modality. File names are shared across folders: clipNNNNNN.pt refers to the same source clip in every folder that contains it.

Folder Clips (5s / 10s) Size Contents
latents/ 10,842 / 43,860 ~332 GB Video latents
audio_latents/ 10,842 / 43,860 ~6.1 GB Audio latents
conditions/ 10,847 / 43,860 ~643 GB Prompt conditioning (text embeds)

Total: ~979 GB, 164,111 files.

File formats

latents/{5s,10s}/clipNNNNNN.pt — video latents

Key Type / shape Notes
latents bfloat16, (128, T, 22, 40) VAE latent; T = 16 (5s) or 32 (10s)
num_frames int Latent temporal frames (T)
height, width int Latent spatial dims: 22 x 40
fps float 25.0

audio_latents/{5s,10s}/clipNNNNNN.pt — audio latents

Key Type / shape Notes
latents float32, (8, T, 16) Audio VAE latent; T = 122 (5s) or 250 (10s)
num_time_steps int Latent time steps (T)
frequency_bins int 16
duration float Seconds (~4.84 for 5s clips, ~9.96 for 10s clips)

conditions/{5s,10s}/clipNNNNNN.pt — prompt conditioning

Key Type / shape Notes
video_prompt_embeds bfloat16, (1024, 4096) Video text-encoder embeds
audio_prompt_embeds bfloat16, (1024, 2048) Audio text-encoder embeds
prompt_attention_mask int64, (1024,) Attention mask for the prompt tokens
Downloads last month
42