datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
imagenet1k-256-wds-latentsThe imagenet1k dataset in the webdataset format
Each image was resized so that the max side resolution is 256, making sure to preserve aspect ratio.
Each image was encoded to latents using the sixteen channel https://huggingface.co/ostris/vae-kl-f8-d16
No cropping was used to encode to latents!
The resulting dataset has images in their original aspect ratio, but much smaller, and encodeded with a vae.
latent-sr-embeddings
Latent-SR Embeddings: Precomputed VAE Latents for Medical Image Super-Resolution
Precomputed VAE latent embeddings from the paper:
"Domain-Specific Latent Representations Improve the Fidelity of Diffusion-Based Medical Image Super-Resolution"Sebastian Cajas, Ashaba Judith, Rahul Gorijavolu, Sahil Kapadia, Hillary Clinton Kasimbazi, Leo Kinyera, Emmanuel Paul Kwesiga, Sri Sri Jaithra Varma Manthena, Luis Filipe Nakayama, Ninsiima Doreen, Leo Anthony Celi.arXiv:2604.12152 (2026)… See the full description on the dataset page: https://huggingface.co/datasets/sebasmos/latent-sr-embeddings.anycam-syncam-480-4shot-latentsanycam-mvhumannet-480-4shot-room-latentsfusionX_480p_wan21_latentsvoxcpm2-ghana-speech-ipa-latents
VoxCPM2 Ghana — Precomputed AudioVAE Latents
Precomputed VoxCPM-2 AudioVAE latents for a Ghanaian multilingual TTS fine-tune,
ready for training with the official train_voxcpm_finetune.py
(train_manifest: ghana-latents). No audio decoding or VAE encoding needed at train
time — the latent feat is fed straight to the VoxCPM-2 model with the IPA transcript.
Each language is a dataset subset:
from datasets import load_dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/voxcpm2-ghana-speech-ipa-latents.pdm3-ht-20260528-flux2-vae-latents-public
PDM-3-HT FLUX.2 VAE latents for ImageNet-256 train
Public research artifact for PDM-3-HT VAE-backend experiments. This repository contains latent cache shards only. It intentionally does not contain raw ImageNet images, ADM-cropped uint8 images, PAE latents, PAE checkpoints, or training checkpoints.
Source and preprocessing
Source dataset: ImageNet-1k train via ILSVRC/imagenet-1k; access requires accepting the upstream ImageNet terms.
Image preprocessing before VAE… See the full description on the dataset page: https://huggingface.co/datasets/LAXMAYDAY/pdm3-ht-20260528-flux2-vae-latents-public.video-dit-latents-hq
Video DiT Latents - Animals (HQ)
Pre-computed VAE latents for training video generation models.
Dataset Info
Property
Value
Resolution
256×256 pixels
Latent Shape
(4, 16, 32, 32)
Frames
16 @ 8fps (2 seconds)
VAE
stabilityai/sd-vae-ft-mse
Classes
dog, cat, bird, horse, fish, lion, elephant, monkey, butterfly, deer
Usage
import torch
from pathlib import Path
# Load a single latent
latent = torch.load("dog/12345.pt")… See the full description on the dataset page: https://huggingface.co/datasets/Jnaranjo/video-dit-latents-hq.MiraData_Wan21_Latentssa-1b_latents_768sa1b-sdxl-latents-1024MiraData_Wan22_Latentsanycam-mvhumannet-480-4shot-latentssa-1b_latents_512imagenet-latents-images-rawimagenet-latents-imagesimport os
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
from datasets import load_dataset
dataset = load_dataset("G-REPA/imagenet-latents-images", split="train")
layergen-eval-latents
LayerGen — Eval-Set Latents (VAE latents + baked text embeddings)
Pre-encoded evaluation-set inputs for the LayerGen layer-decomposition / harmonization models,
so inference can run anywhere (off-AIP) without the raw video → VAE-encode → umT5-encode pipeline.
Each *.parquet is one clip and is fully self-contained:
column group
contents
{composite,mask,fg,bg}_latent_bytes (+ _shape, _dtype)
4-stream Wan-VAE latents, 81f/21 latent-T, fp16, [16,21,60,104]… See the full description on the dataset page: https://huggingface.co/datasets/cs-mshah/layergen-eval-latents.anycam-dl3dv-480-latentsvoxcpm-ghana-latents
VoxCPM Ghana — Precomputed AudioVAE Latents
The exact training-ready data used to fine-tune
ghananlpcommunity/voxcpm-ghana:
precomputed VoxCPM-0.5B AudioVAE latents (16 kHz) for 42 Ghanaian languages +
filtered Ghanaian English, with language-tagged transcripts. Drop-in for
VoxCPM fine-tuning — no audio decoding or VAE encoding needed at train time.
1,756,157 clips · ~3,400 h · 16 kHz
42 Ghanaian languages (incl. Twi split: twi-asante, twi-akuapem) + en
AudioVAE from… See the full description on the dataset page: https://huggingface.co/datasets/ghananlpcommunity/voxcpm-ghana-latents.3DGS_latentsLatentsimagenet-latents-invae-f16d32import os
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
from datasets import load_dataset
dataset = load_dataset("G-REPA/imagenet-latents-invae-f16d32", split="train")
ego10k-vjepa-latents
Ego10k V-JEPA Latents Dataset
This dataset contains compressed, highly-informative Video Joint Embedding Predictive Architecture (V-JEPA) latents extracted from Ego-centric industrial manufacturing videos.
Dataset Structure
The dataset is partitioned into roughly 1GB .parquet chunks using PyArrow.
Data Source and Preprocessing
The latent embeddings in this dataset were systematically extracted from the Ego10k Master Dataset provided by build.ai. The… See the full description on the dataset page: https://huggingface.co/datasets/rookierufus/ego10k-vjepa-latents.molmo-ae-cc12m-latents
CC12M latents — VGT-AE with a frozen SigLIP2 (Molmo) encoder
Pre-encoded CC12M images as (32, 16, 16) float16 latents, paired with their captions.
WebDataset format: 1097 tars, each member a <key>.npy + <key>.txt pair.
samples
6,836,022
shards
1097
latent shape
(32, 16, 16) float16
input resolution
512 px
encoder
VGTAE_Siglip2, siglip2vit_frozen_stage2 (50K steps)
ViT
Molmo SigLIP2-base/patch16, frozen through both stages
decoder (for reference)… See the full description on the dataset page: https://huggingface.co/datasets/dingshizhe/molmo-ae-cc12m-latents.midashenglm-gen-training-latents
ModelsLab/midashenglm-gen-training-latents
Precomputed audio latents for fine-tuning
mispeech/midashenglm-gen,
paired with six-view prompts in the exact format the model was trained on.
This is not an audio dataset and not a caption dataset. Each record is the
output of the model's frozen DashengTokenizer encoder — 768-dimensional latents
at 25 Hz, stored float16 — next to the tagged prompt string built from the
source metadata.
Why it exists
The encoder is frozen… See the full description on the dataset page: https://huggingface.co/datasets/ModelsLab/midashenglm-gen-training-latents.vgtae-cc12m-latents
CC12M latents encoded with VGT-AE (448px)
Pre-encoded CC12M
images in the VGT-AE latent space, so text-to-image training can skip the
encoder entirely.
These are not DC-AE latents. VGT-AE is a hybrid codec — a fine-tuned
Qwen2.5-VL ViT as the encoder, a DC-AE decoder at sampling time. The tensor shape
happens to match DC-AE f32c32 (32, 16, 16), but the space is completely
different; mixing the two silently trains a model against noise.
Contents
1097 WebDataset… See the full description on the dataset page: https://huggingface.co/datasets/dingshizhe/vgtae-cc12m-latents.imagenet-256-flux2-vae-latents
ImageNet-256 FLUX.2 VAE Latents
Pre-computed deterministic, model-facing encodings from the
FLUX.2 VAE (black-forest-labs/FLUX.2-dev)
for the full ImageNet-1K training set at 256x256 resolution, stored as Parquet
shards. Each example includes latents for both the original and horizontally
flipped image, enabling flip augmentation without re-encoding at training time.
Dataset Description
Each example contains:
Column
Shape
Stored type
Description… See the full description on the dataset page: https://huggingface.co/datasets/yuanchenyang/imagenet-256-flux2-vae-latents.h3_stage_latentsdanbooru2024-latents-sdxl-1ktar
Danbooru 2024 SDXL VAE latents in 1k tar
Dedicated dataset to align deepghs/danbooru2024-webp-4Mpixel. "4MP-Focus" for average raw image resolution.
Latents are ARB with maximum size of 1024x1024 as the recommended setting in kohyas. Major reason is to make sure I can finetune with RTX 3090. VRAM usage will raise drastically after 1024.
Generated from prepare_buckets_latents_v2.py, modified from prepare_buckets_latents.py.
Used for kohya-ss/sd-scripts. In theory it may replace… See the full description on the dataset page: https://huggingface.co/datasets/6DammK9/danbooru2024-latents-sdxl-1ktar.sa-1b_latents_256