datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Latent-Earth
Latent Earth: An Atlas of Architecture in Flux.2
200,000 images of 40,000 places on Earth, each rendered by a single
image model in a single state of its training, with five internal
representations recorded for every image while it was being generated.
Nothing else enters. Each prompt contains only a place's name; no
photographs, no maps, no climate records correct what the model proposes.
This is therefore not a depiction of the world but a probe of the model: a
survey of what… See the full description on the dataset page: https://huggingface.co/datasets/Punktiert/Latent-Earth.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.danbooru2024-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.Latent-Resonance-AI-Image-Forensics-Benchmark-N100
Latent Resonance: SOTA Empirical AI Image Forensics Benchmark (N=100 & N=1,000 Scale)
Author: Debdip Bandyopadhyay (Independent AI Researcher, Kolkata, India; M.Tech, IIT Jodhpur, AI & Data Science)Preprint & Paper: Latent Resonance: Zero-Shot Autoencoder Inversion and Azimuthal Spectral Forensics for Diffusion Image Attribution (IEEE Flagship / CERN Zenodo 2026)
Benchmark Overview
This repository provides:
The official verified $N=100$ ground-truth image… See the full description on the dataset page: https://huggingface.co/datasets/DebdipCS/Latent-Resonance-AI-Image-Forensics-Benchmark-N100.Latent-Resonance-AI-Image-Forensics-Benchmark-N1000
Latent Resonance: SOTA Large-Scale AI Image Forensics Benchmark (N=1,000)
Author: Debdip Bandyopadhyay (Independent AI Researcher, Kolkata, India; M.Tech, IIT Jodhpur, AI & Data Science)Preprint & Paper: Latent Resonance: Zero-Shot Autoencoder Inversion and Azimuthal Spectral Forensics for Diffusion Image Attribution (IEEE Flagship / CERN Zenodo 2026)
1. Executive Summary & Diagnostic Suite
This repository contains the complete empirical evaluation records… See the full description on the dataset page: https://huggingface.co/datasets/DebdipCS/Latent-Resonance-AI-Image-Forensics-Benchmark-N1000.imagenet-256-sd-vae-ft-mse-latents
ImageNet-256 SD-VAE-ft-MSE Latents
Pre-computed posterior means (no variance/std) from the Stable Diffusion VAE (stabilityai/sd-vae-ft-mse) for the full ImageNet-1K training set at 256×256 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
Type
Description
latent_mean… See the full description on the dataset page: https://huggingface.co/datasets/yuanchenyang/imagenet-256-sd-vae-ft-mse-latents.latent-CIFAR100
Latent CIFAR100
This is the CIFAR100 dataset that has been latently encoded with various VAEs and saved as safetensors.
Paths are structured:
train/class/number.safetensors
test/class/number.safetensors
The files are in class named folders, each safetensor file contains its image data in the "latent" key, and the original class number in the "class" key (which I don't recommend using.)
I recommend using the sdxl-488 version! The 488 refers to the size of the latents (channels… See the full description on the dataset page: https://huggingface.co/datasets/Verah/latent-CIFAR100.e621_2024-latents-sdxl-1ktar
E621 2024 SDXL VAE latents in 1k tar
Dedicated dataset to align both NebulaeWis/e621-2024-webp-4Mpixel and deepghs/e621_newest-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… See the full description on the dataset page: https://huggingface.co/datasets/6DammK9/e621_2024-latents-sdxl-1ktar.Civ-40M
CivitAI 40M — Image Metadata
Companion dataset to the peer-reviewed article:
Laura Wagner and Eva Cetinić (2026). "Perpetuating misogyny with generative AI: How model personalization normalizes gendered harm." Big Data & Society 13(3). Open access (CC BY-NC 4.0).
DOI: 10.1177/20539517261467361
Dataset DOI: 10.57967/hf/9053 · Cite as Wagner and Cetinic (2025a).
This repository contains the image metadata dataset analyzed in that paper: 40,630,560 records describing images… See the full description on the dataset page: https://huggingface.co/datasets/latentcanon/Civ-40M.