monet

Dataset Card for MONET MONET** (**M**assive, **O**pen, **N**on-redundant and **E**nriched **T**ext-to-image dataset) is a large-scale, curated image-text dataset designed for training text-to-image (T2I) systems. It contains **103.8 million** high-quality image-text pairs distilled from **2.9 billion** raw pairs across nine heterogeneous open sources (6 *real* and 3 *synthetic*) through successive stages of safety filtering, domain-based filtering, exact and near-duplicate removal, and re-captioning with multiple vision-language models, and is further augmented with synthetically generated samples. Each image is released with pre-computed embeddings, structured annotations and pre-encoded VAE latents to accelerate downstream use. A 4B-parameter latent diffusion model trained *exclusively* on MONET reaches competitive GenEval and DPG scores, demonstrating that MONET lowers the barrier to large-scale, reproducible text-to-image research. Table of Contents Dataset Summary Dataset Sources Curation Pipeline Data Fields Usage Splits Supported Tasks Demos - Retrieval & UMAP Building subsets with FAISS indexes Training Biases, Risks, and Limitations Ethical and Responsible Use Maintenance & Contact Changelog Citation 📋 Dataset Summary **Name:** **MONET**: **M**assive, **O**pen, **N**on-redundant and **E**nriched **T**ext-to-image dataset **Primary modality:** image-text pairs **Scale:** 103.8M **Language:** English (`en`) **License:** Apache-2.0 **Intended use:** text-to-image pretraining, image-retrieval, zero-shot image classification **Available captions:** original alt-text plus up to four synthetic captions from Florence-2, ShareGPT4V-7B, InternVL3-8B and Gemini-2.5-flash-lite, spanning short concept-level to long fine-grained descriptions **Available embeddings:** CLIP ViT-B/32 (512-d), DINOv2 ViT-G/14 (1536-d), SSCD disc_mixup (512-d) **Available metadata:** source/provenance, licensing, aesthetic scores (LAION-improved and an internal classifier), perceptual hash (p