license: mit
dataset_info:
splits:
- name: train
num_examples: 382241
task_categories:
- text-to-image
- image-classification
tags:
- midjourney
- synthetic-images
- captioning
- embeddings
- consent-safe
- humanless
- previews
language:
- en
size_categories:
- 100K<n<1M
Dataset Card
Introduction
In recent studies, there has been a significant concern regarding data privacy. While one can scrape data online or create one’s own datasets, there is a major problem concerning the privacy of any individuals depicted in a dataset, their consent to be published, as well as dataset copyright.
Because of this, the present dataset aims to focus on fully human-free and copyright-free material based on the existing published datasets created by Photoroom and the follow-up LLaVA-annotated corpus. The parent corpora are Photoroom/midjourney-v6-recap, published on the Hugging Face Hub in 2024 under the MIT license, and brivangl/midjourney-v6-llava, on which it is based on.
Methodology
We used the original Parquet files captions and text prompts to filter out human-related captions and topics, producing 404,598 caption-clean survivors. To perform more in-deep filtering in ambiguous cases, when caption doesn't contain human tag (for instance profession only), we used a combination of embedding-based metrics and models, including YOLOv8, SigLIP, DINOv2, WD14 tags, applying different techniques, in particular zero-shot filtering in combination.
In particular, YOLO detector boxes identified approximately 21,000 potential candidates for filtering.
To increase the confidence of filtering, we then utilized a combination of a Booru-style tagger and two zero-shot constructions, combining different embedding spaces to achieve high confidence. The best single detector achieved an AUC of approximately 0.91–0.95 across runs, which was sufficient to support the filtering of human-associated objects and retain non-human depictions. Every machine verdict was also subject to human verification.
Figure 1 — One-shot ROC (person-positive vs. clean), re-run on the post-filtering residual distribution.
Figure 2 — Categories found by unfiltered stock detectors per proxy set (confidence ≥ 0.25).
The resulting dataset consists of 382,241 image records, with its structure displayed in the table below.
Dataset Structure
| Property | Value |
|---|---|
| Surviving images (rows) | 382,241 |
| Prompt groups | 146,632 (≈2.6 variants each) |
| VLM captions per image | 3 (LLaVA, Gemini, Qwen), all present |
| SigLIP embedding | so400m-384, 1,152-dimensional, float16 |
| DINOv2 embedding | base, 768-dimensional, float16 |
| UMAP projections | 2-D and 3-D, per encoder |
| Preview tier (optional) | 384 × 384 JPEG, approximately 35 KB median |
| Source image | 1024 × 1024 JPEG, approximately 152 KB median |
| Container | Apache Parquet (Zstandard compression) |
| Columns | 40 metadata columns / 43 with previews |
| On-disk size | 1.74 GB / 15.6 GB with previews |
| License | MIT |
 |
 |
Figure 3 — UMAP projection of the semantic embedding space by subject (SigLIP so400m-384, left), and by rendering style (DINOv2-base, right).
Currently, the publicly released dataset is limited to the image embeddings only. In the further update, the full subset containing the metadata and image previews will be uploaded. This limitation is a consideration of the dataset footprint, and ability to browse and/or download on the limited connections, not copyright.
Because the derived dataset preserves the existing structure and public row identifiers of the original datasets, one can browse the existing datasets, download the required dataset entries from the original parquet files, using streaming download, and filter out the records in which they are interested while consulting the corresponding copyright notices.
Caveats
Several caveats apply. The text-stage filter is English-centric, so non-English prompts reached the vision stage unfiltered, and the depiction-of-a-depiction boundary (statues, mannequins, dolls) remains an unresolved policy question. No stage proves the absolute absence of human depictions — the claim rests on layered, individually characterized filters plus human adjudication and a notice-and-takedown commitment. Finally, person_n_boxes is evidence, not a label: on this humanless set its boxes are dominated by the detector's known false-positive vocabulary, so read it alongside the confidence value.
Copyright, License, and Attribution
This dataset is based on the MIT-licensed Photoroom/midjourney-v6-recap dataset, the CortexLM/midjourney-v6 dataset, captioned with the LLaVA-1.6 model as well as brivangl/midjourney-v6-llava.
The derived embedding set is intended for release under the MIT license inherited from the parent corpus’s MIT-derived licensing chain. If you consider this dataset helpful, please consider supporting and citing the creators of the original datasets: Photoroom and brivangl.
This dataset was released as is. By accessing and using this dataset, you acknowledge and agree that Cortex Foundation and the author of this repo are not responsible for any copyright violations or legal consequences that may arise from the use of these images.