Stable Diffusion data creator cleans up its act
LAION just re-released its flagship training dataset as Re-LAION-5B, claiming it has thoroughly scrubbed suspected child sexual abuse material. Following recommendations from watchdogs and a Stanford investigation, the group removed 2,236 flagged links and urged researchers to ditch the older LAION-5B copies.
Re-LAION-5B is a cleaned re-release of LAION-5B, the massive text-image dataset originally used to train Stable Diffusion and several other frontier vision models.
If you train diffusion models, multimodal representations, or large-scale vision backbones, this release aims to get rid of severe legal and ethical hazards without sacrificing scale.
What got scrubbed
Keep in mind that LAION datasets do not host raw images. They are indexes containing image URLs and alt text scraped from the open web via Common Crawl.
Back in December 2023, the Stanford Internet Observatory revealed that LAION-5B contained at least 1,679 links to illegal child sexual abuse material, prompting LAION to pull the entire dataset offline.
To clean it up, LAION collaborated with the Internet Watch Foundation, Human Rights Watch, the Canadian Center for Child Protection, and Stanford researchers.
Using URL lists and image perceptual hashes from these groups, LAION removed a total of 2,236 flagged links associated with suspected CSAM, fully covering the 1,008 flagged URLs identified in the Stanford report.
Specs and how to get it
Re-LAION-5B contains around 5.5 to 5.8 billion text-image pairs and is published under the permissive Apache 2.0 license.
It comes in two flavors. The standard Re-LAION-5B Research variant removes all identified CSAM links while preserving general web data.
The second variant, Re-LAION-5B Research-Safe, layers on additional filters to strip out adult NSFW imagery, making it safer for general visual research.
If you already host local copies of the original dataset, you do not need to redownload dozens of terabytes. LAION provided the metadata diffs so teams can easily sanitize their existing archives by dropping matched records.
The metadata files are now live and downloadable directly through LAION's official project pages.
Why it matters
For academic labs and open-source practitioners, Re-LAION-5B offers a legally defensible foundation for large-scale multimodal training. Anyone currently hosting or training on the legacy LAION-5B should pull the new metadata and purge their pipelines immediately.
That said, cleaning the index does not erase what past models have already learned. Weights trained on the original dump remain out in the wild, which recently prompted Runway to remove its Stable Diffusion 1.5 weights from Hugging Face.
It is also worth noting that LAION explicitly labels this dataset for research rather than commercial deployment. If you are building enterprise image generation products, you still need properly licensed proprietary datasets rather than relying on scraped web dumps.
Sources (1)
- techcrunch.com(opens in a new tab)Originaltechcrunch.com