No Safe Dose: How Training Data Drives Unsafe Image Generation

Felix Friedrich, Lukas Helff, Niharika Hegde, Patrick Schramowski, Kristian Kersting, [4pt] Black Forest Labs · 2026

Read No Safe Dose: How Training Data Drives Unsafe Image Generation by Felix Friedrich, Lukas Helff, Niharika Hegde, Patrick Schramowski, Kristian Kersting, [4pt] Black Forest Labs (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-28137.txt.

Description

Friedrich and colleagues isolate how unsafe training images drive unsafe image generation. Training identical models on datasets differing only in unsafe share, they find output unsafety climbs steadily with contamination, and the share, not the count, matters. Even at zero contamination a floor remains, traced partly to the frozen text encoder, while filtering costs no image quality.

Catalog record

CREATORFelix Friedrich, Lukas Helff, Niharika Hegde, Patrick Schramowski, Kristian Kersting, [4pt] Black Forest Labs
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55463
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.28137

Browse the full library · About the Tower of Babel