Pristine Models Separates Unsafe Knowledge Into Detachable Branches During Text-to-Image Pretraining

uw-madison · hf · 2026-10-15

Researchers from UW-Madison and Adobe Research propose Pristine Models, a text-to-image architecture addressing the safety-quality tradeoff in training data.

• Instead of post-hoc concept unlearning (which suffers catastrophic forgetting as concepts accumulate), the model uses an attention mechanism to compartmentalize safe and unsafe knowledge into independently parameterized branches during pretraining.
• The unsafe branch can simply be detached at inference time.
• On ImageNet, it enables safer generation while achieving higher image quality and better text alignment than safe-data-only training or post-hoc concept removal.
• The approach scales to CC12M, remaining effective in a larger, more diverse open-world pretraining setting.

Original post →

More from Safety

• AI Doesn't Care If It's Wrong — You'll Face the Court Consequences — adariostrange · 2026-10-15
• Hackers hit 700 French supermarkets, spoiling millions in food — researcher suspects AI-assisted attack — AaronBergman18 · 2026-10-15
• User finds ChatGPT data-sharing silently re-enabled after opting out — AlexKontorovich · 2026-10-15
• Dev's honest scoreboard: pattern rules catch under 25% of prompt injections, guard model catches most — colinmcnamara · 2026-10-15
• OpenAI, Anthropic, Meta, Google all disclosed models accessing systems without authorization since July — petrusenko_max · 2026-10-15
• Neuroscientist Kording: Suing AI Companies Is the Most Useful "Alignment" Action — KordingLab · 2026-10-15

Safety channel →