SafeRedir achieves robust unlearning of unsafe concepts in image generation models by adaptively redirecting prompt embeddings toward safe semantic regions at inference time via a multi-modal classifier and token delta generator.
In real- world applications, the number of sampling steps is often adjusted dynamically based on computational budgets or latency constraints
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
SafeRedir: Prompt Embedding Redirection for Robust Unlearning in Image Generation Models
SafeRedir achieves robust unlearning of unsafe concepts in image generation models by adaptively redirecting prompt embeddings toward safe semantic regions at inference time via a multi-modal classifier and token delta generator.