Uni-AdaVD: Universal Concept Erasure for Visual Generation via Orthogonal Value Decomposition

Uni-AdaVD (Universal Adaptive Value Decomposer) provides a unified, architecture-agnostic approach to concept erasure in modern visual generative models. As generative models are trained on large-scale, uncurated internet data, they often inadvertently learn to reproduce copyrighted, unsafe, or otherwise undesirable content. Traditional erasure methods are typically tied to specific architectures (e.g., U-Net-based diffusion) and often require costly retraining or fine-tuning. Uni-AdaVD addresses this by performing intervention directly during inference, allowing for the removal of concepts from diverse models—including U-Net, DiT (Diffusion Transformer), and autoregressive generators—without modifying the underlying model weights.

The framework operates on the principle that the value space of multimodal attention layers serves as a common interface for injecting textual semantics across different generative architectures. Uni-AdaVD employs three primary mechanisms:

1. Encoder-Aware Target Representation Construction (ETRC): Recognizing that different text encoders (e.g., CLIP vs. T5) distribute semantic information differently, ETRC uses encoder-specific probing. It aggregates scattered semantic features to accurately localize target concepts, whether they are explicit (e.g., a specific character) or implicit (e.g., unsafe content).
2. Orthogonal Value Decomposition (OVD): This technique identifies the semantic directions associated with the target concept within the attention value space and suppresses them by projecting the value vectors into an orthogonal subspace.
3. Layer-wise Adaptive Erasing Shift (LAES): To maintain high-quality generation, LAES dynamically modulates the strength of the erasure across different network layers. This ensures that the target concept is effectively removed while preserving the model's generative priors and non-target visual content.

Uni-AdaVD offers a flexible, "plug-and-play" safety mechanism for the rapidly evolving landscape of visual generation. By avoiding the need for retraining, it significantly lowers the barrier to deploying safer models in production environments. Its ability to handle both single and multi-concept erasure across heterogeneous architectures—from image generators to text-to-video models—makes it a robust tool for developers seeking to align generative AI with safety and compliance standards.