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

2026-07-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Uni-AdaVD is a novel universal inference-time concept erasure framework designed for visual generative models. It addresses the critical issue of undesirable concepts absorbed from uncurated pretraining data, a challenge existing architecture-specific methods often fail to resolve while preserving non-target content and generative priors. Uni-AdaVD operates by treating the value space of multimodal attention as a unified intervention space. It employs encoder-aware target representation construction to localize target semantics across diverse text encoders and integrates orthogonal value decomposition with an adaptive erasing shift to suppress target semantic directions without altering original model weights. Extensive experiments confirm its strong performance on single- and multi-concept erasure across U-Net-, DiT-, and autoregressive image generators, as well as text-to-video models, establishing it as an efficient and adaptable safety mechanism.

Key takeaway

For Machine Learning Engineers deploying visual generative models, Uni-AdaVD provides a crucial safety mechanism. If you are struggling with architecture-specific concept erasure or preserving non-target content, consider integrating Uni-AdaVD. It allows you to universally remove undesirable concepts from models like U-Net, DiT, and text-to-video generators at inference time, without costly model retraining. This approach ensures safer model deployment while maintaining generative quality.

Key insights

Uni-AdaVD offers a universal, inference-time method to erase undesirable concepts from visual generative models without retraining.

Principles

• Concept erasure can be universal and architecture-agnostic.
• Intervention in multimodal attention's value space is effective.
• Orthogonal decomposition suppresses semantics without weight updates.

Method

Uni-AdaVD constructs encoder-aware target representations, then applies orthogonal value decomposition with an adaptive erasing shift in the multimodal attention's value space to suppress target semantics.

In practice

• Erase specific concepts from generated images.
• Remove undesirable content from text-to-video models.
• Apply safety mechanisms post-training without fine-tuning.

Topics

• Concept Erasure
• Visual Generative Models
• Multimodal Attention
• Orthogonal Value Decomposition
• Model Safety
• Inference-time Adaptation

Code references

• QifanZhou/Uni-AdaVD

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• TILDE: TILt-based Distributional Erasure for Concept Unlearning — Concept unlearning in diffusion models is best achieved by explicitly defining the desired post-unlearning distribution, not just suppressing concepts.
• Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows — UVR restricts unsafe information flow in DiT multimodal attention during an early "start-up stage" to prevent harmful image generation.
• Association Restoration Test: Revealing Restorable Shortcuts after Unlearning — Output robustness and feature readability do not guarantee functional unlearning; associations can remain restorable.
• U-CAN: Utility-Aware Contrastive Attenuation for Efficient Unlearning in Generative Recommendation — U-CAN precisely unlearns sensitive data in GenRec LLMs by selectively attenuating LoRA adapter parameters based on activation contrasts and utility.
• UNIVID: Unified Vision-Language Model for Video Moderation — UNIVID unifies video moderation with policy-aware captions, improving interpretability and efficiency over fragmented black-box systems.
• TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation — Generative recommendation unlearning requires token reassignment to resolve conflicts from shared semantic IDs, preserving utility.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.