Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Vysotskyi, Mykola, Kohut, Zahar, Shpir, Mariia, Rumezhak, Taras, Karpiv, Volodymyr
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866914330700677120
author Vysotskyi, Mykola
Kohut, Zahar
Shpir, Mariia
Rumezhak, Taras
Karpiv, Volodymyr
author_facet Vysotskyi, Mykola
Kohut, Zahar
Shpir, Mariia
Rumezhak, Taras
Karpiv, Volodymyr
contents Machine unlearning in text-to-image diffusion models aims to remove targeted concepts while preserving overall utility. Prior diffusion unlearning methods typically rely on supervised weight edits or global penalties; reinforcement-learning (RL) approaches, while flexible, often optimize sparse end-of-trajectory rewards, yielding high-variance updates and weak credit assignment. We present a general RL framework for diffusion unlearning that treats denoising as a sequential decision process and introduces a timestep-aware critic with noisy-step rewards. Concretely, we train a CLIP-based reward predictor on noisy latents and use its per-step signal to compute advantage estimates for policy-gradient updates of the reverse diffusion kernel. Our algorithm is simple to implement, supports off-policy reuse, and plugs into standard text-to-image backbones. Across multiple concepts, the method achieves better or comparable forgetting to strong baselines while maintaining image quality and benign prompt fidelity; ablations show that (i) per-step critics and (ii) noisy-conditioned rewards are key to stability and effectiveness. We release code and evaluation scripts to facilitate reproducibility and future research on RL-based diffusion unlearning.
format Preprint
id arxiv_https___arxiv_org_abs_2601_03213
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion
Vysotskyi, Mykola
Kohut, Zahar
Shpir, Mariia
Rumezhak, Taras
Karpiv, Volodymyr
Machine Learning
Machine unlearning in text-to-image diffusion models aims to remove targeted concepts while preserving overall utility. Prior diffusion unlearning methods typically rely on supervised weight edits or global penalties; reinforcement-learning (RL) approaches, while flexible, often optimize sparse end-of-trajectory rewards, yielding high-variance updates and weak credit assignment. We present a general RL framework for diffusion unlearning that treats denoising as a sequential decision process and introduces a timestep-aware critic with noisy-step rewards. Concretely, we train a CLIP-based reward predictor on noisy latents and use its per-step signal to compute advantage estimates for policy-gradient updates of the reverse diffusion kernel. Our algorithm is simple to implement, supports off-policy reuse, and plugs into standard text-to-image backbones. Across multiple concepts, the method achieves better or comparable forgetting to strong baselines while maintaining image quality and benign prompt fidelity; ablations show that (i) per-step critics and (ii) noisy-conditioned rewards are key to stability and effectiveness. We release code and evaluation scripts to facilitate reproducibility and future research on RL-based diffusion unlearning.
title Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion
topic Machine Learning
url https://arxiv.org/abs/2601.03213