# SHIFT: Steering Hidden Intermediates in Flow Transformers Source: [SHIFT: Steering Hidden Intermediates in Flow Transformers]() Collection index: [llms.txt]() Data updated: 2026-04-14T05:00:04.234Z Authors: Nina Konovalova, Andrey Kuznetsov, Aibek Alanov Published: 2026-04-10T11:07:33.000Z ## Abstract Diffusion models have become leading approaches for high-fidelity image generation. Recent DiT-based diffusion models, in particular, achieve strong prompt adherence while producing high-quality samples. We propose SHIFT, a simple but effective and lightweight framework for concept removal in DiT diffusion models via targeted manipulation of intermediate activations at inference time, inspired by activation steering in large language models. SHIFT learns steering vectors that are dynamically applied to selected layers and timesteps to suppress unwanted visual concepts while preserving the prompt's remaining content and overall image quality. Beyond suppression, the same mechanism can shift generations into a desired \\emph{style domain} or bias samples toward adding or changing target objects. We demonstrate that SHIFT provides effective and flexible control over DiT generation across diverse prompts and targets without time-consuming retraining. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Concept Preservation on Mickey prompts](): Concept Preservation; Mickey prompts; Concept Score (CS): 26.06. - [Concept Preservation on Dog prompts](): Concept Preservation; Dog prompts; CS: 24.18. - [Concept Preservation on Legislator prompts](): Concept Preservation; Legislator prompts; CS: 21.77. - [Concrete Object Erasure on Pikachu Concrete Object Erasure](): Concrete Object Erasure; Pikachu Concrete Object Erasure; CS: 25.14. - [Style Erasure on Old artists](): Style Erasure; Old artists; AccE: 30. - [Concept Preservation on Spongebob prompts](): Concept Preservation; Spongebob prompts; Concept Score (CS): 27.35. - [Concept Preservation on Pikachu prompts](): Concept Preservation; Pikachu prompts; Concept Score (CS): 26.25. - [Concrete Object Erasure on Snoopy Concrete Object Erasure](): Concrete Object Erasure; Snoopy Concrete Object Erasure; CS: 11.98. - [Mickey Concrete Object Erasure](): Concrete Object Erasure; Mickey Concrete Object Erasure; CS: 22.7. - [Concrete Object Erasure on Spongebob Concrete Object Erasure](): Concrete Object Erasure; Spongebob Concrete Object Erasure; CS Score: 26.13. - [Concrete Object Erasure on Dog Concrete Object Erasure](): Concrete Object Erasure; Dog Concrete Object Erasure; CS: 22.98. - [Concrete Object Erasure on Legislator Concrete Object Erasure](): Concrete Object Erasure; Legislator Concrete Object Erasure; CS Score: 17.9. - [Object Erasure on Snoopy prompts](): Object Erasure; Snoopy prompts; CS Score: 18.57. - [Style Erasure on Modern artists](): Style Erasure; Modern artists; AccE: 60. - [Image Generation on MS-COCO 5K (sampled images)](): Image Generation Evaluation; MS-COCO 5K (sampled images); FID: 33.9. - [Nudity Erasure on I2P (dev)](): Nudity Erasure; I2P (dev); Common Count: 155. ## Original sources and implementation [Original paper]() [PDF]()