Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions

ICLR 2026

Ada Görgün* 1, Fawaz Sammani* 2, Nikos Deligiannis 2, Bernt Schiele 1, Jonas Fischer 1
1 Max Planck Institute for Informatics, Saarland Informatics Campus, Germany
2 ETRO Department, Vrije Universiteit Brussel, Belgium
* Equal Contribution
Paper Code arXiv

Prompt-conditioned intervention (PCI) over diffusion timesteps. We suggest to study when noise turns into a specific concept through the lens of concept insertion success, i.e., the chance of inserting a concept at a certain timestep successfully. By switching the base-prompt (top) at different time points of the diffusion process with a prompt composed of the base prompt and the concept of interest (old), we can measure this success rate (CIS curve) across different seeds and base prompts to analyze temporal dependency and influence of concepts in diffusion models.

Abstract

Diffusion models are usually evaluated by their final outputs, gradually denoising random noise into meaningful images. Yet, generation unfolds along a trajectory, and analyzing this dynamic process is crucial for understanding how controllable, reliable, and predictable these models are in terms of their success/failure modes. In this work, we ask the question: when does noise turn into a specific concept (e.g., age) and lock in the denoising trajectory? We propose PCI (Prompt-Conditioned Intervention) to study this question. PCI is a training-free and model-agnostic framework for analyzing concept dynamics through diffusion time. The central idea is the analysis of Concept Insertion Success (CIS), defined as the probability that a concept inserted at a given timestep is preserved and reflected in the final image, offering a way to characterize the temporal dynamics of concept formation. Applied to several state-of-the-art text-to-image diffusion models and a broad taxonomy of concepts, PCI reveals diverse temporal behaviors across diffusion models, in which certain phases of the trajectory are more favorable to specific concepts even within the same concept type. These findings also provide actionable insights for text-driven image editing, highlighting when interventions are most effective without requiring access to model internals or training, and yielding quantitatively stronger edits that achieve a balance of semantic accuracy and content preservation than strong baselines.

Prompt-Conditioned Intervention (PCI)

Prompt-Conditioned Intervention (PCI) tracks how and when visual concepts solidify during the diffusion process. Starting from the same noisy seed, we denoise with a base prompt until a chosen timestep, then switch to a concept prompt that introduces the target concept. The model finishes generation under this new prompt, and a VQA model checks whether the concept appears in the final image. By repeating this across seeds and timesteps, PCI measures the probability that a concept can still be inserted, revealing the point in the trajectory when concepts become stable and resistant to change.

Category-Level Analysis

Intervention samples across timesteps

CIS reveals cross-category and cross-architecture differences. ● denotes CIS at τ50 and ● denotes CIS at τ70 across multiple concept categories and diffusion models.

Fine-Grained Analysis

Revealing context-dependent differences for the same concept. We show how CIS curves vary when the same concept is placed in different contexts within the base prompt, demonstrating that contextual framing can significantly shift concept editability over time.

Text-Driven Image Editing

Intervention samples across timesteps

Examples of text-driven image editing on SDXL. The edited images are shown at four different points with their respective CIS probabilities: τ30, τ50, τ70, and τ90. High probabilities until a certain point ensure the intended modification but reduce preservation of the original image. We observe that CIS probabilities above 0.7 start to noticeably compromise the original content, and probabilities between 0.5 to 0.7 as suggested by our analysis (red rectangle) are best for editing while preserving the original image.

BibTeX

@misc{gorgun2025temporalconceptdynamicsdiffusion,
      title={Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions}, 
      author={Ada Gorgun and Fawaz Sammani and Nikos Deligiannis and Bernt Schiele and Jonas Fischer},
      year={2025},
      eprint={2512.08486},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.08486}, 
}