ParamsDrag: Interactive Parameter Space Exploration via Image-Space Dragging

TL;DR

ParamsDrag tackles the challenge of efficiently exploring numerical simulation parameter spaces by enabling direct interaction with predicted visualizations. It builds a parameter-to-visualization generator with a StyleGAN2-inspired architecture and introduces a structure-based patch system, feature supervision, and a gradient-descent parameter inversion to edit outputs in a discrete latent space. The approach demonstrates competitive image-quality metrics against state-of-the-art surrogates and enables intuitive, stepwise control of simulation parameters through image-space dragging. The framework has practical implications for accelerating parameter tuning in real-world simulations, while acknowledging domain-dependent patch design and discretized latent-space editing as areas for future refinement. Overall, ParamsDrag provides a novel, interactive pathway for parameter-space exploration with measurable gains in efficiency and user-centric control over visualization outputs.

Abstract

Numerical simulation serves as a cornerstone in scientific modeling, yet the process of fine-tuning simulation parameters poses significant challenges. Conventionally, parameter adjustment relies on extensive numerical simulations, data analysis, and expert insights, resulting in substantial computational costs and low efficiency. The emergence of deep learning in recent years has provided promising avenues for more efficient exploration of parameter spaces. However, existing approaches often lack intuitive methods for precise parameter adjustment and optimization. To tackle these challenges, we introduce ParamsDrag, a model that facilitates parameter space exploration through direct interaction with visualizations. Inspired by DragGAN, our ParamsDrag model operates in three steps. First, the generative component of ParamsDrag generates visualizations based on the input simulation parameters. Second, by directly dragging structure-related features in the visualizations, users can intuitively understand the controlling effect of different parameters. Third, with the understanding from the earlier step, users can steer ParamsDrag to produce dynamic visual outcomes. Through experiments conducted on real-world simulations and comparisons with state-of-the-art deep learning-based approaches, we demonstrate the efficacy of our solution.

Figures (10)

• Figure 1: The high-level pipeline of ParamsDrag.
• Figure 2: The model structure of the generator, with inputs being the simulation parameters and visualization parameters, and the output being the visualization image.
• Figure 3: The workflow of interacting with visualization images. Users can directly obtain target parameters and visualization images by selecting feature structures and editing it.
• Figure 4: Continuous changes to the latent vectors of ParamsDrag generator can lead to the collapse of the generated images.
• Figure 5: Schematic representation of the distribution of latent vectors for the same amount of training and test data under two different models.