// publications

Ehsan Pajouheshgar*, Hyunsoo Kim*, Sabine Süsstrunk, Wenzel Jakob, Jinah Park

We introduce Neural Particle Automata, a Lagrangian generalization of Neural Cellular Automata where each cell is a particle with a continuous position and internal state, using differentiable SPH operators and memory-efficient CUDA kernels to scale learned self-organizing particle dynamics.

Ehsan Pajouheshgar, Yitao Xu, Ali Abbasi, Alexander Mordvintsev, Wenzel Jakob, Sabine Süsstrunk

We pair a coarse-grid Neural Cellular Automata model with a lightweight implicit decoder that maps cell states and local coordinates to appearance, enabling self-organizing NCA outputs to render at arbitrary resolution in real time while preserving regeneration, robustness, and spontaneous dynamics across 2D, 3D, and mesh domains.

Ehsan Pajouheshgar, A. Bhardwaj, N. Selub, E. Lake

We study binary 2D cellular automata that robustly store one-bit memories under noise, and use a modified Neural Cellular Automata framework to discover many new memory-preserving rules beyond Toom's rule.

M. Larsson, H. Yamaguchi, Ehsan Pajouheshgar, I-C. Shen, K. Tojo, C-M. Chang, et al.

We present Mokume, a solid-wood dataset of 190 cube samples with high-resolution photos, ring annotations, and CT scans, and introduce a three-stage inverse pipeline that reconstructs globally consistent 3D growth fields and volumetric color textures from exterior photographs using procedural and NCA-based synthesis.

D. Wang, Ehsan Pajouheshgar, Y. Xu, T. Zhang, S. Süsstrunk

We present VNCA, an efficient volumetric style-transfer model that synthesizes multi-view-consistent stylization features for target smoke in real time while preserving temporally coherent transitions across frames.

Y. Xu, T. Zhang, Ehsan Pajouheshgar, S. Süsstrunk

We identify Canonical Latent Representations in conditional diffusion models and use them in CaDistill to transfer core class semantics through compact representative latent samples, enabling strong robustness and generalization with only a small fraction of the original training data.

Ehsan Pajouheshgar*, Yitao Xu*, Alexander Mordvintsev, Eyvind Niklasson, Tong Zhang, Sabine Süsstrunk

MeshNCA extends Neural Cellular Automata from regular grids to triangle meshes, enabling local self-organization and real-time dynamic texture synthesis directly on 3D surfaces without UV unwrapping or global communication.

Ehsan Pajouheshgar, Yitao Xu, Sabine Süsstrunk

We analyze NCA texture models in the continuous space-time limit, show that standard training overfits discretization near the seed, and use uniform-noise initialization to learn dynamics that stay consistent across resolutions, are robust to stochastic updates and additive noise, and behave like PDEs.

Y. Xu, Ehsan Pajouheshgar, S. Süsstrunk

We analyze the relationship between NCA architecture and emergent motion in the dynamics by varying cell-state dimensionality and MLP width, showing that their disparity and proportionality strongly correlate with motion strength and provide a practical design rule for dynamic NCAs.

Ehsan Pajouheshgar*, Yitao Xu*, Tong Zhang, Sabine Süsstrunk

DyNCA enables real-time and controllable dynamic texture synthesis, producing arbitrarily long and arbitrary-size video textures while improving synthesis speed and quality by orders of magnitude over prior optimization-based methods.

Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, Ariel Shamir

CLIPasso abstracts any object into a sketch by jointly optimizing the position and shape of Bézier strokes, using a CLIP-based perceptual loss to preserve semantic identity across arbitrary levels of abstraction.

Ehsan Pajouheshgar, Tong Zhang, Sabine Süsstrunk

We propose LELSD, a method that learns semantically meaningful directions in GAN latent space for localized image editing, enabling precise attribute changes in specific spatial regions without affecting the rest of the image.

A. Khodadadi, A. Hosseini, Ehsan Pajouheshgar, F. Mansouri, H.R. Rabiee

ChOracle models churn through user return-time prediction by combining temporal point processes, recurrent neural networks, and latent loyalty variables, and uses an efficient variational backpropagation-through-time training method that generalizes across diverse real-world datasets.

Ehsan Pajouheshgar, Christoph H. Lampert

We propose an auto-regressive spatio-temporal CNN that outputs explicit probability distributions over future trajectories from visual input, and show that this simple probabilistic baseline matches or outperforms more complex methods on MNISTseq and Stanford Drone benchmarks.