Ehsan Pajouheshgar

I am a postdoctoral researcher at EPFL Lausanne, working jointly with the Image and Visual Representation Laboratory (IVRL) and the Chair of Statistical Field Theory (CSFT) with Clément Hongler. I recently completed my PhD at EPFL under the supervision of Sabine Süsstrunk. Before joining EPFL, I obtained my bachelor's degree in Computer Engineering, with a minor in Mathematics, from Sharif University of Technology in Iran.

Research interests. I am broadly interested in Artificial Life and Computer Vision/Graphics, and in mixing ideas between these areas. In Artificial Life, I am interested in self-organization, self-replication, and in discovering or designing systems that exhibit unbounded, open-ended growth in complexity. In computer graphics and vision, I have worked extensively on textures and am particularly interested in the role textures play in visual perception. Much of my recent work uses Neural Cellular Automata (NCA) as a minimal, trainable model of self-organization and pattern formation.

Update: I am looking for postdoc or research scientist positions. Feel free to reach out to discuss potential opportunities or collaborations.

Email  /  CV  /  Google Scholar  /  Github  /  Twitter

MeshNCA is a type of Neural Cellular Automata that can operate on cells arranged on a 3D mesh. MeshNCA can simultaneously generate multiple textures (albedo, normal, roughness, ...) for Physically-based-rendering (PBR). It also enables grafting multiple NCA models on each other to create mixture of textures. Checkout the demo!

We propose a small change to Neural Cellular Automata (NCA) and show that the update rule learned by NoiseNCA actually correspond to a continuous space-time Partial Differential Equation (PDE). Utilizing this emergent space-time continuity, NoiseNCA allows changing the speed of the pattern formation and the scale of the patterns at inference time. Checkout the demo!

Dynamic Neural Cellular Automata (DyNCA), is a method that enables real-time and controllable dynamic texture synthesis. Checkout the demo!

Our work converts an image of an object to a sketch, allowing for varying levels of abstraction, while preserving its key visual features.

VNCA is a novel model for efficient volumetric style transfer that synthesizes multi-view consistent stylizing features on the target smoke in real-time, while exhibiting temporally coherent transitions between stylized frame.

We introduce Mokume, a dataset of 190 solid wood samples with high-resolution photos, ring annotations, and volumetric CT scans for realistic wood texturing. Using this data, we reconstruct a 3D growth field and synthesize volumetric color textures from a few exterior photos, combining procedural modeling with Neural Cellular Automata (NCA) to closely match real wood.

Trained Neural Cellular Automata (NCA) models exhibit many emergent behaviors, such as self-organization, and spontaneous motion. In this paper we investigate the spontaneous motion property of NCAs. We find that the ratio between the number of channels in the cell state and the number of hidden neurons in the update rule is a key factor that controls the NCA stability and emergent motion in the NCA generated patterns.

We introduce the Locally Effective Latent Space Direction (LELSD) framework, a novel approach to localized image editing in Generative Adversarial Networks (GANs), which utilizes a new objective function incorporating supervision from a pre-trained segmentation network.

We present an uncomplicated non-parametric baseline that attains the lowest error based on a widely adopted metric within the field. This serves to demonstrate the misleading nature of this particular metric.

Webpage Template