Reveries

Rêveries – Generative Neural Paintings by Orkhan Mammadov

An Exploration of AI Aesthetics, Texture Synthesis & Simulation, and the Evolving Dialogues Between Machine and Human Creativity

1. Artistic and Conceptual Context

1.1 Core Concept and Inspiration

At the heart of Rêveries lies a fascination with how artificial neural networks—originally designed for tasks such as image classification and pattern recognition—can be repurposed to generate complex, painterly textures. Drawing on early inspirations from Google’s DeepDream experiments and subsequent developments in convolutional neural networks (CNNs) & Texture Synthesis Using Convolutional Neural Networks, Orkhan Mammadov’s work examines the ways machines “see” and abstract the world.

This series finds philosophical footing in the concept of dreamscapes: spaces where recognizable forms give way to fluid, evolving patterns. The artist embraces the partial disintegration of representational imagery, allowing intuitive play between human intention and machine processes. In this liminal zone, Rêveries proposes an aesthetic that is neither purely algorithmic nor entirely human—rather, a symbolic collaboration revealing the hidden “mind’s eye” of neural networks.

1.2 Evolution from first experiment with AI

Rêveries directly returns to Orkhan Media Art Lab's formative venture into AI-driven artwork, which began in 2016. While earlier projects integrated snippets of his custom neural code—such as creating carpet textures—this series marks the first time he fully revisits, refines, and applies his original model with the benefit of today’s more advanced knowledge. By circling back to his roots, Mammadov captures a narrative of personal and technological evolution, revealing how both the artist and the medium have matured.

1.4 Intended Audience Engagement

Rêveries is intended to function as a contemplative space that invites viewers to question how machines synthesize memory, texture, and form. The aim is to provoke wonder and introspection: What does it mean to “see” in the age of AI? How do we, as human viewers, reconcile the familiarity of natural textures with the uncanny distortions produced by deep learning algorithms? Mammadov hopes audiences will not only ponder the psychological and cultural dimensions of AI but also reflect on art’s enduring role in navigating the uncertainties of technological revolutions.

2. Creative Process and Technique

Rêveries builds directly on the techniques outlined in the seminal paper

2.1 Texture Synthesis

Rêveries's custom texture synthesis software inspired by the seminal paper "Texture Synthesis Using Convolutional Neural Networks" by Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge, the project investigates how the latent structures of artificial neural networks can be transformed into generative painting simulations.

At the core of the series is a custom generative approach that utilizes the feature spaces of convolutional neural networks (CNNs), originally developed for object and pattern recognition. Instead of classifying images, the network is reimagined as a creative engine that synthesizes complex textures by capturing the internal correlations between feature maps across multiple layers. Mammadov's studio demonstrates how textures evolve through the network’s hierarchy—gradually distilling the statistical essence of natural images while abstracting away recognizable forms. The algorithm thus becomes both a tool for creative exploration and a window into the deep representational spaces of artificial neural systems.

2.2 Texture Simulation

Building on the synthesized textures generated via CNN feature correlations, Rêveries extends its generative system into dynamic texture simulation—transforming static outputs into living, evolving surfaces. At the center of this process is a feedback loop architecture powered by GLSL shaders, allowing for real-time manipulation of visual material.

Each iteration begins with a generated source image and its corresponding depth map, from which a vector field is derived. This vector field encodes directionality and flow based on spatial and chromatic cues within the image. Simultaneously, binary and gradient masks are constructed to govern localized blending and displacement behaviors.

The source image is then displaced along the vector field twice—once with nearest-neighbor interpolation, once with linear. These two variants are fused using the generated mask, yielding an image that oscillates between sharpness and blur, order and fluidity. A secondary mask reintroduces fragments of the original image, seeding recognizable texture into the shifting abstract surface. The result becomes the input for the next frame, re-entering the system to generate new displacement fields and mask structures.

This perpetual recursion creates a visual ecology: textures flow, fragment, and coalesce in a state of continuous transformation. By offloading the computation to a chain of custom GLSL filters, the system operates at interactive speeds—inviting exhibition viewers to witness the real-time dreaming of the machine.

Credits:

Orkhan Mammadov

Slava Rybin