Optimization and Mobile Deployment for Anthropocene Neural Style Transfer

TL;DR

AnthropoCam tackles real-time neural style transfer on mobile to stylize Anthropocene textures while preserving semantic content. It adopts a pre-trained VGG-16 backbone with Gram-matrix style representations and a multi-objective loss, minimizing $L_{total} = \alpha L_{content} + \beta L_{style} + \gamma L_{tv}$; a feed-forward network enables 3–5 s high-resolution stylization on mobile. The main contributions are a systematic parameter study of layer choice, style set consistency, and loss weights; plus a practical mobile deployment pipeline using React Native and Flask with a GPU backend. This work enables in-situ, participatory visualization of Anthropocene landscapes, bridging technical NST optimization with environmental discourse.

Abstract

This paper presents AnthropoCam, a mobile-based neural style transfer (NST) system optimized for the visual synthesis of Anthropocene environments. Unlike conventional artistic NST, which prioritizes painterly abstraction, stylizing human-altered landscapes demands a careful balance between amplifying material textures and preserving semantic legibility. Industrial infrastructures, waste accumulations, and modified ecosystems contain dense, repetitive patterns that are visually expressive yet highly susceptible to semantic erosion under aggressive style transfer. To address this challenge, we systematically investigate the impact of NST parameter configurations on the visual translation of Anthropocene textures, including feature layer selection, style and content loss weighting, training stability, and output resolution. Through controlled experiments, we identify an optimal parameter manifold that maximizes stylistic expression while preventing semantic erasure. Our results demonstrate that appropriate combinations of convolutional depth, loss ratios, and resolution scaling enable the faithful transformation of anthropogenic material properties into a coherent visual language. Building on these findings, we implement a low-latency, feed-forward NST pipeline deployed on mobile devices. The system integrates a React Native frontend with a Flask-based GPU backend, achieving high-resolution inference within 3-5 seconds on general mobile hardware. This enables real-time, in-situ visual intervention at the site of image capture, supporting participatory engagement with Anthropocene landscapes. By coupling domain-specific NST optimization with mobile deployment, AnthropoCam reframes neural style transfer as a practical and expressive tool for real-time environmental visualization in the Anthropocene.

Figures (10)

• Figure 1: Overall system architecture for neural style transfer: Image Transform Net is the image style transfer system, belonging to the feed-forward computation system; Loss network is the loss calculation system, and the calculated loss is backpropagated to update the Image Transform Net.
• Figure 2: System Overview: This system combines a front-end application interface (mobile device) with content and style selection, and a pre-trained style transfer model on the back-end to achieve near real-time style transfer on mobile devices.
• Figure 3: The left column shows features from shallower layers with more detailed textures. The right column shows features from deeper layers, resulting in a more blocky image. (From top to bottom: plastic bottle, eutrophication, and urban building style)
• Figure 4: Effect of training set consistency on resulting visual textures. The left set of images show the output using more visually consistent style images, while the right uses more contrasting style images. The bottom row shows the style image sets used during training. It can be seen that similar styles maximize the presentation of style features (notice also the difference in clouds).
• Figure 5: Images in the top row feature semi-transparent objects, resulting in a hazy, ethereal quality, compared to the more vibrant colors in the bottom row.