AI Neural Painting — Lucas Marques

The Problem

A personalized-gift store needed to turn customer photos into stylized artwork. Its existing workflow took approximately 26 hours and cost $5 per image. The goal was to automate image preparation and rendering while preserving the quality needed for printed products.

Automated AI neural painting workflow pipeline — The photo-to-gift workflow connects image preparation, neural rendering, and product output.

The Engineering

I developed a full automation pipeline using OpenCV, Pillow, and Stylized Neural Painting deployed on AWS. The system utilizes a dual-pathway neural renderer that translates images into paintings using vector representations. It incorporates Optimal Transport Loss to ensure similarity to the input and Differentiable Rendering to fine-tune the artistic output. The pipeline handles everything from face detection and background removal to the final application of oil painting effects.

Local Binary Patterns face detection visualization — Visualization of the Local Binary Patterns (LBP) algorithm used for precise face detection and alignment.

Dual-pathway neural rendering architecture diagram — The dual-pathway neural rendering architecture that transforms vector representations into stylized artistic strokes.

The Results

Processing time fell from approximately 26 hours to 15 minutes per image, about 100× faster. Processing cost fell from $5 to $0.25 per image, a 95% reduction. Automating the workflow allowed the store to handle more orders with a shorter turnaround.

Input customer photograph before neural painting — Original portrait supplied to the rendering pipeline.

Animated portrait being painted by the neural renderer — Watch individual strokes accumulate into a painted portrait, then compare the result with the input photo beside it.