Evolving Line Drawings

 

Ellie Baker

Margo I. Seltzer

Harvard University

Division of Applied Sciences

May 19, 1994

Goals

• Explore the power and limitations of interactive evolution
• Produce an artist's assistant
• achieve subtle highlighting and textural effects
• use a compact representation that is easily modified and transformed

Outline

• Introduction
• Interactive Evolution
• The Drawing Evolver
• Conclusions

Genetic Algorithms

• Model the process of biological evolution.
• Use random perturbations of a genome to create a population of "creatures."
• Apply a fitness criteria to select surviving creatures
• Repeat process
• Successfully applied to:
• Traveling Salesman Problem
• Graph Coloring
• Newspaper layout
• Animation of physically modeled figures

Interactive Evolution

• Use a human to provide fitness criteria
• Applicable where criteria is difficult to express computationally
• Previous applications
• biomorphs (Dawkins)
• face generation (Caldwell & Johnston)
• 3D sculptures (Todd & Latham)
• abstract color images (Sims)
• Key component:
• evaluation of visual data

Drawing Evolver

• Use interactive evolution to create drawings.
• User need not be able to draw, just select desirable images.
• Use mutation to affect small changes to an existing drawing.
• Use mating to create a drawing with components of two parent drawings.

Key Questions

• Can we use interactive evolution to create specific images?
• Does this technique produce images that would be difficult to produce with MacDraw-like tools?
• Is the tool engaging?
• What is needed to make it useful?
• What are the areas in which interactive evolution is particularly powerful? weak?

Representation

• Drawing is represented as a collection of strokes.
• A stroke is:
• a collection of points
• stroke type
• a symmetry property
• a connection type
• a perturbation factor
• a mutation rate

Getting Started

• Two modes: Random and User-Input
• Random: Initial Population

• Random: Evolved Drawings

Getting Started (2)

• User-Input: Initial Image

• User-Input: Evolved Images

Mutating

• Specify constraints to keep images in "face space".
• Randomly perturb points.

Mating

• Uniform Mating
• Independently select each stroke in each parent.
• Optionally weight stokes for inclusion.
• Face Application uses weightings of 0.3 - 0.7
• ID-Based Mating
• Group strokes into units (e.g. eyes, nose, mouth).
• Select entire group from one parent.
• Hybrid Mating
• For each set of strokes, select either Uniform or ID.

Uniform Mating

ID-Based Mating

Hybrid Mating

Resulting Images

Conclusions

• Achieves effects that are difficult with MacDraw-style drawing tools.
• Goal-oriented evolution is very difficult.
• For most people, a collection of pre-evolved images made the tool more engaging.
• Engaging for exploration.