Supplementary Material for "Evolving Three Dimension (3D) Abstract Art: Fitting Concepts by
Language"
Colab Notebook for code:
[Open in Colab]
[Download]
Author: Yingtao Tian
Abstract:
Computational creativity has contributed heavily to abstract art in modern era,
allowing artists to create high quality, abstract two dimension (2D) arts with a high level of controllability and
expressibility.
However, even with computational approaches that have promising result in making concrete 3D art, computationally
addressing abstract 3D art with high-quality and controllability remains an open question.
To fill this gap, we propose to explore computational creativity in making abstract 3D art by bridging evolution
strategies (ES) and 3D rendering through customizable parameterization of scenes.
We demonstrate that our approach is capable of placing semi-transparent triangles in 3D scenes that, when viewed
from specified angles, render into films that look like artists' specification expressed in natural language.
This provides a new way for the artist to easily express creativity ideas for abstract 3D art.
Table of Content
-
Evaluation 1:
Our method places semi-transparent triangles in three dimension (3D) spaces using
Evolution Strategies.
-
Evaluation 2: Our method is capable of making a 3D art
following the spatial abstract art
style, that looks like what humans can compose in natural language text.
-
(There is no Evaluation 3, to keep the numbering of evalutions in line with Figures in paper.)
-
Evaluation 4: Our method leverages the budgets of
triangles in the increasing order of
granularity, by first using triangles for general shape and then moving towards fine-grained details.
-
Evaluation 5: In our method, different runs lead to
equally plausible yet largely
different 3D art.
-
Evaluation 6: In our method, the fixed transparency
setting allows more global control of
the scene, and the
learnable one provides great
flexibility in how triangles are related to the space.
-
Evaluation 7: Our method produces one 3D art, and
successfully allows it to look differently from different angles.
Evaluation 1: Our proposed method places semi-transparent triangles in three dimension (3D) spaces using
Evolution Strategies.
Back to top
This corresponds to the Figure 1 in the paper:
Our proposed method places semi-transparent triangles in three dimension (3D) spaces using Evolution
Strategies.
Leveraging ray-tracing based rendering Mitsuba 3,
the rendered film at possibly multiple cameras is compared with its corresponding, user-specified text prompt
using distance between their representation embedded by CLIP.
Such distances, aggregate by average, are used as the fitness in sense of Evolution Strategies, which optimize the
parameters of triangles to archive better finesses.
Evaluation 2: Our method is capable of making a 3D art following the spatial abstract art style, that
looks like what humans can compose in natural language text.
Back to top
This corresponds to the Figure 2 in the paper:
Several examples of the evolved 3D art produced by our method, where the evolution process places triangles inside
the unit cube space visualized by black frame and sets triangles' colors and transparencies, forming a spatial
configuration.
In each example shown here, four cameras look at the space from four sides, although this is an arbitrary decision
and cameras can have different numbers and directions.
The film from each camera, capturing the rendered images, is compared with the prompt.
It could be observed that our method is capable of making a 3D art, which follows the spatial abstract art
style,
that looks like what humans can compose in natural language text.
Evaluation 4: Our method leverages the budgets of triangles in the increasing order of granularity, by first
using triangles for general shape and then moving towards fine-grained details.
Back to top
This corresponds to the Figure 4 in the paper:
Our method generating with text prompts "Walt Disney World" with four cameras,
with different numbers of triangles, namely 10, 25, 50 and 100 respectively.
It could be shown that our method leverages the budgets of triangles in the increasing order of granularity, by
first using triangles for general shape and then moving towards fine-grained details.
Evaluation 5: In our method, different runs lead to equally plausible yet largely different 3D art.
Back to top
This corresponds to and extends the Figure 5 in the paper:
Our method generating two configurations, each with two independent runs.
The first configuration is with text prompts "A bright, vibrant, dynamic, spirited, vivid painting of a dog" from
four cameras,
while the second configuration "Walt Disney World" .
Different runs lead to equally plausible yet largely different 3D art.
An artist user could thus be "in-the-loop" by choosing from different variants from these runs.
Evaluation 6: In our method, the fixed transparency setting allows more global control of the scene, and the
learnable one provides great
flexibility in how triangles are related to the space.
Back to top
This corresponds to the Figure 6 in the paper:
Our method generating with text prompts "Walt Disney World", with four settings of transparency.
We show here three fixed transparency of 0% 50% and 80% compared with the default setting of learnable
transparency.
While the fixed transparency setting allows more global control of the scene, the learnable one provides great
flexibility in how triangles are related to the space.
Evaluation 7: Our method produces one 3D art, and successfully allows it to look differently from different
angles.
Back to top
This corresponds to the Figure 7 in the paper:
Our method generating with different text prompts at cameras.
The text prompt for camera 1 and 3 is "Walt Disney World" and for camera 2 and 4 is "an annoyed cat".
Our method produces one 3D art, and successfully allows it to look differently from different angles.