M2R 2016-2017 project:
"Studying and solving the strange visual artifacts occuring when procedural texturing with paradoxical requirements"

( see video or shadertoy )


Fabrice NEYRET   -   Maverick team, LJK, at INRIA-Montbonnot
Joelle THOLLOT - Maverick team, LJK, at INRIA-Montbonnot
Romain VERGNE  - Maverick team, LJK, at INRIA-Montbonnot


Procedural textures such as Perlin noise are a convenient and cheap way to provide natural looking random details. But animating their parameters along time, or making them change continuously in space, induces very visible and distracting artefacts that our eyes interpret as secondary shapes and motions doing totally unexpected things. A part is the way « the maths » « solves » paradoxical requirements ( e.g. « keep the same »and « zoom »), and a part relates to what the human visible system is very sensitive or very blind about.
E.g. : See various artifacts in « Interpolation and blending tricks » shadertoys ). In the
video and shadertoy where we « zoom and keep the same » Perlin noise, islands appear suddenly for B&W images… but no longer if we allow just a little bit a bit of grey.
We propose a very delimited study workbench,  but the problem of temporal coherency and parasitic secondary motions occurs in many situations like paint-like rendering [1,2,3,4], fluid particles [5], advected textures [6,7], and many times a user just innocently wanted to modulate a noise parameter along a shape (cf shadertoys above).

Description of the subject

Since clandestine secondary motion is already obvious in very basic cases such as infinite zoom in black and white Perlin noise, this allows to settle a convenient experimental bench to test the interactions between various algorithmic combinations, textural characteristics, and families of perceivable artifacts. A first step would be characterizing and organizing the families of artefacts, to somehow initiates a domain of artefactology :-) .

A part of the work will consist in experiments with the bench above, testing the perceptual effect of various combinations and tunings, possibly some light perception experiments, but also trying to tell the relation between the effect (e.g. secondary motion, parasitic wavelength) and the maths of the model.

Beside the beginning of a general understanding of the issue, we expect from this studies to get recommendations helping to improve time coherency in various applicative algorithms: Which kind of transformations are unperceived or extremely perceived, or which range of parameters make them so.  Plus possibly, to suggest some slight changes or new variants in tools, normalizations, and tunings to avoid the artifacts.