Morphological Antialiasing on GPU

Abstract

Morphological antialiasing is a post-processing approach which does note require additional samples computation. This algorithm acts as a non-linear filter, ill-suited to massively parallel hardware architectures. We redesigned the initial method using multiple passes with, in particular, a new approach to line length computation. We also introduce in the method the notion of topological reconstruction to correct the weaknesses of postprocessing antialiasing techniques. Our method runs as a pure post-process filter providing full-image antialiasing at high framerates, competing with traditional MSAA. Our complete implementation is freely available.

NEWS !

  • NEW 03/08/2011 : Check this very good presentation / explanation of our technique here
  • NEW 24/03/2011 : New version of GPU TMLAA available with cleaner code and english translated commentary ! Thanks to L. Rahyen for the translation.
  • 30/12/2010 : A new version of GPU TMLAA is available on the download page. The new shaders are finally here (we wanted to wait for GRAPP conference results) and accelerate quite the computation (x2/x3)! The new TMLAA compete now with classical MSAA in terms of speed. Moreover, a new algorithm to complete some missing pixel, due to the aliasing, is proposed (check the GRAPP publication for details). Depth buffer can also be loaded and used (in the code proposed, solely for the pixel reconstruction but it can be used for discontinuity detection too).
  • 21/11/2010 : A new version of GPU MLAA is available on the download page. Bug fix for 64-bit compilation + english comments on shader files + minor bug correction on area computation.
  • 12/11/2010 : A new version of GPU MLAA will be available soon with an algorithm 2x faster than the previous (on my laptop ! should be even better on newer GPU)
  • Lang Fox has done an implement of our MLAA GPU method on the Ogre Rendering Engine. Thanks to him !

Related research publication

Communication at GRAPP'11 (to appear - preprint paper)

Presentation slides at SIGGRAPH'10

Presentation paper at SIGGRAPH'10'

Results in image

example research image example research image


Two images presenting differences between the MLAA (at left) and the MSAAx2 (at right) in a Sponza like scene (note that shading differs a lit bit between the two images).

example research image example research image example research image


Three exemple of MLAA in various shading situation (bloom, shadows, toon shading). Note that TMLAA corrects also aliasing du to shading (like in the shadow exemple at center image)

example research image example research image




Differences without (left) and with (right) topological reconstruction (see GRAPP paper for further explanations)


Results in video

example research video


This video points out the use of our algorithm with deferred shaded scene since the antialiasing is applied as a post process. The video shows the source image (top left square), the corrected image (top right square), the discontinuities detected (bottom left square) and the alpha map generated (bottom right square).


example research video




This video is another example of the use of MLAA in a real time rendering engine.