A system for real-time watercolour rendering

Abstract

A s S T R A C rWe present a system for the real-time rendering of scenes to produce images like watercolour paintings. Our system uses a hybrid of object space rendering techniques and image plane post processing to attain extremely good results very quickly. We achieve this through Graphics Processing Unit (GPU) programming, bardware accelerated shadow volumes, and the use of noise for various dif€erent tasks. We manage to circumvent problems found in both previous image space and object space methods by combining the two.Nonphotorealistic rendering ("R) is the family of techniques that can be used to produce scenes that do not exactly resemble the real world. Standard, realistic graphics techniques do not produce image features that are necessary in many NPR fields. The field that we are interested in is painterly rendering -the rendering of scenes in the style of paintings. Some of the characteristics of paintings that are missing h m standatd approaches to rendering are: 0 The sketchiness inherent in certain styles of painting.The presence of brushstrokeS. Complex colour interactions caused by paint on canvas rather than by light in the real world. NPR metbods that attempt to model these characteristics have typically worked on finished images [1][2][3]. That is to say, the system would be provided w i t h a rendered scene or photograph, and it would then ptocess it so thaf it resembled a painting. Our method does much of its work in object space.There are two key problems to painterly rendering in the image plane that we have dealt with thmugh object space * June 22-24,2005, Stony Brook, NY, USA 0-7803-9330-9105l$20 .OD 02005 I EEE techniques. Firstly, it is usually very slow. The speed of such processes is not a necessary feature of image plane techniques.Such methods are t y p i d y slow because of the large mount of processing needed at every pixel. The second problem is that the post-processing step usually has only the information contained in a single fiame. Because of this, generated features, such as brush strokes, are not coherent.We present a hybrid system for rendering images in the style of watercolour paintings. Working in hardware and doing much of the 'bard work' in object space, our system is very fast. Object space techniques produce h e -t o -8 a m e coherency almost automatically.Our system first produces various pieces of infomation about each pixel in the scene through object space methods, and then performs a post-processing step on this information to produce a final W e . In section 2 we look a! the features of real watercolour paintings, examine previous work in painterly rendering, and discuss the limitations and benefits of working on the GPU. In section 3 we. Iwk at how we can rapidly produce shadows using a technique found primarily in video games, and in section 4 we examine our paint layers algorithm €rom beginning to end. In section 5 we examine our results, and in section 6 provide a brief discussion of our success up to this point with an eye towards future work. BACKGROU...

Interactive watercolor rendering with temporal coherence and abstraction

“…Line integral convolution is applied along the principal curvature directions to render brush-like textures. Burgess et al [2005] created a similar system, using Wyvill noise to create stroke textures. Luft and Deussen [2005] abstract ID images at their boundaries to create flow patterns and edge darkening.…”

Real-time watercolor rendering of 3D objects and animation with enhanced control

“…Not long afterwards, Burgess et al [2005] present a procedural combination of object-space and image-space techniques which also takes advantage of the GPU and programmable shaders for its watercolor approximation. While highly inspired by the approach of Lum and Ma [2001], the method developed by Burgess et al approaches watercolor with real-time performance in mind, avoiding the use of Line Integral Convolution.…”

“…Burgess et al [2005] most likely proposed the first approach towards pigment turbulence, by extending the work of Lum and Ma [2001] that used Line-integral Convolution (LIC) [Cabral and Leedom, 1993] to emulate strokes along the ob-ject's surface. Instead of strokes, Burgess et al [2005] aimed to generate textured thickness (pigment turbulence), by distorting Wyvill noise [Wyvill, 1999] along the surface of the geometry. The resulting thickness would then modify the color using the Kubelka Munk color model [Haase and Meyer, 1992].…”

“…Their approach darkens the edge, but it does not concentrate the color or showcase a gradient accumulation of pigments. Burgess et al [2005] improved upon this by proposing the use of color and depth to find edges. A set of 25 sparse neighboring pixels were sampled and, depending on their differences and a set of thresholds, an edge was identified.…”

“…Bousseau et al [2006] resorted to the use of a simple symmetric kernel to detect edges, which were later used to concentrate the pigmentation using their proposed color transmittance modification model. Similarly to Burgess et al [2005], their approach concentrated color at the edges, but does not incorporate the gradient pigment accumulation that happens with the natural medium. Luft and Deussen [2006a] used a Difference of Gaussians to subtract the alpha contribution of segmented elements, leaving the edges with the original color, but lightening the inside of the segments, instead.…”