Sunflow is an open source rendering system for photo-realistic image synthesis. It offers biased and unbiased rendering options. It is fairly robust for general situations although is lacking sophistication (as of Aug 2008) in some certain key areas for an optimized production pipeline. Certain problems could be resolved through joint development with the Sunflow community however the feasibility to resolve core internal render engine issues remains an open question. http://sunflow.sourceforge.net/

Osario Test Renders using Sunflow

This section shows various technical tests made using test scenes of the Osario. The test scene was created to give a base example of the general components of a typical scene: High-poly decimated mesh derived from scan data, flora, HDR sky, and various texture inputs. Currently this scene contains around 900,000 polys and the Osario wall maps are 2K each while the grass texture in front of the camera area up to the ruins is 4K. If decimated poly models are being used in the final render the final poly count will likely be an order of magnitude higher while the textures will be closer to 4K. Our main goals with Sunflow is to first get a feel for the workflow involved just to render a sequence, secondly to determine what Global Illumination options (biased rendering options) would be ideal (and if they have breaking thresholds), and lastly see what kind of issues arise in regards to shader limitations, rendering times, flickering issues and general stability.

render test v01

This first test was actual a pre-test to the Osario scene in which we just took one scan, converted it to polygons, placed UV's and applied a base texture from ortho-rectified photographs taken on site. It was converted from Blender as we did not have a functioning Maya2008 exporter for Sunflow. All lighting is derived from the HDR sky and it is using a biased method called 'Instant GI'. It was part of an animated camera sequence to test for flickering. The mirror ball was just side test to see behind the camera and how reflectivity holds up.

render test v02

This image was a geometry/GI test to see how Sunflow would deal with a 900,000+ poly scene. It was also a test to see how the exporter from Maya2008 was working as the initial compiled version for the Sunflow community was done at INSIGHT. Sunflow has 3 main options available for the Global Illumination (GI) methodology, 2 of them biased and 1 unbiased. This implementation of path tracing is not a practical option for this project (slow, grainy + flickering) so we focused on the 2 unbiased approaches 'Instant GI' and 'Irradiance Caching'. Irradiance Caching was the preferred method but proved unstable with this amount of polygons (in fact didn't work at all) so the only option that works currently is the 'Instant GI' option. For more details on these GI types please see the Sunflow Wiki GI section.

shader test v01

This image represents a quick and dirty attempt to control the BRDF of the test shader with some modified existing code available from the Sunflow Wiki. It is basically a combination of the standard lambertian diffuse shader with a basic fresnel calculation for returning the color component from 2 different textures. The 'quick' idea was that we could take the base color texture as the main color component and then at more glancing angles use a modified color texture to unscientifically 'simulate' the change in reflectance. Obviously this is not acceptable for real production but was an attempt to at least bring the Sunflow test renders closer to the tests being done in other packages which already include BRDF control. It was also an exercise in modifying a 'Janino' shader which is a java compiler compiling on the fly when encountered in the scene description. For more information about Janino within Sunflow see here. For more information about this specific 'textured fresnel' shader see the following forum posting.

render test v03

This image represents a test of the 'textured fresnel' shader introduced into the actual scene on one face. The effect of the shader would be more apparent if the camera was moving. A basic grass color texture was added to a diffuse shader to see how Sunflow would deal with resolving texture noise towards the horizon.

animated camera test v01

This is a small movie showing one of the fixes we did on the original exporter seen on left (0.9.6). When we rendered out the original camera move it was introducing a lot of rotational noise into the camera so we pinpointed the problem in the source code and re-compiled a new exporter (along with other small fixes) which is available from the the Sunflow forums here. If no QT controls are visible please click below the lower left hand corner of the image (in the black).

render test v04

This represents the first 'quasi real' test of Sunflow with the same test scene being used in Fryrender, mental ray and Radiance. While we were at SIGGRAPH we let this roll on one dual core machine which was doing about 1 frame every 3.5 hours for a high rez image (1920x1440). For unknown reasons it stopped after 27 frames (out of 48). We did not redirect the output into a text file so we do not have any record of what happened, something to include in the batch script for the next test. This scene contains all diffuse shaders, it does not include the 'textured fresnel' shader due to time constraints before leaving (setup is rather involved and manual). Also noteworthy are the tree leaves, which are not leaves at all but rather 4 sided polygons with a single green color. Currently Sunflow does not support alpha channels (bitmap or from the render output) so transparency is out of the question - all flora needs to be pure geometry.

render test v04 quicktime movie(11.3MB)