I'm currently implementing a new UberPOV feature that allows bypassing 
the caching of radiosity samples, effectively resulting in purely 
stochastic unbiased Global Illumination computation. (See recent 
discussion on povray.unofficial.patches.)

The first image shows the cornell.pov radiosity sample scene, with 
slight modifications to the radiosity settings to eliminate artifacts. 
The scene rendered at 3018 pps on my machine. Anti-aliasing was set to 
+am2 +a0.3.

The second image shows virtually the same scene, but with radiosity 
caching disabled, the radiosity count parameter set to a much lower 
value, and rendered with UberPOV's oversampling ("anti-aliasing") mode 3 
with settings chosen to render in about the same time; speed was 3076 
pps in this case. Oversampling was set to +am3 +a0.056 +ac0.9.

The third image shows exactly the same scene as the second one, but with 
+a0.01 (in mode 3, this specifies the accepted noise level). Render 
quality has obviously indreased, but so has render time: With these 
settings, the scene rendered at only 229 pps.

It should be noted that relatively simple scenes with plenty of even 
surfaces are easy prey for the radiosity algorithm, so it is no surprise 
that this new feature's performance compares quite poor with this scene.


Some benefits of this approach over radiosity:

- Much lower memory footprint especially at high-quality settings.
- Much fewer interdependent quality settings to mess around with.
- No systematic limitations to quality; the result can always be 
improved simply by using more aggressive oversampling.
- Unlike radiosity, the algorithm does not require any communication 
between threads, and will therefore scale effortlessly in a distributed 
rendering environment.
- It is well-suited for a "keep investing render time until I like the 
output" mode of operation.

Post a reply to this message

Attachments:
Download 'cornell_rad.png' (77 KB) Download 'cornell_st.png' (313 KB) Download 'cornell_st01.png' (226 KB)

Preview of image 'cornell_rad.png'

Preview of image 'cornell_st.png'

Preview of image 'cornell_st01.png'