Warp <war### [at] tagpovrayorg> wrote...

>   - Your patch certainly makes the so-called "color bleeding" more
evident.
> This, however, doesn't make the image look better or more realistic, only
> kind of overexposed.

From what I understand, Mr. Marty's Stochastic model has the potential to
produce more realistic results than MegaPov.  MegaPov's approach sacrafices
a bit of realism for speed and good-looking results.

Actually, the two approaches are very similar.  Stephane's model gathers
samples from the scene for every intersection.  MegaPov, on the other hand,
only gathers samples between 1% to 10% of the time (or more, if higher
quality is required), and interpolates between these gather points to fill
in the missing data.  The result is a much smoother image.  The overall
"correctness" of the MegaPov solution should be very good, but dome detail
is lost because of the interpolation.  Fortunately, this loss of detail is
more natural and pleasing to the eye than the noise that is introduced by
other stochastic (monte-carlo) ray-tracing techniques.  The Radiance engine,
which is known for its realistic output, uses a technique similar to the one
employed by POV-Ray and MegaPov (cached & interpolated stochastic
gathering).  Unfortuantely, POV's radiosity code still contains some bugs,
often leads to splotchiness when the error_bound is decreased.

>   - As seen in the two example images made with my scene, the brightness
seems
> to need some fixing. At least to me the megapov version looks more like it
> should be (I'm talking about the white wall).

Again, we are talking about how it "should" be even though we don't really
know what it really should look like.  Until someone converts the Cornell
Box to POV format, we won't really know.

>   - Your patch seems to have some graininess problem which is somewhat
evident
> in the example image in the www-page of the patch.  If I understood
correctly,
> it happens when the sample count is too low. In megapov even an extremely
> low sample count creates almost perfectly smooth illumination.

The graininess exists because samples are taken for each intersection point.
MegaPov removes this noise by interpolating between multiple sampling
points.  MegaPov's indirect illumination is actually smoother than it should
be (but because there is no high-frequency noise, the image looks better).

>   - Your patch is very slow compared to megapov

Actually, I'm interested in how _fast_ it is, considering how much work it
is doing.  800 samples for

>   Somehow I also have the feeling that using antialiasing would have made
the
> render time even longer compared to megapov (I may be wrong here, of
course).

I believe that you are correct.

-Nathan

Post a reply to this message