Christoph Hormann wrote:
> 
> After so many of you have tried randomized landscape i of course also 
> had to try some improvements - here are a few test renders.  In contrast 
> to most of the other tests i have seen (which really look good BTW) i 
> have made more closeup looks this time which poses additional problems 
> (at low sun angles i regularly get completely shadowed images for example).
> 
> The difficulty of a truely random system (i.e. where you really render 
> the image defined by a random number and not select one of the stream 
> that looks best) is to restrict the parameter ranges to regions 
> resulting in interesting images and at the same time not make all 
> renders look the same.
> 
> Since this work seems rather boring to me i got the idea that one could 
> have the computer automatically do this using standard optimization 
> methods.  There are mainly two problems about this: 1) you need a lot of 
> tests to find 'interesting regions in parameter space', especially in 
> high dimensional problems like this (i currently have about 100 
> parameters). 2) you have to measure the 'quality' of the resulting image 
> somehow, preferably automatically of course which is quite impossible.
> 
> So my idea is to render a set of images (maybe 10-20, something like a 
> reasonable population size for an evolution strategy) in small size 
> every week and have a public voting on them and use this vote for the 
> ranking.  Then maybe render the highest rated image in larger, generate 
> the next step population and start again.
> 
> I have not yet worked this out compeletely but it could be an 
> interesting thing to try out.
> 
> Christoph

These are great images!
And the idea is interesting, I would participate to the voting.

Sebastian H.

Post a reply to this message