Enhanced sphere tracing paper

[subblue]

This paper discusses some techniques to speed up and improve the quality of sphere tracing ray marchers.
Some interesting snippets in there
http://erleuchtet.org/~cupe/permanent/enhanced_sphere_tracing.pdf

[cKleinhuis]

nice, thank you for sharing!

ehrm, i did not fully get the idea, it is using screen space to determine larger steps by inducing an error tracking ? something like that, can you give the idea of the paper in one sentence ?
and furthermore have you applied it to any raymarcher yet?  

i think the provided raymarching implementation is as easy as copy and paste to any existing script when i find some time i will take a look into it! 

[eiffie]

The main improvement suggested in the paper is over stepping. Simply put you can multiply the distance estimate by something like 1.5 and then if the next distance estimate is larger than the amount you over-stepped you know you didn't skip over anything. (because the distance estimate is omni-directional and covers the area you stepped over)
The trick is automating this overstep value depending on if you are approaching or moving away from a surface. I haven't seen an implementation of that yet.

[eiffie]

Here are some results of testing a simple auto-overstep with the following code:

Code:

float t=0.0;    //total distance
float d=1.0;    //estimated distance
float pd=10.0;//previous estimate
float os=0.0;  //overstep
for(int i=0;i<64;i++){	//march loop
	d=DE(ro+rd*t);	//get the distance estimate
	if(d>os){	        //we have NOT stepped over anything
		os=d*min(1.0,0.5*d/pd);//calc overstep based on ratio of this step to last
		t+=d+os;   //add in the overstep
		pd=d;    	//save this step length for next calc
	}else{		//we HAVE stepped over something
		t-=os;	//so remove the overstep
		pd=100000.0;d=0.1;os=0.0;//and reset so the next overstep will be small
	}
	if(t>MAX_DIST || d<MIN_DIST)break;
}

I would estimate about a 10% improvement. This image shows the number of steps as glow...

[cKleinhuis]

so eiffie, nice having you back, i can not spot any visual error, so you confirm speed improvement more or less without losses ?!

[eiffie]

The method assures nothing is stepped over unless the Distance Estimation function is faulty (returns a non conservative value). If you run out of march steps you get errors as usual. Some pixels may require MORE steps using this method so in that sense you can increase the number of visual errors. I didn't see this in my test.
Nice to be back.