Caustics quality

[vlado]

Caustics quality

What is your QMC noise threshold? I hope to finally get some time to play with the scene this weekend.

Best regards,
Vlado

[Thomas An._cg]

Caustics quality

What is your QMC noise threshold? I hope to finally get some time to play with the scene this weekend.

Best regards,
Vlado

Image Sampler: Adaptive QMC [1,4] Antialiasing [Area 1.5]
QMC Sampler: Ad.Am=0.02 Noise=0.002 MinSmpl=5 Mult=1
Irradiance Map: MinR=-6, MaxR=0, SubD=80, Samples=20, Colr=0.3, Nrml=0.1, Dist=1, Delone, DensityB 10
Light Cache: SubD=2500, sampleSize=0.002, Scale=Screen, ph=1, Reconstr=None

Vlado, if you do get around to try this scene, is it possible to try it on a single processor PC having similar stats to this --> PC specs = AMD 64 3500 1GB.

[flipside]

Caustics quality

Adaptive amount 0.02 is just crazy. That means you're almost not using adaptive sampling, you're making some kinf of brute force rendering by using that low value (probably not the right explanation but I guess you know what I mean )

I also can't get the noise on the blurry coins to go away, I'm not sure if it is related to the material, or to the caustics falling on it.

-6/0 is also a bit extreme. I don't think the first 2 or three passes wil actually help. Remember how it works, your first pass will be computed on a 800/2/2/2/2/2/2=12.5px wide image. Don't think vray will get much usefull info out of that. It will not improve this image much, it's just a general tip

One thing to try, to avoid blurring your ir map solution too much, is decreasing the 'samples' value in IR map controls. The bigger this number, the more surrounding samples get interpolated together, resulting in smoother images, but less detailed GI. So for this type of caustics, lowering it from 20 to for example 12, will keep your GI very sharp. Try it with 4 or smth on a simple scene, you'll see the drastic impact it has

For Vlado, this image is extremely difficult if you keep the 'clamp colors' turned off. It's hopeless to get clean AA on the very bright reflections.

[Thomas An._cg]

Caustics quality

Adaptive amount 0.02 is just crazy. That means you're almost not using adaptive sampling, you're making some kinf of brute force rendering by using that low value (probably not the right explanation but I guess you know what I mean Smile )

You are right. I should have used 0.2 or something, but other higher values would introduce blotchy GI. I have found that high values (like 0.85) also decrease the caustic density.

-6/0 is also a bit extreme. I don't think the first 2 or three passes wil actually help. Remember how it works, your first pass will be computed on a 800/2/2/2/2/2/2=12.5px wide image. Don't think vray will get much usefull info out of that. It will not improve this image much, it's just a general tip

I wouldn't worry about it I have done a lot of test the past weeks on these values and it turns out that they do not really add to the render times. (I can tell with good certaintly that the culprit is not there). Also, based on my tests it turns out that adding those low passes -6, -5, -4 it also adds an extra layer of sample points to the irradiance solution. In other words the irradiance map ends up being more dense than if you choose -3 0. In this case the more dense the map the better, because anything else creates incomplete (irregular) not smooth caustics. These values have been researched quite extensively.

One thing to try, to avoid blurring your ir map solution too much, is decreasing the 'samples' value in IR map controls. The bigger this number, the more surrounding samples get interpolated together, resulting in smoother images, but less detailed GI. So for this type of caustics, lowering it from 20 to for example 12, will keep your GI very sharp. Try it with 4 or smth on a simple scene, you'll see the drastic impact it has

Less samples would make it more grainy (I have tried it )

So, the IR settings have been researched for highest IR caustic density. (I have spent over 90 hours on this image)

For Vlado, this image is extremely difficult if you keep the 'clamp colors' turned off. It's hopeless to get clean AA on the very bright reflections.

Yes, the aliasing is very difficult. I have tried all aliasing algorithms and and at various settings (from 1 to 4) ... but it doesn't seem to help. Clamping the colors is simply not an option in this image... I really want to see those burned out areas !

[flipside]

Caustics quality

Hey thomas,

I know the passes don't help much, that's whay I said it was just a general tip, not really for this image. There are better ways to add more samples to the map than by using extreme low passes.

Higher values for adapt amount indeed means more blotchy GI. But remember that lowering this value affects everything, also your DOF and glossies for example. So if you want better GI only, this is not the way to go because it will improve quality of a lot of other stuff too, that you don't even notice.

Less samples, I mean the samples value, not the hsph. It will indeed be more grainy, but grainy means sharp. Increase other values to get rid of the grain, but still retain the sharp GI.

Delone triangulation also means you need higher GI settings, as it is a non blurry method. I believe the samples value from above has less effect when using delone, so maybe if you keep delone, samples doesn't matter much.

I'm going to dig up this scene again

[Thomas An._cg]

Caustics quality

Higher values for adapt amount indeed means more blotchy GI. But remember that lowering this value affects everything, also your DOF and glossies for example. So if you want better GI only, this is not the way to go because it will improve quality of a lot of other stuff too, that you don't even notice.

Exactly ! The coines would not look as good with higher values either... and... despite all that and even at these extreme low adaptive values and at 80 smaples per coin (as well as Adaptive QMC sampling) they are still garinny.

Less samples, I mean the samples value, not the hsph. It will indeed be more grainy, but grainy means sharp. Increase other values to get rid of the grain, but still retain the sharp GI.

Delone is the only acceptable one, I have tried the others but they blur the little caustic spots around the coins and I want those really sharp. I found the other methods harder to produce a good cautic result.

I'm going to dig up this scene again

If you do ... can you try it on a single processor PC ?
I mean if your dual can get you a 12500 photon map in 1hr and mine gets it in 4hrs .... its not fair

[flipside]

Caustics quality

Hehe, just multiply my rendertime with 4 then

[flipside]

Caustics quality

Have you tried QMC GI for first engine? Becausein your case, if you want really exact GI, this could be faster than IR map. The use of IR map is to speed up GI, at the cost of some detail. Since detail is all you want here, maybe IR map is not the best engine. (it is better however if you want a really big render, because you can then save the IR map and reuse it on the big render, and that is not possible with QMC GI or PPT)

[Thomas An._cg]

Caustics quality

hmm, no I haven't tried QMC GI ... I mean.. I did try it briefly, but gave up on it for some reason which I can't remember. Somehow I do not think QMC will be faster than PPT fot this scene though.

P.S. These kind of heavy caustic scenes are quite important for me, beacuse some of my work involves jewelry.

[flipside]

Caustics quality

Aaaah that explains a lot

QMC was just a though, didn't try it extensivly too. I did one test, which was quite nice, but just like ppt, to get the caustics ok, it needed much more rendertime.

Here is that old test, 1h50min on my pc:

[Thomas An._cg]

Caustics quality

I see you fixed the mesh problem on the coins (sorry about that)

Yeah, the QMC looks promicing quility-wise, but that image will take a lot longer than 2 hours... I mean those caustics look nice and we might be tempted to think "it shouldn't take much longer to fill in the rest of the grain" ...but... from my experience with brute force solutions, it will take quite longer.

[flipside]

Caustics quality

Here's a similar test as yours, IR+LC only.

I used 0.5 for adaptive amount, -3/1 for min/max, clr=0.25, hsph=110, raised all glossy subdivs to 40, raised the dof subdivs a bit.

These were just 'guessed' settings, the only goal was to prove that you don't need such a low adaptive amount. Maybe the result would have been the same if I left it at 0.85. I don't use delone and I used sample value of 12 instead of 20. Also here, you see this doesn't always mean more noise, if you use high enough hsph. It simply keeps GI sharper. I know you spent 90h on this image, but I think I will be close (not on this last test though, that was just 5 minutes rasing every setting). But believe me, I spent way more than 90 hours on testing IR map and LC on vray for max

Now you will say that there is still noise on the coins, but the same in your render, so this seems to be impossible to clear up, even with that low adaptive amount and huge amount of glossy subdivs.

The cause of it is that really bright area light you have in the scene. If you would use a different lighting setup, which isn't necessarily less nice, you could avoid this problem. Or maybe even use directional light instead of an area light for that. But then again, there is no area shadow option on that yet. It would make caustic creation much easier though. Maybe you have another scene to test vray with, because I don't think it is really fair making a decision on only one scene.

Also as I said before, you will not suddenly get a really quick render if you are after this kind of quality. The methods you are using are designed to make renders faster with the drawback of less accurate GI or glossies or whatever. You simply cannot have both: quick render and perfect accuracy. So either choose for a 16H render, or satisfy with a bit lesser quality.

Here my latest test:

[Thomas An._cg]

Caustics quality

Six hours is not too bad, but then I remember that your machine is 2 to 4 times faster than mine... this means that it might go from anywhere between 12-24 hours if I were to try your settings.

You might be right about the adaptive amount, but on the other hand the 110 hmsmpl is a bit on the exotic side as well... and ... you are forced to go -3 / 1 (going from 0 to 1 is substantial bump, much much much more than going from -6 to -3. The -6 to -3 is no more than 10 min of calculation, but from 0 to 1 can be more than an hour or two at 110 samples )

The cause of it is that really bright area light you have in the scene. If you would use a different lighting setup, which isn't necessarily less nice, you could avoid this problem. Or maybe even use directional light instead of an area light for that.

Well, that would be almost unacceptable in my book. I refuse to let a render engine dictate what I can and can't do with my scenes.

The glow on the right-hand coin was a planned feature attraction as well as the burned out areas. It is these subtleties (or not so subtle) of light that can make a jewelry image stand out.

The main thing that bugs me now is the aliasing on the bright areas. Vlado, if you are following this thread.... is there any room for improvement on that aspect ?

You simply cannot have both: quick render and perfect accuracy. So either choose for a 16H render, or satisfy with a bit lesser quality.

Well, I wouldn't give up just yet. One thing that is quite clear from your test is that the GI caustics method is 6h versus 1.5 hours for the caustics-photon-mapped method on your machine. This means that the photon mapped caustics are about 4x faster (so I was off by a bit on my previous estimate; that they should be comperable)

So, the next thing is, I will try to render again and let it compute and finish the photon map... if I can get it to go down to 6 hours for the same quality then it will be considered good progress (6h is much preferable than 18 )

[flipside]

Caustics quality

About the -6 to -3, this will not add much to the render time, but it will also not add that much to the quality of GI. I mean if you compare it with the effect of going from 0 to 1. It's the second number that has the most effect, that is the 'max' value. This also determines the GI sharpness, for example going from -4/-1 to -4/0 will be much better quality than going from -4/-1 to -3/-1.

I use high hsph because I wanted better GI only, not better DOF of better glossies. By using low adapt amount, you increase quality of everything, and that I wanted to avoid. 110 was just a guess, I didn't test anything in between. Remember that going from 50 to 100 for example is not twice as good or twice as slow. 50 means 50*50=2500, and 100 means 100*100=10000! I believe previous test I did had hsph=70.

Problem is as you also mentioned somewhere that you cannot go and test all in between values, having to wait several hours and then see that you needed higher hsph.

However if you need high res output, ir map is a good option, because in my example if I needed to render at 3200*2400 px, I know that I will have same GI detail with -5/-1 as I had with -3/1 at 800*600. In other words, on the 3200px image, the IR map will take exactly as long to render than on the 800px image. This is not the case with ppt or qmc GI, that will take waaaaaay longer to render (but quality will become even better of course)

[Thomas An._cg]

Caustics quality

However if you need high res output, ir map is a good option, because in my example if I needed to render at 3200*2400 px, I know that I will have same GI detail with -5/-1 as I had with -3/1 at 800*600. In other words, on the 3200px image, the IR map will take exactly as long to render than on the 800px image. This is not the case with ppt or qmc GI, that will take waaaaaay longer to render (but quality will become even better of course)

That is a very good point !
I was thinking about possible render times from high res images... and this is a good tip. Thanks

Also, thanks for investing a whole lot of your time on this scene, this experience has helped me a lot; figuring out the settings and the values.

-Thomas