Well, the reason for needing such high samples is the extremely high Edge Sharpness in this. It's up at 20, the default is 1. I realize it may be necessary to achieve the desired effect, but it's good to know *why* the scene you're trying to create is so demanding, and important to realize that this is one of the hardest problems to solve in computer graphics: realistic cumulus clouds.
With that out of the way, I did find that Raytraced Atmosphere is beneficial here. You will see long render times regardless, but RTA with the right settings does help, assuming you want fairly high quality. I'll try to illustrate what I mean with the attached images. Please note that due to not using a GI cache (something I forgot to do until half-way through the tests, unfortunately), there are some dramatic lighting differences between tests. You should ignore the lighting and focus on *edge quality* and *noise level*, as these are what is primarily controlled by Detail and Quality (Samples), and in the case of RTA, also controlled by Antialiasing. Antialiasing for all renders was 3 in these tests (including RTA tests).
Image #1: The clouds at a Quality level of 2, which I believe is what is in the TGC. This gives you fairly high samples. It was rendered at a main render Detail of 0.5. Take note of 2 things. First, the *edge* quality is not very good. Second, there is still noise in the shadow of the cloud on the left. This noise is due to the *atmosphere* samples, not the cloud samples, in this case. Atmo samples were set to 48 here as in the clip file. This render took about 20 minutes on a Core i7 with 4 cores at 2.7Ghz.
Image #2: The clouds at a Quality of 1, my recommended value. This gives you a more reasonable number of samples. Likewise rendered at Detail 0.5. Note that the edge quality is basically the same, and that while noise is increased in the cloud shadow, it is not a *significant* increase, and more importantly there is not a notable increase in noise elsewhere in the clouds. This again reinforces the idea that it's *atmosphere* samples that need to be increased to get rid of that noise. This render was about 12 minutes.
Image #3: Reduced cloud Quality back to 1, but increased main renderer Detail to 1 (from 0.5 in the previous 2 tests). The noise appears more fine-grained overall, but there's still a good amount of it especially on that left side, so we can see it doesn't solve that problem (again, increasing atmosphere samples should). But what's important to notice here is that the *edge quality* is improved a noticeable amount. That's the same reason that the noise is more fine-grained; with increased main render Detail, the micropolygons being rendered are smaller, so you get better edge quality and finer noise grain. This render took *35 minutes*, obviously not ideal especially since there is still noise that needs to be dealt with, which would further increase render time.
Image #4: This is the first test with RTA enabled. Cloud quality is at 1. You can see that the edge quality is high, similar to the Detail 1 render above. There is also far less noise everywhere, despite having lower cloud Quality than the first test image above. This was with the TGC's 48 atmosphere samples, which as I said is the likely source of the noise, and this I think verifies that. This render took about 30 minutes, so it was faster than Image #3, has equivalent or better edge quality, *and* far less noise from the atmosphere. Clearly RTA is winning here.
Image #5: This final test just shows RTA with Quality 1 on the clouds and *16* atmosphere samples. You can see the noise is back but take note of 2 things. First, *it maintains high edge quality* on the clouds. Second, even though there is noise, it is fine-grained like the Detail 1 test above, and is arguably even less noisy than that by a small amount. Render time on this was about 20 minutes. So it seems you could increase atmosphere samples to maybe 24 or 32, get rid of the noise, and still have lower render time than the best quality non-RTA render. Note also that we didn't even match the result from Image #4 above with *any* of the non-RTA tests; we would have had to increase atmosphere samples on the Detail 1 render (Image #3) for that, and render time would have gone up even further, probably above 40 minutes. Again, clearly RTA is the winner.
So, my recommended settings after all that would be: Cloud Quality at 1 for all cloud layers, Atmosphere Samples at 24-32 (start low and increase if necessary), and enable Raytrace Atmosphere. Keep AA at 3.
Now here's where it gets potentially more challenging and complicated. IF your scene is going to include objects, especially plants or other complex objects that need higher AA values, then RTA starts to get harder to tune properly as your AA needs increase. At AA6 or so you might find RTA starting to lose to non-RTA rendering for overall render time and roughly equivalent quality (besides cloud edge quality, which requires Detail near 1, which is very render-time intensive). But if you don't have complex objects or other reason for high AA settings, then these recommendations are very good for you.
One key to keep in mind is that it always gives you high edge quality, even when there is noise, so in that way it is always higher quality than non-RTA; you get that for "free". *However*, if you care more about noise than edge quality, this could be a disadvantage because unlike non-RTA, you can't sacrifice edge quality (lower main render Detail) to save render time *while keeping samples high to reduce noise*. So depending on your specific needs, non-RTA might be something you could balance better.
Hopefully this has taught you a bit about how RTA works and when it's most useful. It is not always the right choice, but sometimes it can really give you better results in less time.
- Oshyan
