After 100 passes:
[Edited by - Max_Payne on June 14, 2005 9:37:36 PM]
Photon Mapping : I still Believe I'm Right
Author
I am quite sure *someone* suggested I render an image with sharp caustics. I chose to render a common and fun example to see if it would converge properly:
After 1 pass (approx 100K photons stored):
After 100 passes:
[Edited by - Max_Payne on June 14, 2005 9:37:36 PM]
After 100 passes:
[Edited by - Max_Payne on June 14, 2005 9:37:36 PM]
Any expiriment needs a control, can you show us one pass of a high photon count reference image?
Nice looking result. However, it's meaningless without any comparison basis. If each pass was done with an effective 100,000 photons, you should have a total of 100,000,000 photons effective in the final result. How does a single-pass render with 100,000,000 photons compare? How about 50,000,000? 1,000,000? 500,000? How does the approach behave in a non-trivial case like the headlight example in this paper?
You'll also note I've said repeatedly that the averaging will converge to something that looks like the correct solution - but slower and less reliably than a single-pass approach. Without numbers and comparisons your statements really don't mean much. I mean, I could post and say that I got this image with 20 photons, but it wouldn't mean much [wink]
You'll also note I've said repeatedly that the averaging will converge to something that looks like the correct solution - but slower and less reliably than a single-pass approach. Without numbers and comparisons your statements really don't mean much. I mean, I could post and say that I got this image with 20 photons, but it wouldn't mean much [wink]
Author
I should have a total of 100 x 100,000 = 10,000,000 photons in a comparable image.
I have done a quick render with 1.3M photons for somewhat of a comparison basis. No AA because it would have been too slow without irradiance caching:
The headlights look quite complex to make, since I don't have a model support at the moment. Can you show me an image of a checkered cornell box render?
I have done a quick render with 1.3M photons for somewhat of a comparison basis. No AA because it would have been too slow without irradiance caching:
The headlights look quite complex to make, since I don't have a model support at the moment. Can you show me an image of a checkered cornell box render?
With antialiasing, I highly doubt it would take a single-pass renderer an additional 8.7 million photons to converge to the same effective result as the averaged sample... you'll also notice in the reference render that the area between the ring and the dim edge of the caustic is slightly brigher than the averaged render. This is evidence of data being lost.
You can see a somewhat decent example of a checkered-wall Cornell box at http://wfmh.org.pl/thorgal/jatrac/img/cb01.png I recommend making the pattern smaller in your test cases than in this shot to demonstrate high-frequency color bleeding.
You can see a somewhat decent example of a checkered-wall Cornell box at http://wfmh.org.pl/thorgal/jatrac/img/cb01.png I recommend making the pattern smaller in your test cases than in this shot to demonstrate high-frequency color bleeding.
Author
Quote:With antialiasing, I highly doubt it would take a single-pass renderer an additional 8.7 million photons to converge to the same effective result as the averaged sample...
I do not believe that would be a good comparison basis, because at 100 passes, the render is already nearly noiseless (I can hardly notice any on my high contrast monitor). I will provide you a shot of 10 passes at 100K photons each instead, which should be almost equivalent to 1 pass with 1M photons:
As you can see, it looks extremely similar. I can't really see any difference... (Well, of course, the noise isn't exactly the same). The only minor graphical difference, if any, could be attributed to anti-aliasing.
Quote:you'll also notice in the reference render that the area between the ring and the dim edge of the caustic is slightly brigher than the averaged render. This is evidence of data being lost.
I don't see any data being lost ;) The 100 pass shot is very accurate. You can see all details of the caustic, and it does not appear blurred.
Quote:You can see a somewhat decent example of a checkered-wall Cornell box at http://wfmh.org.pl/thorgal/jatrac/img/cb01.png I recommend making the pattern smaller in your test cases than in this shot to demonstrate high-frequency color bleeding.
Thanks for that shot. I will program a better implementation of lighting in my renderer (right now its hardcoded). Once that is done, I will make sure to give a go at rendering something very similar.
[Edited by - Max_Payne on June 14, 2005 10:24:31 PM]
Quote:Original post by Max_Payne
I do not believe that would be a good comparison basis, because at 100 passes, the render is already nearly noiseless (I can hardly notice any on my high contrast monitor). I will provide you a shot of 10 passes at 100K photons each instead, which should be almost equivalent to 1 pass with 1M photons...
Why wouldn't it be a good comparison? According to your viewpoint, they should be identical - they both have the same number of effective photons. I don't understand why you think this would be a bad comparison?
Author
Quote:Original post by ApochPiQQuote:Original post by Max_Payne
I do not believe that would be a good comparison basis, because at 100 passes, the render is already nearly noiseless (I can hardly notice any on my high contrast monitor). I will provide you a shot of 10 passes at 100K photons each instead, which should be almost equivalent to 1 pass with 1M photons...
Why wouldn't it be a good comparison? According to your viewpoint, they should be identical - they both have the same number of effective photons. I don't understand why you think this would be a bad comparison?
Fine :P The render should be ready by tommorow. 10M photons. I will add AA to it, since otherwise it will look bad.
Quote:Original post by Max_Payne
I am quite sure *someone* suggested I render an image with sharp caustics. I chose to render a common and fun example to see if it would converge properly:
The image doesn't look right to me - is the bowl emmisive? What reflectance function are you using?
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
