Would we want to change the values sampled from X for every pixel/frame or do we generate these values beforehand and use those for all the frames/pixels?
olliep24
I'm curious how to measure the tradeoffs between the number of rays and computational cost. Obviously with more rays, the cost to render the image increases, but what are the ways that we can measure the clarity/realness of the image? I feel like up to this point in the class, we've just used the eye test, but as we go into more realistic and physically modeled renderings, can there be a way to measure the "realness" of an image?
AnikethPrasad
@olliep24 This is a very interesting question. In my opinion "realness" is a very human quality. It might be possible to train a neural network to identify "realness" however I believe it would be quite difficult to train a model to evaluate such a biased metric.
myxamediyar
I think that the very first topics and this one connect in a lot of ways. The Nyquist theorem seems heavily tied into this - we can miss a lot about the image if we sample too slow. Say we completely skip the rectangle.
rcorona
If I'm understanding correctly, the images are respectively generated with 1 random sample per pixel and 16 averaged random samples per pixel (where each sample randomly selects a point on the light as the source).
Assuming that's the case, then in both images there is a constant number of samples per pixel.
I'm wondering if importance sampling is ever used to vary the number of samples per pixel? My question is motivated by the observation that pixels near the center of the shadow are going to have very low variance in values across samples (i.e. all of them are black), whereas pixels near the border of the shadow are going to have higher variance in values.
Therefore, can one save compute cost by weighting pixels to take additional samples from based on the variance of their previous samples?
carolyn-wang
I'm curious if there is also a technique to only sample 1 shadow ray but to run gaussian blurring filters on the surfaces that have shadows. Maybe this can get the same effect while saving computation costs?
MillerHollinger
I also have @ElShroomster's question. If you resample from X each frame in a video, wouldn't that cause the result to be noisy? Especially for a low number of shadow rays, I could imagine the shadow changing shape slightly each frame, which at high framerates would look very noisy.
Would we want to change the values sampled from X for every pixel/frame or do we generate these values beforehand and use those for all the frames/pixels?
I'm curious how to measure the tradeoffs between the number of rays and computational cost. Obviously with more rays, the cost to render the image increases, but what are the ways that we can measure the clarity/realness of the image? I feel like up to this point in the class, we've just used the eye test, but as we go into more realistic and physically modeled renderings, can there be a way to measure the "realness" of an image?
@olliep24 This is a very interesting question. In my opinion "realness" is a very human quality. It might be possible to train a neural network to identify "realness" however I believe it would be quite difficult to train a model to evaluate such a biased metric.
I think that the very first topics and this one connect in a lot of ways. The Nyquist theorem seems heavily tied into this - we can miss a lot about the image if we sample too slow. Say we completely skip the rectangle.
If I'm understanding correctly, the images are respectively generated with 1 random sample per pixel and 16 averaged random samples per pixel (where each sample randomly selects a point on the light as the source).
Assuming that's the case, then in both images there is a constant number of samples per pixel.
I'm wondering if importance sampling is ever used to vary the number of samples per pixel? My question is motivated by the observation that pixels near the center of the shadow are going to have very low variance in values across samples (i.e. all of them are black), whereas pixels near the border of the shadow are going to have higher variance in values.
Therefore, can one save compute cost by weighting pixels to take additional samples from based on the variance of their previous samples?
I'm curious if there is also a technique to only sample 1 shadow ray but to run gaussian blurring filters on the surfaces that have shadows. Maybe this can get the same effect while saving computation costs?
I also have @ElShroomster's question. If you resample from X each frame in a video, wouldn't that cause the result to be noisy? Especially for a low number of shadow rays, I could imagine the shadow changing shape slightly each frame, which at high framerates would look very noisy.