Why does the shadow based on sampling random points look so goofy? Is there any way to make it look better without too much of a hit on performance?
henrykhaung
I think it just looks goofy because of the nature of random sampling which introduces noise. To make it look better, maybe we can use the techniques we learned from the rasterization section like increasing the sample count or bilinear filtering.
epark272
I wonder if we could do some post-processing on the Monte Carlo shadows to make it look less noisy. For example, applying a Gaussian blur, or perhaps some other kernel that we can convolve.
RishSharma7
I would assume the center of light sampling yields a distinct shape in the shadow because of its consistency in sampling, not because we're sampling from the center. If we sampled from the corner, I think the shadow would still look similar, right? And if that's true, it should explain in contrast why a random point on light type of sampling would result in this pixelated (for lack of a better word), blurry shadow that we see on the right. I still like the way it looks, though. I assume it can have its own artistic use cases.
brianqch
Following up on henrykhaung's comment, I wonder if we can apply these sampling techniques after the shadows have been generated. Is supersampling still valid in this situation because it will also have an effect on the rest of the image? Or is it possible to apply supersampling during the generation of the shadows?
eugene-yoojin-han
Following up on epark272's comment, I also wonder if there would be a way to make it look less noisy. But if we can apply Gaussian blur to the shadow, why don't we just apply the filter after rendering the shadow with the setting on the left image?
buggy213
the best solution here is just to increase the number of samples per pixel, though this can be expensive. it's quite a difficult scene to render due to the sharp edges of the geometry, so some sort of denoising pass could also be good. generally, these denoisers are a bit more sophisticated than just using gaussian blur, since that would be problematic for sharp edges where we want them to occur (e.g. the edges of the plane itself, rather than its shadow). more details can be found at https://alain.xyz/blog/ray-tracing-denoising, this guy's blog is a goldmine for lots of graphics things
stang085
Looking at the other comments, I feel like we can apply some kind of bilinear sampling or something to smooth out the shadow on the right in order for it to look more realistic. While the shadow on the left is very crisp and clear, that's not really how shadows look most of the time in real life unless the object is very close to the surface.
Why does the shadow based on sampling random points look so goofy? Is there any way to make it look better without too much of a hit on performance?
I think it just looks goofy because of the nature of random sampling which introduces noise. To make it look better, maybe we can use the techniques we learned from the rasterization section like increasing the sample count or bilinear filtering.
I wonder if we could do some post-processing on the Monte Carlo shadows to make it look less noisy. For example, applying a Gaussian blur, or perhaps some other kernel that we can convolve.
I would assume the center of light sampling yields a distinct shape in the shadow because of its consistency in sampling, not because we're sampling from the center. If we sampled from the corner, I think the shadow would still look similar, right? And if that's true, it should explain in contrast why a random point on light type of sampling would result in this pixelated (for lack of a better word), blurry shadow that we see on the right. I still like the way it looks, though. I assume it can have its own artistic use cases.
Following up on henrykhaung's comment, I wonder if we can apply these sampling techniques after the shadows have been generated. Is supersampling still valid in this situation because it will also have an effect on the rest of the image? Or is it possible to apply supersampling during the generation of the shadows?
Following up on epark272's comment, I also wonder if there would be a way to make it look less noisy. But if we can apply Gaussian blur to the shadow, why don't we just apply the filter after rendering the shadow with the setting on the left image?
the best solution here is just to increase the number of samples per pixel, though this can be expensive. it's quite a difficult scene to render due to the sharp edges of the geometry, so some sort of denoising pass could also be good. generally, these denoisers are a bit more sophisticated than just using gaussian blur, since that would be problematic for sharp edges where we want them to occur (e.g. the edges of the plane itself, rather than its shadow). more details can be found at https://alain.xyz/blog/ray-tracing-denoising, this guy's blog is a goldmine for lots of graphics things
Looking at the other comments, I feel like we can apply some kind of bilinear sampling or something to smooth out the shadow on the right in order for it to look more realistic. While the shadow on the left is very crisp and clear, that's not really how shadows look most of the time in real life unless the object is very close to the surface.