Hacker News Reader

Comment by nickpsecurity | Hacker News Reader

How am I supposed to compare these three on the top as a non-expert in this area? The middle one looks noisier than the other two. I notice the one on the left has darker color in dresser which may be an advantage of correct, color handing but might be incorrect with one on right being brighter due to the correct handing of window's light on it. And what's up with the graininess of all of them?

Just a few things that popped into my mind looking at them. Wouldn't mind learning a little more in objectively evaluating the renderers.

View on HN

The color differences are from different "tone mapping" implementations after the render is stopped and can be ignored.

The point of that page is to evaluate renderer correctness (think of them as visual unit tests), not to really compare "which one looks better". For example, some of the images show appleseed with hard shadows, when they should be soft.

If you want to know which one is "right", Mitsuba is the one to look at as it's generally the most correct.

The graininess comes from not letting the machine renderer longer—the longer it goes, the less grainy it will be. That said, all of these renderers employ "tricks" to reduce the graininess that can make it practically impossible to remove whatever graininess slips through.

You can't say any of these is "more correct" from looking at the renders. Those hard shadows stem from a difference in light size, which is almost certainly a problem with the scene parameters being treated differently. The light size may have been "lost in translation", turned to 0, creating a perfectly sharp shadow.

One limitation of appleseed today is that our physically-based sun model is a purely directional light, not a far away disk with a finite radius. This explains the hard shadows on some of the scenes.

> And what's up with the graininess of all of them?

I am not a rendering expert either but IIRC "physical rendering" works by shooting photons across the scene based on a random distribution, and in the end, the finished rendering is an average version of all photon passes. If you don't shoot enough photons, the averaged out version will be noisy because there will be some hotspots were significantly more photons hit, relative to the total photon count. More photons = longer rendering time, so the graininess is in the end probably caused by a time constraint enforced for each of the images.

Typically, it's not using photons but the other way round, paths are started from the camera primarily.

The noise stems from the simple fact that those renderers try to solve an infinte dimensional integral (all light reflected by all surfaces) with stochastic Monte Carlo methods, the most popular being path tracing. There are other methods (finite element radiosity for example) to simulate light transport that do not exhibit stochastic noise, but those have fallen out of favor.

Without knowing the settings which were used by each renderer (and thus stuff like the BSDF materials, sampling techniques, random number generator sequences, etc, etc), it's very difficult to say.

All these things play a crucial part in how efficiently a renderer can produce an image with as little noise as possible.

But it's also possible to look through the noise (ignore it), and look for issues - i.e. the Classroom scene is very odd as Appleseed doesn't have soft shadows (what light type was used? PhysicalSky/Sun? HDR IBL?), and the Tungsten illumination from the environment is less warm.

The classroom is lit by a physical sun+sky combination. A limitation of our current implementation of physical sun is that the sun is a directional light instead of being a far away disk of finite radius. This is why we don't have soft shadows in this scene.

In theory, for "unbiased path tracers", the results of all three should be the same, at infinite samples. Rendering for a finite amount of time, lower noise suggests the renderer is more efficient.

In practice, every renderer has different features and settings, so the actual scenes will only be approximately the same.

Reply on news.ycombinator.com