direct light sampling

[Animator89]

Hello!
It would be nice to implement the coefficient of distribution of samples between direct and indirect light. This is in order, so in the scenes where a lot of direct light to remove noise from it. I hope it is possible to implement :)
Thank you!
Sorry for my bad English.

[Ondra]

this is exactly what lights multiplier does.

[Animator89]

I have almost all the pictures there are noises from direct light.
What should be the multiplier for complex scenes, fully Illuminability direct light?
I set value of 4 but the effect is not visible
thanks

[maru]

You should experiment with this value.

witX.jpg - where X is the amount of light multiplier. PT+HD

wit32_pt.jpg - 32 light samples, pt+pt only.

[Ondra]

If you cannot get rid of the noise, then the problem is either in too difficult scene, or there is a bug in Corona. Post pictures (result with time and settings, and scene overview) to be sure.

[Animator89]

Example scene. It is not the whole illuminated by direct light but the noise is very noticeable on the wall.
In this scene, combined kitchen, staircase, living room, hallway of 18 million polygons and about 110 lights.
80% of the materials in the scene are composite materials with AO texture blending
weight 2 GB scene without textures
rendertime 4 hr. Pathtracing with HD cache. Light mult 4
i7 950 20Gb ram
thanks

[lacilaci]

Nice scene :) btw try 16-32 for light multiplier (2 vs 4 is not that big difference in some complex scenes)

[maru]

Show render settings. Especially pt samples and HD settings.

Maybe it's because of AO?

Or maybe it's because of translucent lamps?

[Animator89]

I do not think that this is because of the transparent tubes so that the noise can be seen in places where exposed to direct light from the lamp. Maybe it's because of the small light sources and in this case would help bidir. Bidir but does not work with blend materials and HD cache. In any case, any corona faster renders with whom I worked on the calculations of brute force. :) for this scene maxwell give me similar result in 30-35 h.

with 32 samples i feel real difference!
Thanks to all responded for the help!

[Ondra]

yep, this is a hard scene to render ;). But since it is only direct illumination, you can get away with brute force (extremely high light samples multiplier).

[Javadevil]

I do remember a time when Corona lights would take longer to render than using geometry/material lights.
Try a low poly sphere/material light, see how that goes.

cheers

[Polymax]

I think, in the meshlight(mesh/material light) number of lights depends on the subdivision mesh.
In my scene, which is exhibited in the gallery here, I used only hi-poly meshlight and render info shows me that I have about 400 thousand light sources, and the render time was 1 hour 20 minutes per frame, I used a 10 light samples.
With a corona light, I had a lot of noise and don't use it.

[Animator89]

You are right. Mesh light actually faster than the corona light object. I thought that this is the same but with a more convenient parameter setting. With a mesh lights noise is indeed less. Thanks for help!
Sorry for my bad English

[Ondra]

You are right. Mesh light actually faster than the corona light object. I thought that this is the same but with a more convenient parameter setting. With a mesh lights noise is indeed less. Thanks for help!
Sorry for my bad English

I disagree. For planar lights it is the same, and spherical lights are MUCH better as Corona light than actual geometry in the general case. The sole reason that it seems better in some particular settings is that geometry light is interpreted as a separate light for each triangle. So you end up with 320 lights instead of single one. If you have 100 other lights, you end up having higher probability of sampling one of the 320 lights from 420 lights total in light-as-geometry case, than the probability of sampling 1 light out 101 (it is actually only about 50% higher).

This just means, that because there is some defensive sampling involved, you can give higher importance to one light by turning it to geometry. It will render cleaner, but the rest of the scene will be worse. I don't like this solution, but I understand that the sampling in big scenes (>15 lights) is far from optimal. It would be much better to simply have some "visual importance" slider for each light that you can adjust for different shots, but I've been trying to avoid that. I'll try to improve it some other, automatic way first.

[iliyang]

The sole reason that it seems better in some particular settings is that geometry light is interpreted as a separate light for each triangle. So you end up with 320 lights instead of single one. If you have 100 other lights, you end up having higher probability of sampling one of the 320 lights from 420 lights total in light-as-geometry case, than the probability of sampling 1 light out 101 (it is actually only about 50% higher).

I didn't get the last thing in the parentheses, but from your reply it seems to follow that the probability for choosing a light source that has been defined as a mesh depends on the tessellation of the mesh? If so, this is certainly undesirable. You either need to have a sampling scheme that picks a point on any emitting surface proportionally to the total emitted power at that point, and/or have mesh light objects, and not break the meshes into individual light sources.