OK, so working on getting a nice render of my go to test scene with filmic render. However, I’m seeing more noise in the render. As comparisons:
This is filmic @ 2,500 samples, no caustics and clamp indirect @ 1.0
This is standard sRGB @ 1,500 samples, reflective and refractive caustics both on and clamp indirect 0.5
Whilst the tone mapping with filmic is much better (and yes, it’s better generally with my postage stamp sized previews), the filmic render is noisier, and will probably need 4,000 - 5,000 samples to get similar.
Obviously, the lights have been adjusted, but in both cases I’m relying on HDRI and a single sun lamp. The segment shown is lit entirely by bounce light, but as an interior, a lot is lit with bounce light rather than directly. The adjustments to the lights are to intensity. Colour (black body on the sun, and out of the box with the HDRI) are the same.
First thing to remember: filmic blender is a grading tool, it is not part of the render. It’s applied AFTER the render, when saving files. So it’s not really causing more noise per-se, just making noise that was already there more visible.
All of that is nice academic knowledge, but it doesn’t really help the noise in your scene. You could try adjusting colors in the compositor to hide the grain. Or you could set clamp indirect back to 0.5, although that will darken your lighting a bit. Are you using portals for your HDRI sky?
I am using portals - on every entry point for the light. I let it finish to a full render overnight, and whilst there were a couple of issues (not with the render, but at some point I must have moved some chairs by accident because they were floating), the noise overall was less noticable.
I’m going to let it go in a progressive render today with an insane number of samples set so that it keeps going whilst I’m out at work, and see how it gets on.
Clamp is the same tradeoff it always was. You just have to adjust your values upward due to the higher white point. For example, if you clamped indirect at 0.5, you might want to clamp at something like 6-12 instead. And consider only clamping indirect - most things you want to clamp are indirect (fireflies), most things you DON’T want to clamp are direct (direct specular highlights, bright objects).
@Roken: if you have a lot of small openings, try one big portal stretching across them instead of separate portals. Don’t worry about a wall over part of your portal, it’s still better than evaluating a bunch of little portals. Or a couple of big portals, such as one portal for all the skylights on one slope of the roof, another for all the windows on one wall, etc.
Cycles’ rendering algorithm is not well suited for scenes that are mostly indirect lighting. If possible, consider interior lighting. (if you have a client who specifically asked for daylighting…welp)
@MZGarmi. I’m aware that you shouldn’t use clamping, but I’m of the same mindset as J_the_Ninja, it’s a necessary evil. Only indirect is clamped. Without it, it’s incredibly bad. I’d have to render this scene for days to get a decent result.
@J_the_Ninja. I didn’t think of trying one large (or in this case it would be four) portals rather than individual on each window. Next render I’ll try. If I can eliminate clamping at the same sample rate, I’ll be happy.
For info, caustics have been turned off, which I really didn’t want to do (the whiskey bottle and glass should have a caustic effect, but in all honesty, it’s barely made a difference), and there’s no glossy filtering.
I deliberately use this scene because of the technical difficulties. PBR materials and indoor without interior lights. Once I’m happy with the daylight render, I’ll move on to a night time render using the interior lights, but that should be much easier. Early Blender tests required the interior lights on, or worse, a bunch of fake lighting, but now I’m reaching the point where accurate lighting is starting to work.
I’m also trying to minimise post work. The latest render is straight out of the engine, and looks much better to me. Still a very high number of samples, and I estimate (didn’t check) 11.5 hours render time on GPU.
