You’ve likely bemoaned the CPU-only limitation on the cycles SSS node as well as I have. It’s so slow. I have a complicated scene that I’d like to GPU render, but an SSS character occupies about 10% of the pixels.
So I’d like to render just the SSS portion on the CPU and composite it in to the remainder rendered on the GPU. But if I isolate just the SSS material in its own scene, it has no contexual lighting, reflections. For that, I’d like the rest of the scene to be there-- but not waste CPU cycles rendering to the camera. Given that Cycles originates trace rays from the camera (backward from reality) and computes the outcome it encounters objects in the scene in an ever-spreading graph, it seems this should already exist somewhere. Like, just canceling the render immediately after it finishes the patches I care about, but smart.
Is there a way to do that?
Thanks in advance
The newest trunk builds should now have the ability to render out SSS passes, so you can just save the results you get from the CPU render and composite them into the GPU-rendered scene.
Not exactly sure if this is what you want though.
Cool. That helps to isolate the material for later compositing, but wouldn’t I still be forced to chose between wasting time CPU-rendering the rest of the scene or doing without the environmental context on my SSS geometries?
You would still need to CPU-render the scene if you want to obtain the full environmental context for the SSS passes, though you might be able to get away with removing parts of the scene and paring the effects/lighting by shifting objects to another layer, reducing shading complexity, or decreasing the bounces.
If the SSS is only part in part of the scene, you can also speed things up by doing a border render, the border render should still contain all the pass information and shouldn’t affect the actual SSS like what’s done to an extent in BI.
I’ve gotten halfway there.
I’ve been experimenting with placing objects in mask render layers, which evidently kill the camera ray where present, thus saving tons of time on the cpu render. And I put a mirrored ball in the scene, and, indeed, it reflects the whole environment. So I’m getting the benefit of the surroundings without spending time doing throw-away computations.
So, there’s this, a plane mask with the rectangular cutout in the middle and a mirrored ball on top of the character, which rendered many times faster than the other attached image, the original work in progress.
Or, there’s not this. For some reason, it’s pretending I didn’t just upload my artwork. Nothing uploaded. Nice little counting up animation, though. Do I have to disable cookies? Are we still in the 90’s? Ugh.
In any case, I’d still prefer if it computed the masking shape on its own. It would be easy for it to do. I expect it does, somewhere.
Actually I’ve done exactly this before:
What you do is put the character in it’s own scene layer and then make a new render layer and select that scene layer to be the only layer that renders in that render layer.
1.) Select the character and hit M and pick the second layer to move it to. everything else not SSS related should remain in layer one.
2.) Then go to the render layers panel and hit the plus button to add a new render layer.
3.) Underneath that you’ll see four sets of layers, “Scene”, Layer", “Exclude” and “Mask Layer”.
4.) In the layer set listed as “Layer”, turn off all the other layers except for layer two (the character layer).
5.) If the character passes behind anything in the scene (whatever is still in layer one), you’ll need to enable layer one as a “Mask Layer”.
6.) Now select the first render layer and turn off all the other layers except for layer one.
7.) Last step is to make sure layer one and two are selected in the “Scene:” layer set for both render layers. This makes it so, even if it’s not rendered, it will still contribute to the ray interaction like shadows, reflections and GI. It’s really the secret to making this work. 
If you want to enable SSS passes in your character pass, go ahead but it’s not really needed.
All that’s left to do now is render the character layer on the CPU and the background layer on the GPU. You can do that by hitting the little push pin icon with the right layer selected and then go to the main render panel and select CPU or GPU. Depending on which layer you have selected.
EDIT: PS. Don’t forget to set your film to “Transparent”
Splendid. You’re da best!
I think this tidbit would help a lot of people… perhaps almost as much as the tidbit about the compositor not being a bygone relic of the past, but rather a valuable tool in achieving quality output.
For the benefit of the conversation, here are references to those images I mentioned. First the scene (kitchen setup compliments of Jay.Artist ), followed by the scene rendered through a peephole to a mirror ball. Again, the later finished, like, 20 times faster.
Awesome! I l love the huge grin she has on her face. I’ll record a tutorial on how to do this at some point.
Thanks! I called that shape key “badass gufgraw”. The scene is entitled “pushing back into a gufgaw at viewer’s cut finger whilst chair snags cord dangling from the cubbord, dragging attached toaster and shelf of pots onto said gufgawer’s head.” Obviously, not done… but I found a better use for the scene. The lighting is too nice to blow on slapstick.
I grew up on Norman Rockwell’s photo-realistic painting style coupled with his tendency toward impossible, over-the-top cartoon exaggeration on the expressions. Like in this work:
Oh wait, that’s a photo of Norman Rockwell.
In any case, the contradiction is visually interesting. It’s also an attempt to stay on the this side of the uncanny valley… though my family says the result is horrific-- which works too.
