CPU/GPU render different outcomes



  • @piersyf Yeah, We've discussed this before. I think one of the threads about render speed (maybe the one I started). We had a standard scene made by @Raven and some folks noticed that not all the lights would be reflected in windows and other surfaces. I think it comes down to BPT being off in GPU that turns off "sample all lights" etc. We also concluded that if BPT is on in GPU that you will get a nasty crash when the GPU starts to render transmapped hair. So gotta run GPU with that off.



  • @ghostship said in CPU/GPU render different outcomes:
    that if BPT is on in GPU that you will get a nasty crash when the GPU starts to render transmapped hair. So gotta run GPU with that off.
    If it's you display GPU. Not if it's CUDA only GPU, at least on Win7. As Win10 reserves memory on all GPU it can be different there too.



  • Will experiment later, but for me BPT and PT gives very similar results, in PT SSS and diffuse clears first, glossy second and vice versa in BPT. Reflections fully formed near 50 samples in PT and clear up to 100 and more.

    Re starting scene: how light fixtures are made?



  • Same settings as before;

    0_1486857429029_hallway bpt.jpg

    on the left, CPU render, BPT on. On the right, GPU render, BPT on.

    0_1486857531366_hallway BPTX.jpg

    on the left, CPU render, BPT off. On the right, GPU render, GPU off.

    Rendering without BPT is killing specularity and reducing light intensity. To go back to my original post, I haven't noticed this degree of difference before. It has occurred after the last SR, but also I've had several driver updates for the GTX970.

    The lights used are 3 spot lights. Area lights give entirely the wrong coverage for this scene.



  • @phdubrov Good to know! Thanks! I'll look into that as I was thinking of building a new computer with two video cards for Poser renders.



  • @piersyf SM claimed changes in glossy in last SR's readme.
    But is seems something scene/materials/lights dependent, will post some images later.



  • test scene https://www.dropbox.com/s/5660jcast1iwsp3/specular test.pz3?dl=0

    result BPT on left, PT on right
    0_1486883841779_2017-02-12 10_13_55-specular test - Smith Micro Poser Pro  (64-bit).png



  • I got the same results as you regarding the test scene. I also don't get nearly the same difference in my normal character creation scene, so yes, it appears to be scene specific. That doesn't answer why, though. If the light settings and material settings are identical, the only difference being whether BPT is on or not, I should not be getting such differing results. I'd like to know why I am.



  • @piersyf have you tried deleting all but one light and seeing if there is a difference then? I was under the impression it had to do with sample all lights etc. so maybe it has an issue when there are multiple light sources.



  • Not a single light, no, but as few as 3. If Poser can't handle specularity with 3 lights, it's fundamentally broken.



  • Not the numbers of lights issue - I add 7 spotlights in test scene, same results. Are lights behind the glass?



  • No. Free and clear of the ceiling. It also makes no difference whether spot, point or area light.



  • Materials?
    Poser, Physical or Cycles roots? If poser, is AltSpec involved?



  • It's a house, dude... 40 or more textures, so mixed root. Mostly Poser. I don't use alt spec. It does the same thing to people, so normal maps and EZskin is effected the same way.

    Sorry, I know you're trying to help. Just getting frustrated.



  • Probably it's down to one or more firefly nodes or PoS inputs broken in PT or BPT. So I try to figure which.
    Wall and floor mats are first to check.



  • I ran several tests comparing CPU and GPU renders with and without BPT, always using the same settings (Windows 7, nvidia GTX 1060). As far as I can see, there was no visible difference.

    I really do prefer the GPU render if I can leave my computer for a while because it is so much faster than CPU (approximately 10 times). The downside is that the UI is totally laggy while rendering, so I can hardly do anything else on the PC, not even watch a movie.
    I like rendering with BPT even better because it is so much faster, adds less noise and has better specularity on many materials (as already mentioned by @piersyf). So the fastest combination is definitely GPU & BPT. Unfortunately, transmapped hair reliably crashes the driver when using this dream team.

    Here is my solution until Poser and nvidia can get this to work properly:

    1. I make the hair figures/objects invisible and render the whole scene with GPU and BPT.
    2. Then, using the same settings, I switch to CPU, make the hair visible again and use the area render to render only these parts of the image that have hair in it. You have to think of hair shadows and reflections, but it is usually only a small fraction of the whole picture.
    3. Finally I combine both pictures in Gimp or something.

    Sounds like a lot of effort? But I can render a screen size image in 45 mins instead of 8 hours this way! And I get the specularity that I like, sampling the direct and indirect lights. No matter how many samples you throw at this in GPU rendering, you will never achieve a comparable gloss.



  • In my case, the floor is a physical surface (root node), so I CAN increase specularity by reducing roughness, but even at 0.01 it is still less shiny than using BPT. It still doesn't explain why it effects skin on characters as well, in this particular scene.

    I agree that GPU is generally faster, but I'm using a GTX970; to run GPU and BPT I need to reduce bucket size to 32 or it crashes (I haven't done the coding fix on timing).



  • Found it! Physical surface is broken. BPT on top, PT on bottom. Look for lights reflections on top of the ball.
    1_1487061519762_pys-bpt.png 0_1487061519761_pys-pt.png

    Pure Cycles surface is like BPT in both modes.
    Setup
    0_1487061657634_setup.png



  • Nice... worth an experiment. I'll have a look at doing the material using a different node set up.
    Even if it pans out, it doesn't explain why figures (people) render 'flatter' than with BPT...
    As eager as I am for an answer to this, my study is being torn apart tomorrow and a new door fitted (to the outside; through a wall). I'll get back later...



  • @piersyf if you use EzSkin there are good chances that something in PoserSurface broken same way as PhysicalSurface.