10

This image is the result of the exact same file, one rendered on my computer and the other on sheep-it. The quality of lighting in the sheep-it version is superior, but I cannot understand what it actually is.

Rendered locally: Rendered locally

Rendered with sheep-it: Rendered with sheep-it

Animated gif of the above (6.8M download)

The local render is GPU (CUDA) but I tried a small part of it on the CPU (because it would take weeks to try rendering it all) and the lighting is the same. I have also tried Windows on the same computer and a Mac.

For a specific difference look at this detail. There is distinct difference in the shadows, as is on the upper bar shelf.

Detail

Reducing the intensity of the overhead lights does introduce those shadows but changes the overall image in other ways. Plus changing the value of one node in one material doesn't make sense. It's probably some global variable that is different.

Blender version is 2.78a downloaded from blender.org or the Thomas Schiex ppa (same result).

qwazix
  • 1,885
  • 11
  • 24
  • I'm thinking colour management - maybe different gamma? Look in the Scene properties under colour management. What settings do you have there? Do you have a non-standard device like having the filmic addon installed? Could the sheep-it blender not have OCIO in the build? – sambler Dec 25 '16 at 03:27
  • I do. And that's what I thought at first too, so I removed it. The images above are without filmic (see the saturated spotlight cone over the counter — with filmic it becomes almost white). I also tried with all the different settings and it just becomes lighter-darker, no difference in the shadows. Exposure is 0 and Gamma is 1 in the file. – qwazix Dec 25 '16 at 10:35
  • To me it looks like sheep-it does more or more accurate indirect bounces. – VSB Jan 27 '17 at 04:03
  • I'll try tweaking that and report back. – qwazix Jan 27 '17 at 08:08

2 Answers2

3

Doing a difference of the two images is a little more revealing: OP's renders subtracted from each other

What do you notice? For one, the light fixtures themselves are black. And there are no reflections of lights, only of other objects.Further, the direct shine from a light is seen through the glass (center right), but nowhere else. My guess is something in the sheep-it version has more indirect light. I'm not sure WHY this is the case, (I would expect it to render the file as is) but it may be using more diffuse/glossy bounces, or have increased/disabled the threshold for the "clamp indirect" setting.

JtheNinja
  • 9,122
  • 1
  • 31
  • 49
  • If not overall indirect, is it plausible it is a different transfer function being applied? The highest emissions look identical, yet the darker regions might be lifted? – troy_s Dec 24 '16 at 18:20
  • @JtheNinja Hm... testing now. At first impression it seems you are correct. – qwazix Dec 24 '16 at 18:29
  • @troy_s I have a curve under color management but I also tried without and the image still isn't the same. I don't think any amount of curves can change the relationship of the shadows as it happens in the detail above. – qwazix Dec 24 '16 at 18:29
  • 1
    @qwazix Curves will influence the relationship of the shadow. The reason is that the display referred transform will bend the ratios more or less dramatically based on whatever transform is in use. In this case, I suspect JTheNinja may well be closer to the solution. There are a few different contexts here though. – troy_s Dec 24 '16 at 19:10
  • @qwazix How are you encoding the render on Sheepit? – troy_s Dec 24 '16 at 19:34
  • @qwazix try rendering as EXR and see if there are any differences. –  Dec 24 '16 at 20:57
  • Updates: the indirect clamp doesn't seem to help. The direct clamp however does, but from a first try, it messes up with the spotlight cones, I'll try to tweak the values. – qwazix Dec 25 '16 at 01:18
  • I have the .blend file to render to jpg, however sheep-it spits out a .png. I don't seem to recall any setting on the site for the output, but I can look again. If I set the blend file to exr will it honour it? Btw, even if I manage to make sheepit render the same file as locally, I'm still interested in replicating the sheepit result because I find it more realistic. – qwazix Dec 25 '16 at 01:22
  • @qwazix Better to know what the hell it is doing wrong. At the same time, it would be wise to pinpoint exactly what you like about the result and arrive there properly. As it stands currently, it is a bogus result. – troy_s Dec 25 '16 at 03:44
  • Ok. I'm waiting for the clamped render to finish and will upload to sheep-it with exr output. What I like I can only describe as more AO (well not exactly), but the more complex shadows, the less gloss on the back of the office chairs, the less diffused light on the ceiling. I thought (hoped) maybe the sheep it version was the right one but if the consensus here is that the local is the right one I can work towards that goal by tweaking the lights individually. (The first is the local by the way and the second one the sheepit version, just so that we're sure we are talking the same thing) – qwazix Dec 25 '16 at 10:54
  • This is locally rendered with clamp direct and clamp indirect both set to 5.0, and this is it's difference with the sheep-it version. Sheep it ignores the output option and always renders in png (however this time the image is 16bit (46mb instead of 12.7mb). I didn't even know 16bit png's were a thing. – qwazix Dec 26 '16 at 11:13
  • Dang, this is really weird. You should definitely report it. I've been dealing with some glitches on Sheep It also, but it's an amazing tool. – Anson Savage Dec 31 '16 at 04:05
2

Found it!

Sheep it! instead of splitting a frame to parts and stitching, it splits the sample count and averages. And as @YOLOonline said here averaging different sample seeds isn't equivalent to just rendering with the total samples.

Kinda proof: I rendered 157 equal frames at 20% resolution and averaged them. I resized both the sheep-it image and the locally rendered image to 20% with gimp and here are the differences with the averaged frame.

sheepit-averagemine-average

The sheepit - averaged image is almost completely black except near edges (which are artifacts of resizing) while the local-averaged image looks like the local-sheepit image JtheNinja posted

I believe that the edges will not be visible if I render the 156 frames in full resolution but I don't intend to waste that much processing power just to make sure.

Now the question becomes, why?. I thought that cycles sampling was like dice roll, and the previous result didn't influence the next.

Community
  • 1
qwazix
  • 1,885
  • 11
  • 24