Reproducible Images?

[PabloMack]

I just built a render machine and have been using it to speed up my sequences. Unlike some other TG users, I am unlikely to render very high resolution stills. Instead, animation and video is the ultimate product of my work. One of the first things I did was to benchmark my two systems to see how much faster they can produce a sequence when working in tandem. Each of the two systems produced ostensibly "the same" frame from the same unchanged scene and data files. But when I flip between the two frames that were generated by the two different systems, they are not identical as I can see small differences in the vegetation. This is a little disturbing as I was hoping that the frames would be exactly reproducible given the very same data to start with. This makes me wonder whether there is any pseudo-random numbers that are generated independently from just the data in the scene and other source files. Has anyone else experienced this phenomenon?

[Oshyan]

Yes, there are random numbers involved, but the differences will be more severe at lower detail settings. Use of GI Caching is also essential to ensure the most similar output.

- Oshyan

[Kadri]

Yes, there are random numbers involved, but the differences will be more severe at lower detail settings. Use of GI Caching is also essential to ensure the most similar output.
- Oshyan

just curious, no way to make them exactly Oshyan?

[jo]

Hi,

It is certainly possible to have images/frames consistent. After all TG is used in animation for major motion pictures and broadcast work amongst other things.

As Oshyan says the biggest difference is usually in GI, and that's where GI caching comes in.

Regards,

Jo

[Kadri]

I thought the same until now. I was just curious if it is my English and asked to be sure. Thanks Jo

[PabloMack]

I had started a long HD render using GI caching on my two render nodes for the first time before Oshyan and Jo posted their replies. I guess there is more than one strategy to use this technique. Any experiences/opinions/corrections are welcome. One strategy I think would be to generate GI lighting on every nth frame and then interpolate. This would slow down flicker so that it might become unnoticable. The averaging on every frame might accomplish the same thing except it requires more caching frame storage. I am doing a combination of the two. It would seem to me that skipping too many frames might leave some parts of the frame poorly lit because the camera might move too far away from corners that are not represented with nearby cached lighting points. However, I presume that averaging involves looking ahead as well as looking behind temporally proximal cached frames(?) [...if you catch my drift].

[Tangled-Universe]

What you basically  "made up" is the way it basically works.

[PabloMack]

What you basically  "made up" is the way it basically works.

Thank you. It is always reassuring to I feel like I understand what is happening.

[PabloMack]

It is certainly possible to have images/frames consistent. After all TG is used in animation for major motion pictures and broadcast work amongst other things.
As Oshyan says the biggest difference is usually in GI, and that's where GI caching comes in.

I just rendered the same frame of the same scene on three different systems using the same GI caching files and the vegetation looks very noticably different on all of them. There is a hill in the background and it goes from light to medium to dark. In these tests the vegetation in the foreground looks consistent, though. But this should not be happening. In my mind, the images should be identical. I have all content on one of the systems on drive "E". On the other two systems, I have mapped a network drive as "E". There is no other place for the two remote systems to get any resources for the scene. They all start with the exact same data. It makes sense that the frame number might be used as some sort of seed for internal pseudo-random number generators. But on the same frame with all the same data, the images should be identical, especially considering that they are using the very same GI caching files.

[Oshyan]

Can you post the 3 images you came up with? Are you using very low detail?

- Oshyan

[PabloMack]

The forum software complained that they were too big so I cut the dimensions in half and resaved them as jpegs. The differences don't look so obvious now but when you flip through them the lighting on the ferns in the center of the picture shimmers. It is the mountain in the far left background where the lighting is extremely different:

Phenom II x4 955 [attachmini=1]
A10-7850K APU [attachmini=2]
FX-8350 x8 [attachmini=3]

[Matt]

It looks like GI differences, or possibly cloud shadows, but I can't explain why it would do that. Can you send us a clip with your render node, or copy and paste the render node into your reply?

Matt

[Oshyan]

Looks kind of like cloud shadow differences. Are you using any of the acceleration options, Voxel Buffer or the Acceleration Cache?

- Oshyan