Implement an overscan support for tiles, so that adaptive sampling can
rely on the pixels neighbourhood.
Details
Details
Diff Detail
Diff Detail
- Repository
- rB Blender
Event Timeline
Comment Actions
Personally I think we should disable intermediate denoising of tiles. If you render without tiles, denoising also does not happen until the end. I don't think it's worth the performance cost.
| intern/cycles/render/tile.cpp | ||
|---|---|---|
| 381 | I think this should be lower until there is really a need for it. I suggest using 4 instead. | |
Comment Actions
Lower overscan to 4 pixels.
Personally I think we should disable intermediate denoising of tiles. If you render without tiles, denoising also does not happen until the end. I don't think it's worth the performance cost.
For the command line rendering I agree we should disable denoising of tiles.
For the rendering from the interface I am not so sure yet what's the right thing to do. It feels like it would be nice to treat F12 render similar to the viewport, so that artist can stop rendering once the denoised result feels good.
Perhaps its time to have a quick design session and figure out answers in this field :)
Comment Actions
This looks good now, however the patch needs to be updated after the bugfix in rBb314d3e7877f: Fix T91639: Cycles crash rendering high resolution images with multiple passes.
Comment Actions
Rebased on top of latest master branch.
I've also disabled overscan with denoising, in the anticipation of D12662.
Had a second tough, and don't think we'll have enough time to design and implement the workflow I've described above. Until then lets prefer better render performance. Also noticed that 4 pixels overscan is not enough at a very low number of samples.