I met up with a few friends recently at a Flame meeting, and one thing that came up was that quite a few people are still happily working on older versions of Flame. Another one was thad it. would be good to have it working with new Flame versions and GPU’s.
That got me thinking that it might be useful to package the new ML Timewarp, FluidMorph and new Reference Warp tools as a standalone PyBox. So I’ve put together a first release that should run directly in Batch on both fairly old setups and current systems, have additional controls for initial scale and bi-directional processing as well as model confidence output helpful for spotting.potential artefacts.
There are currently several NVIDIA/CUDA builds:
-
cu118 — aimed at older systems
-
cu126 — including Pascal cards such as the P6000
-
cu128 — for newer GPUs; this build doesn’t support Pascal
-
cu130 — for Blackwell-generation GPUs
Cuda has minor version compatibility so for example 12.2 should be fine with both cu126 and cu128.
For ML Timewarp, I’ve attached a couple of images on Github showing the Batch setup. Flame’s native Timewarp should be in Timing mode, with a second copy of the source fed through a Mux shifted by one frame. The Timing channel from Flame’s Timewarp is then linked to the ML Timewarp PyBox.
Reference Warp is a slightly different tool. Given an image and a reference, it produces forward and backward UV maps describing the warp between them. The forward map can be useful for transferring/copying motion from one image to another, while having both directions makes it useful for stabilize / paint / destabilize workflows.
The Reference Warp model is still somewhat under-trained, so consider this an early version, but it may already be useful for some shots. I’ll update the weights as training progresses.
Mac versions are coming soon.
if you’d like to support this work: https://buymeacoffee.com/talosh
P.S. I have not tested all cuda versions yet so please let me know if there’s any issues with builds.
