Autodesk Flame 2027 Update introduces two machine learning models for generating depth maps from indoor and outdoor shots. These models are available in the Selective object, where they can be used to isolate parts of an image; in the Z-Depth Map object, where they can be used to place objects within a scene or apply defocus effects; and as standalone nodes.
Please can we train these lovely tools to recognise trees. Just had to use Runway ML on about 120 shots to get a depth maps which were not always accurate. Rotoscoping trees in this day and age is upsetting. Loving your work though. Thank you.
Yeah I gotta say the model seems solid the couple times I’ve used it! Thank you! The Automatte node is solid too. What’s next?
I recently had a shot where I had to wet down a shirt (like it had been raining). But the talent was doing a 360º turn. The existing motion vectors system in Flame didn’t cut it and sort of fell apart where he turned. I ended up going with RAFT to generate the vectors (in Nuke). So maybe a better motion vectors model could be useful?
where did you use those vectors?
I used them in Nuke, feeding directly into the VectorDistort node swapping out the SmartVector node. There’s probably a way to adapt them for Flame but I didn’t tackle that given the usual time pressure. They’re a significant step up from the built in Nuke/Flame vectors.
Interesting. Haven’t worked with RAFT yet. There may be a way of importing them into Flame via the Montion Convert node.
There is a barebones ComfyUI node to generate them. But there are some gaps.
RAFT only does produce two u/v vectors in one direction. So to make it more solid, we’ll have to build something more solid that does both forward/backward and maybe confidence if Flame can use it. Seems like a custom ComfyNode project in the near future.
Here’s a comparison of the motion vector landscape with various options: Claude
Yeah I tried RIFE too, but that model’s sweet spot is fast good quality retiming I believe, using it for tracking wasn’t much better than the built in Nuke tool in my use case which was ‘edges rotating out of view’.
There’s an interesting point in that summary, it’s essentially the same data in retiming and tracking, so perhaps it’s possible to re-use the ML Timewarp data for tracking?
I’ve been working on an OFX that uses RAFT/SEA-RAFT as an alt to Flame’s MVs. It’s a pair of nodes, analysis and apply. They are fairly eponymous…
Old demo… it’s progressed a lot since. I’m also using a post-processed calculated failure as an alpha to mask back in areas which will fall over. Since then I’ve added multiple patch holds which fade over their hold ranges at defined intervals. Just need a week to wrap it up.
Always ahead of the game 
I wish I was a little bit taller…
…I wish I was a baller…
Trust me being tall isn’t all it’s cracked up to be
I wish you understood my sarcasm 
Shit translates poorly over text man
True that
I wish there was a girl on support
I would call her