Automatic transform matching in Baselight - cool!
7 Likes
@finnjaeger
yep - this tool is really nice and exceptionally fast.
you might want to examine https://www.kornia.org/ and applying the results to a 2d transform timelinefx
[cracks knuckles and opens Claude code….]
13 Likes
i wish adsk would do the same and just use ai coding assistants to fix flame 
1 Like
I would take even AI agents just to test Flame, as that part seems to have stopped awhile ago.
1 Like
much easier to have people pay to be beta testers its just smart business.
I also let my clients QC our comps - they love it.
/sarcasm warning
2 Likes
It makes bean counters happy. But it really messes with your coverage. You get good coverage in the popular minimum viable feature set and then it flattens and goes to zero in the complex edge cases. Not a recipe for success.
beta testing is good for feedback and uncovering blindspots, not a replacement for the basics.
subscription pricing seems to be cancerous to software development, especially when there is no competition
1 Like
Watching this video reminded me of the fact that, long ago, a junior NUKE compositor believed the MatchGrade node was a tool for transform matching…
2 Likes
Talk about a quality of life improvement! Yes please!
Wow. Yeah you’d think this would be at the top of the planned features list. How would you even go about developing a script for this? Does it start with a right-click of the clip in the timeline? Even a single frame matcher would be immensely helpful.
I have been doing exactly this, a simple frames 2D match, scale, rotation and translation but I find the real help would be to also match the time so I will finalise it and then port it to Flame.
Just as a few questions to you all;
- Is it ok for you guys if I use numpy, scipy, openCV libraries? (clearly will add extra memory requirements but I just want to double check.
- How much would you be happy to pay for it? I am not intending to add yet another bloody license model so I would rather have a friendly, a coffee a month kind of cost done without fuss, those who can’t pay it, well, so be it.
Is this something you guys would like? I am keen in making it super fast so a normal feature based approach of FFT based approaches tend to be very slow at 4K, therefore I have a cool solution that, once I can trust 100%, I will start tidying it up.
Let me know please.
@jordibares - since numpy, scipy and opencv wobble between BSD 3 and Apache Licenses and are fully open to commercialization there should be no objection from anyone to charging for your development time and expertise, but perhaps you may find more fulfillment through open-sourcing in such a way that people may contribute back and improve the code (perhaps a switch to GPLv3 for your wrapper?)
two useful tools are:
and
ORB and AKAZE on a 4K frame are fast—like well under a second per frame pair on a modern CPU, and most flames likely have decent cpu bound performance. The memory footprint is the bigger concern than compute time, but with a thread pool you control how many frames are in memory simultaneously and in the case of a first frame or a first frame and last frame analysis we’re not talking about anything overly batshit. Every n-frames might be more of an issue though.
The transform step is literally nothing if you’re using action—you avoid resize and 2d transform and go straight into native pixel coordinates the whole way, so what comes out of the homography decomposition goes directly onto the keyframes with no correction…
I’ve got a basic framework going just need a little more time to fuck about.
5 Likes
After a bit of mucking about and some false starts I’m seeing some good results now using SIFT keypoint detection with three modes:
- Similarity—uniform scale + rotation + translation
- Affine—non-uniform scale + rotation + shear + translation
- Homography—full perspective
It’s a hook, runs local and is pretty light weight. Anyone wants to test hit me up with your git username. If it works for enough folks I’ll flip the repo.
18 Likes
It’s another beautiful solution brother.
Keep ‘em coming.
Now add timewarps too. JK lol. (Well, it would be cool though)
1 Like
@cnoellert - so excellent…
so useful for connected conform and normal aspect ratio / social variants.
huge deal.
1 Like
Cheers man
1 Like
@cnoellert - keep doing the good work brother - this is a very sophisticated and useful tool. i appreciate you.
1 Like