I have to track a logo on some ai shots containing people drinking a energy drink. There is a lot of movement and swirling liquid in the bottles. I can’t figure out how to track this. I’m doing a manual track and it’s floating ofcourse.
Is there any other way to do it?
I remember there was machine learning tool in flame that you feed a single frame reference to and it would replicate that on the whole shot.
Having an option to use Syntheyes or similar match moving package so maybe so you can track geo. Or KeenTools, for Blender perhaps for object tracking with the help of geometry.
Otherwise, every shot is differnt off course. Is there something in the similar space, like hand holding a bottle that you can track and go from there? Mocha you mentioned has great adjust track module and now with point ML trackers so that is another option or you can manually try to adjust it on reference frames.
Hard to tell since every footage or in this case AI stuff is going to be differnt. Reflections, transparency, occlusions etc. Speaking of occlusions, make sure you make some occlusion masks for tracking and see if there is any offset tracking that can be done, Like hand holding a bottle or something firmly in the same place as bottle.
Make sure you take advantage of the adjust track in mocha, it should help with any slipping and use mesh tracking since its not a plane, but you can extract the data to a plane in latest Mocha. If you can track hand that is holding the bottle firmly, that is half way there. Not sure if it would work to solve the scene and export alembic and use that to project the label. But something to try.
This might be a job for AI for a quick fix… Get a good still frame painted up, feed it to (probably) Kling o3 Edit Video. I’ve done some proof-of-concept tests swapping out soda cans, and it was pretty decent.
Based on conversations with people who’ve worked on more ML generated images than I have, tracking in the “traditional” sense will likely be harder than normal since there are no optical rules governing the image.
My advice is start with a 1pt track to stabilize the image, and keep doing more tracks/adjustments in downstream actions until you’ve got the image stable. Then copy all the axis nodes you made into a single action, chain them up and use that action to stabilize. Comp the label and then destabilize.
It’s easier to manually stabilize and image than it is to track it. Try to stabilize it in a way you can invert. I have a demo and a downloadable setup someplace that makes this easier.
But equally how do “clients” get so attached to image sequences? (We can fall down the rabbit holes of absence of imagination/skill/etc)
It’s not as if they staged and shot them.(see above)
Why not adjust the prompt, supply some things, and get an acceptable end result.
Then upscale because the idea was so immaculate.
clients
Can’t live with ‘em
Can’t burn those clowns to make electricity to power ai spewing machines in a data center to make pictures that nobody will EVER care about for a product that NOBODY will remember…
That’s exactly what I suggested to them @aaronneitz they finally did it in ai, the problem was ai was not generating the logo according to brand guidelines.
I imagined that was possible @philm , atleast if there were trackers that could get covered by the logo, it would have been faster and possible in such little time
