It would be very very useful and doesn’t seem like it would be a giant stretch for the ML stuff. But I’m just a naive picture guy.
2 Likes
Based on basic search, the tools exist: AI Color Grading - Match AI | Color.io
I have not tried it. It will apparently will generate a LUT which you could import into Flame.
Questions remain on you operationalize that - if you had 3 clips with slightly changing conditions you would need 3 LUTs and the tool would have to have good consistency between multiple runs.
Shot matching has existed before, it’s not clear that ML has a particular advantage, since you actually have to train a neural network with various grades, vs. figuring out a classic algorithm that does it the non-ML way.
You could also try out this: https://revisionfx.com/products/rematch/flame/#features which does run as an OFX plugin on Flame.
I was at an Adobe event a couple of weeks ago, and apparently the new release of Premere has an ML color match tool. I haven’t played with it yet though.
Here’s a basic technique using the AFX_ReverseGrade to match two shots:
Here is the same match using AI Color Grading - Match AI | Color.io - use your own judgement. The video above provides the necessary context.
I’m waiting to fix an install problem on re:vision rematch to compare the same with that plugin.
4 Likes
That vimeo link doesn’t work correctly
Sorry about that, fixed now. Didn’t test it in a separate browser, and was logged in where I tested it 
2 Likes
Here’s a comparison of all the tools.
This should be prefaced, that there are two use cases - one where two cameras shot the same scene (either separate takes in the same scene or a multi-cam capture), and one where you have separate scenes of totally independent origin, and you’re trying to harmonize the look & feel.
This test is the latter (and harder). Also the Re:Vision plugin is optimized for the former, but can create some results for the latter.
As becomes obvious - there is no right or wrong answer, and it depends on what qualities in the image are considered important, and what becomes the flex.
The key things in the reference shot in my mind is the contrast range, and the more pastel blue sky with a bit of tint to purple.
The other shot is overlay contrasty and moody, also has purple tints everywhere (sky & and the railing).
The Re:vision Rematch captured the lower contrast feel and had a balance between the purple and the bluer skies. But blew the highlights in the clouds (there are some controls)
The AFX Reverse Grade has better highlights but they sky skews too purple still.
The color.io result has a very different palette. Much more cyan. The clouds are even more cyan than the reference, and the railing too. For me it has gone too far in that direction.
Also the color.io tool I believe is a one-shot take it or leave it. Whereas the other two tools have several controls to refine the result.
(ps: original post had two shots swapped, fixed image and text above)
Maybe over the weekend I’ll try to find some shots from the same scene or multi-cam and repeat the experiment with something that should match very closely.
(updated: here’s a pass at a manual match - just with scopes and one image node)
1 Like
What provoked this thread in me was I was reading up on this-
The folks behind this have a decidedly mixed reputation and a penchant for overselling based on numerous discussions on our colorist discord.
Everyone should do their own research. I wouldn’t expect a magic bullet.
3 Likes
If an ML or AI tool doesn’t exist by now, it’s probably because nobody thinks they can make any money on it.
1 Like
You can easily train nuke copycat to match grades.
Done this a bunch now for different occasions.
its like super easy for a AI to grab it if you have before and after.
I usually train it on the graded rec709 and simple rec709 lutted source that is as close as possible to final grade, then boom, takes like maybe 1 hour on a 4090 to get a decent result
I have trained it on a full commercial once where I had to put in a extra shot - and it worked super well.
Sounds like ti’s cheaper to hire a flame artist…
1 Like
I can totally see how that can match a grade from the same reel. Won’t have to deconstruct it, just rinse and repeat without knowing what knob was turned.
But does it also work between the A and B cam, if they weren’t painted correctly on set?
And does it work if the shot is from a different day, time of day, etc.?
Presumably somewhat for the former, and not really for the latter. But would love to be wrong.
as long as you have enough varriing inputs and output the AI does a pretty good job, but it always depends on the material, like if you suddenly give it a very blue-cast input to fo inference on it will crap out
So if you have a lot of different shots before-after it will do a very good transfer to B cams or even phone shots, stuff reshot later and much much more