Resolve doesn’t provide a line on the wipe, but both screenshots are wipes between a gallery image of the grade, and the original with LUT exported by Resolve.
In fairness Resolve doesn’t analyze the shot. It has a record of the operators and can generate a LUT representing the operations. That is a lot more accurate and easier to do.
There are more sophisticated LUT tools, I think Lattice is one, though I recall another one just can’t locate it. They support larger samples and more controls.
There’s a new algo. Try similar settings to what I attached. I haven’t checked if the .cube export is right yet, but just see if this captures the look better in the preview.
the match grade node is only in NukeX FYI. To clarify unless there is some AI magic under the hood I don’t expect the result to be pixel perfect. Just needs to be close and be able to generate .cube file. We often run into this when we are working ungraded and client goes to do a look set with an out of house colorist, and that colorist doesn’t want to build LUTs for us They will eventually return to do a final grade, it’s just that after the clients sit with the colorist they expect all the reviews to be in color look from that moment on. There are other use cases but that’s the biggest IMHO
I would tend to agree. Because if it’s not a close match (as in my 2nd test case) you still have to go through and figure out what’s off.
Rather than a grade node (or maybe in addition to), what would be great is if the AI model could do the analysis and come back with something like this:
To match the two images, please use these grade operations:
Exposure +15%
Gamma: -5%
Overall Saturation: -5%
Reduce Saturation in Reds by 10%
Move the Green Hue slightly towards yellow.
Or something like that. Get to the essence of the match. Because (a) you learn more about what the difference is, it’s easy to translate into various tools, and then you know which of these you need to tweak further after looking at the result.
I know this is more for individual shots, not a batch of 500. But it would be way more useful. I might see what Claude can come up with if I task it with comparing two shot.
As to the use case of @TimC described - that makes sense. My only caution would be, if the magic LUT misses one thing but has the rest right, does your client then start going ‘you missed this color note we had in the session’? Like it’s too good and yet not good enough, and thus keeping it dumber on purpose may be safer.
Compared to applying the basic Camera colorspace IDT, and maybe restricting yourself to basic CDL operations. That way it’s more obvious that this is not a grade, but just making the footage look a bit more like what it looked like on set.
Maybe that’s what your aiming for already. And maybe that’s all the Matchgrade node is capable off anyway as we saw.
Step 2 in the Claude experiment, I asked Claude to build me a LUT with everything it suggested. At first it hesitated, suggesting I do the LUT myself, but then offered to write python script to generate the LUT and actually run it, so I had a downloadable LUT file.
Result:
Not perfect, but more nuanced than the Nuke Matchgrade node. Mostly underpowered the blue/red hue rotation. For what it’s worth…
Original / Color Target / Claude Version (w/ LUT Claude generated)
Some colorists push around stuff in the primaries for a lookset which is an easy match but some of them will get into the weeds with secondaries and keys, windows and what not which is more work to bake down to a LUT so they wont do it. also more difficult for a tool to do a precise match in comp but my experience has been clients don’t balk if it’s in the ballpark and they know it’s gonna go for the proper color grade later where they can really noodle. They do freak out if they’ve set a look in color, come to us for a review of the conform and see something completely different. (like warm vs cool, etc)
for sure the dream is to have a magical pixel perfect match - lets shoot for the moon here - but the results from Nuke match grade node has worked for me pretty well…. and flame currently has ZERO ability to one-click match something and bake out a .cube file.
Well now it has Greater Than Zero ability to one-click match and bake out a .cube file!
I looked into ML options to get better results, and didn’t find a repo that would work. There’s an AI Color Match plugin for After Effects. Maybe that guy would port to OFX?
Yesterday i tried your color match tool eric and also trying out Colourlab tool. the OFX installed for yours very easily in linux flame but the colorlab didnt so im waiting to hear back from colourlab as to why the OFX doesnt show up. i used that as a standalone instead so still had the chance to test. My main problem with both from the standing of the actual shot im trying to match ( same frames same plate - client changed grade on a cleanup plate and cleanup company not willing to swap out source - so trying to grade an already cleaned up rec709 to the new grade) that its all missing selective match. different keyed areas are not matched. so overall hue, sat, density etc are trying to get matched but for instance the skin tones of people are way off. Faces are too red instead of a yellow tint etc. is this the nature of those areas being selective and keyed. Does the color matching need parallel keyed areas to match - ie to select the back wall to match, then another for faces, then another etc etc. all to combine for final match the same way as we actual grade an image? instead of a seemingly overall attempt to match all colors.
I guess my image is complicated maybe? 5 or 6 people in a boardroom so many areas where the original grade would have been adjusted or is the problem something else. i also tried a flat camera source and still didnt solve it well.
Yeah, I think the nature of what I built here is just very general. It’s not going to be able to do those kind of localized matches. I tried building that too, but it was a crash and burn. LUTs just don’t have that level of detail.
