Thanks for the very well made video and great explanation! I roto the hard edge as well , as I believe it gives me the biggest control down the line to final the shot.
I really appreciate that you demonstrated your technique on a real shot. As someone still learning and practicing roto, it feels impossible to learn with just theory like "just roto the middle of the mb." I'm still deciding which method I like best and need to spend time practicing with both, but I like that the hard edge is a definitive line to shoot for rather than the ambiguous middle mb. Also, thank you for your tools!
The number of keyframes required depends on factors such as the movement, camera settings, and the shape of the object in the shot. In Carlos's video, the ball was round and moves in a linear path, whereas in your video, it is not. Therefore, more keyframes are needed, but still fewer than if we were to maintain a hard edge for the subject My clever choice would be doing roto keeping edge in the middle of MB and hardedge and using PXF_Vectoredgeblur for edgefix! Never mean the method you showed is right or wrong! Really liked the way you recreated the hardedge!
The ball in Carlos' graph where he shows the keyframes is definitely not moving in a linear path. 🙂 My point is exactly what your are saying, in real life where things deform, move in non-linear paths/speeds, camera shakes, etc... the difference in number of keyframes between method A and method C becomes negligible in my opinion. However, when doing method C, you end up dealing with all the cons that I mention in the video. From a roto artist point of view, method C is the best for sure... you never see the hard matte, so you don't have to QC it, finesse it, etc... Aim roughly for the middle, turn on moblur and you're done! Pass to comp! Problems start to appear in comp, after the gamma down, when you realize that the hard edge overshoots the ball. Then the edge extend/vector blur starts to smear the sky instead of the color of the ball... then you start noodling with the roto (which sucks because moblur will screw the prev/next keyframe as you tweak the animation, and your roto overlay is now meaningless). If you don't have the splines, then you have to introduce erodes that you have to animate, etc... which also sucks. So at the very least, if you must use method C, I would include a QC step where roto has to render a premult through a gamma down alpha to make sure that no obvious parts of the BG are visible in the hard matte.
@@pixelfudger You are partially correct, non linear movements will require marginally more frames, but not a frame by frame, like pretty much the one you swapped by the end saying 'this would be horrifying to do by experts'......oh hell yeah, not just horrifying to do but horrifying to pay!!! .... we're talking about in the absolute best case scenario a double the amount of frames and in 95% of the general shots something like 10x more, do you know how much this represents when multiplied by a cost per day in a full show with 500 shots? But your point really is not the number of keyframes, once I explain what you're really missing in your workflow you will realise there's no guessing involved, there's verification. But I do believe the biggest concern here is the method of extruding and the damage created. While I like the directional extruding, absolutely correct approach, we disagree drastically when you destroy the original hard edge and then bring it back with an arbitrary grade....that grade has to be animated frame by frame since it will be wrong right on the next frame. Another frame by frame workflow.... :) I believe you might not be friends with many VFX Producers , you cost them a pretty penny ;)
I never claim that my swapped roto at the end is the most efficient... hence my 'horrifying' comment. The main reason I have so many keyframes is because, as I adjusted my hard edge with moblur 'on', I kept destroying adjacent keyframes... requiring more tweaking than necessary. As I said on linkedIn, you're more than welcome to give it a try and show everyone how it's done. Not just the roto but the actual comp of the ball including edgeExtend etc... I agree that the arbitrary grade to 'paste' the original ball on top is a bit dodgy, but it doesnt need to be animated. In hindsight I think playing with the gain is a mistake, but gamma is fair game. Since I did the video I have a slightly more refined solution that is gamma-only built-in to the PxF node itself that removes the guesswork from that part and doesnt require frame by frame animation. Let's remember that I would be perfectly happy with the incorrect 'inward' moblur, or a small erode on the hard matte before edgeExtend/vectorBlur to lower the amount of precision required... but *someone* in the roto department is not happy about the moblur not being perfectly pixel accurate to the plate. So I'm not sure who is costing producers so much money at this stage. ;-)
Hello Xavier! Thank you for this interesting video, even if I have some points of disagreement, it makes us think about our work habits. Always looking for the best way ! Concerning the roto, I am definitely in favor of placing it in the middle. Having the most precise mask possible seems essential and the most adaptive to me regardless of the type of FX to be done. I detect some beliefs or small personal orientation regarding the difficulty of establishing a roto in the center. I would like to point out that if we adjust a frame and move the next frame by a few units, even in large movements, the difference on the previous frame will be very tiny. And tiny difference, in a moment of extreme motionblur in most cases will be without consequences. Which is to say imperceptible. And playing on this terrain, with a little dishonesty, we could say the same for the roto in A. It is not always obvious during big motionblur to perfectly find position A. We could place it in a zone with 10% transparency without being sure. Even if I admit it, it remains visually obvious. With method C (middle), you have to do a second quick roto adjustment pass with the motion blur active in visualization to adjust its placement according to the motion blur (only concerns moments with a lot of movement in general) . As you can point out yourself, your alpha compensation grade is a bit arbitrary. I would tend to manipulate blackpoint and whitepoint simultaneously and then add softness to the extremes with a colorlookup (H shaped interpolation, to be adjusted if necessary). To avoid sharp aliasing (middle edge to hard edge), it is possible to use mask vectors. In high values the grade is active (above 10 = 100% active), between 4 and 10 = feather and in small values (between 1 and 4) inactive grade. (Arbitrary values for understanding). So the hard edge processing will only apply in moments/areas with strong motion blur and we will not break the aliasing when we are in focus. Applying motionblur to the RGB via vectors is definitely the right way to do it when you have so much contrast. Thank you again for this high quality video and sharing your knowledge. You are a great source of thinking and knowledge. Alexandre Ribeiro.
I think you've convinced me in this case with a moving motion blurred object. What about a static defocused object? How would you deal with for example a person very out of focus in a shot? Vector blur isnt going to help in this case? Eroding this roto edge out also wouldnt work depending on what shapes are present in the object.
For heavily defocused subjects, I would do the exact same technique except I would replace the vectorblur with a defocus node. Roto the core, premult, defocus RGBA, comp the core on top to avoid double-defocus.
@@pixelfudger Hmm maybe I'd need to give it a go to understand how it would work. What I'm imagining though is that the defocus would blur inwards as well as outwards so when placing the hard edge back on top you'd have a hard line because the new defocus is starting inside the hard line. Hopefully that made sense... 😅
A new version of PxF_VectorEdgeBlur is available to make adjusting the edge much easier: ua-cam.com/video/mGIMNoZ7l-0/v-deo.htmlsi=bxTka6N1XqQhUC0D
Thank you very much for the great explanation! The motion blur problem was bugging me for years. I am going to practice with your tool.
Glad it was helpful!
Thanks for the very well made video and great explanation!
I roto the hard edge as well , as I believe it gives me the biggest control down the line to final the shot.
I really appreciate that you demonstrated your technique on a real shot. As someone still learning and practicing roto, it feels impossible to learn with just theory like "just roto the middle of the mb." I'm still deciding which method I like best and need to spend time practicing with both, but I like that the hard edge is a definitive line to shoot for rather than the ambiguous middle mb. Also, thank you for your tools!
That's a great attitude. I encourage you to experiment with both techniques and see the pros and cons of both for yourself!
The number of keyframes required depends on factors such as the movement, camera settings, and the shape of the object in the shot. In Carlos's video, the ball was round and moves in a linear path, whereas in your video, it is not. Therefore, more keyframes are needed, but still fewer than if we were to maintain a hard edge for the subject
My clever choice would be doing roto keeping edge in the middle of MB and hardedge and using PXF_Vectoredgeblur for edgefix!
Never mean the method you showed is right or wrong! Really liked the way you recreated the hardedge!
The ball in Carlos' graph where he shows the keyframes is definitely not moving in a linear path. 🙂
My point is exactly what your are saying, in real life where things deform, move in non-linear paths/speeds, camera shakes, etc... the difference in number of keyframes between method A and method C becomes negligible in my opinion. However, when doing method C, you end up dealing with all the cons that I mention in the video.
From a roto artist point of view, method C is the best for sure... you never see the hard matte, so you don't have to QC it, finesse it, etc... Aim roughly for the middle, turn on moblur and you're done! Pass to comp!
Problems start to appear in comp, after the gamma down, when you realize that the hard edge overshoots the ball. Then the edge extend/vector blur starts to smear the sky instead of the color of the ball... then you start noodling with the roto (which sucks because moblur will screw the prev/next keyframe as you tweak the animation, and your roto overlay is now meaningless). If you don't have the splines, then you have to introduce erodes that you have to animate, etc... which also sucks.
So at the very least, if you must use method C, I would include a QC step where roto has to render a premult through a gamma down alpha to make sure that no obvious parts of the BG are visible in the hard matte.
@@pixelfudger You are partially correct, non linear movements will require marginally more frames, but not a frame by frame, like pretty much the one you swapped by the end saying 'this would be horrifying to do by experts'......oh hell yeah, not just horrifying to do but horrifying to pay!!! .... we're talking about in the absolute best case scenario a double the amount of frames and in 95% of the general shots something like 10x more, do you know how much this represents when multiplied by a cost per day in a full show with 500 shots? But your point really is not the number of keyframes, once I explain what you're really missing in your workflow you will realise there's no guessing involved, there's verification. But I do believe the biggest concern here is the method of extruding and the damage created. While I like the directional extruding, absolutely correct approach, we disagree drastically when you destroy the original hard edge and then bring it back with an arbitrary grade....that grade has to be animated frame by frame since it will be wrong right on the next frame. Another frame by frame workflow.... :) I believe you might not be friends with many VFX Producers , you cost them a pretty penny ;)
I never claim that my swapped roto at the end is the most efficient... hence my 'horrifying' comment.
The main reason I have so many keyframes is because, as I adjusted my hard edge with moblur 'on', I kept destroying adjacent keyframes... requiring more tweaking than necessary.
As I said on linkedIn, you're more than welcome to give it a try and show everyone how it's done. Not just the roto but the actual comp of the ball including edgeExtend etc...
I agree that the arbitrary grade to 'paste' the original ball on top is a bit dodgy, but it doesnt need to be animated. In hindsight I think playing with the gain is a mistake, but gamma is fair game.
Since I did the video I have a slightly more refined solution that is gamma-only built-in to the PxF node itself that removes the guesswork from that part and doesnt require frame by frame animation.
Let's remember that I would be perfectly happy with the incorrect 'inward' moblur, or a small erode on the hard matte before edgeExtend/vectorBlur to lower the amount of precision required... but *someone* in the roto department is not happy about the moblur not being perfectly pixel accurate to the plate. So I'm not sure who is costing producers so much money at this stage. ;-)
Hello Xavier!
Thank you for this interesting video, even if I have some points of disagreement, it makes us think about our work habits. Always looking for the best way !
Concerning the roto, I am definitely in favor of placing it in the middle. Having the most precise mask possible seems essential and the most adaptive to me regardless of the type of FX to be done.
I detect some beliefs or small personal orientation regarding the difficulty of establishing a roto in the center. I would like to point out that if we adjust a frame and move the next frame by a few units, even in large movements, the difference on the previous frame will be very tiny. And tiny difference, in a moment of extreme motionblur in most cases will be without consequences. Which is to say imperceptible.
And playing on this terrain, with a little dishonesty, we could say the same for the roto in A. It is not always obvious during big motionblur to perfectly find position A. We could place it in a zone with 10% transparency without being sure. Even if I admit it, it remains visually obvious.
With method C (middle), you have to do a second quick roto adjustment pass with the motion blur active in visualization to adjust its placement according to the motion blur (only concerns moments with a lot of movement in general) .
As you can point out yourself, your alpha compensation grade is a bit arbitrary. I would tend to manipulate blackpoint and whitepoint simultaneously and then add softness to the extremes with a colorlookup (H shaped interpolation, to be adjusted if necessary).
To avoid sharp aliasing (middle edge to hard edge), it is possible to use mask vectors. In high values the grade is active (above 10 = 100% active), between 4 and 10 = feather and in small values (between 1 and 4) inactive grade. (Arbitrary values for understanding).
So the hard edge processing will only apply in moments/areas with strong motion blur and we will not break the aliasing when we are in focus.
Applying motionblur to the RGB via vectors is definitely the right way to do it when you have so much contrast.
Thank you again for this high quality video and sharing your knowledge.
You are a great source of thinking and knowledge.
Alexandre Ribeiro.
I've added new features in PxF_VectorEdgeBlur to eliminate the guess work/'arbitrary' grade. See this video: ua-cam.com/video/mGIMNoZ7l-0/v-deo.html
Instead of adjusting of gamma, I would use the clamp node with minimum value of 1 and minclampto enabled in order to get the core matte. 12:51
I think you've convinced me in this case with a moving motion blurred object. What about a static defocused object? How would you deal with for example a person very out of focus in a shot? Vector blur isnt going to help in this case? Eroding this roto edge out also wouldnt work depending on what shapes are present in the object.
For heavily defocused subjects, I would do the exact same technique except I would replace the vectorblur with a defocus node.
Roto the core, premult, defocus RGBA, comp the core on top to avoid double-defocus.
@@pixelfudger Hmm maybe I'd need to give it a go to understand how it would work. What I'm imagining though is that the defocus would blur inwards as well as outwards so when placing the hard edge back on top you'd have a hard line because the new defocus is starting inside the hard line. Hopefully that made sense... 😅
Nice job Xavier !
Merci Valérie!
GG
you made a good poin, convinced me at least ^^
how i can blend mask to background and no see the eadge?
✨👌💪😎😮😎👍✨
'PromoSM'