got a shot: 50fr long shot by still cameras in a horizontal line of some people standing still. A kind of bullet time without the arc.
we are being asked to slow it down. Natch.
anyhoo: motion and both ML time warps are failing. As is Kronos. Yet to try twixtor. Have had mixed results with Kling. Some frames it works and some it doesn’t. Besides which for Kling it has to be 8 bit and it comes back all banded. It would be more helpful if it was an aces solution.
So I thought: are there any comfy ui solutions that can stretch time? I’d like to do a 25% timewarp. Or at least a 50%. Any ideas?
How about you repeat each frame then move each second frame a little left or right (depending on the direction of the original move) to make the motion look smooth?
You might also try repeating each frame and putting a gap effect morph on top of each repeated frame. I describe how to make one in a tutorial, or depending on your version, I could send you the effect. 2026.2
Similar. Does it not work because it looks bad, or does it just not work at all? Mine “works” but I have no control over whether it looks good or not.
I also have a single frame version that uses the ML timewarp.
Pretty much. I think there might be some subtle differences if it only has 2 frames to judge the in between rather than the whole segment at once, but I’ve never really tested it. I only use this to fix single bad frames. I thought I’d just toss it out there on a lark in the hopes that it got someone out of a jam. It didn’t. So it goes.
Without seeing the frames and how they are failing it is hard to offer any meaningful advice. Can you describe how it is not working?
some thoughts.
Do you need to stabilise/align each frame before applying the speed effect?
Was the capture a line of cameras or did they just use one, get people to stay still and just move the camera bit by bit not perfectly measured? Similar to the align suggestion, if the distance between the camera and each frame is different then it might be worth trying to make the position of each frame an equal distance manually.
Have you tried dropping frames? Sounds counter productive but by dropping every second frame it may help.
Do you need to break up the shot into foreground & background get something working for the foreground then rebuild the background and animate that for a 2.5D effect? Hard to know if this is plausible without seeing the shot.
Have you tried Resolve’s AI speed effect? I think it’s called Speed Warp? I usually found Flame’s to be better but have seen it work better occasionally. Also have you tried both the Flame internal as well as the brilliant one @talosh has given the community? I’m assuming so given your post but thought I’d throw those suggestions in there too.
If those people are standing still and you’re essentially panning across, if I understand it right. Could you bring all the shots into some mega plate (Flame or Mocha), and then pan across that in an action via keyframes. Add motion blur to taste, and maybe automatte the people and treat the background separate for any real parallax you may need to sell it? May require some experimentation.
As we mentioned, without seeing the actual shots, hard to tell. So just spit-balling ideas.
The lenses are far enough apart to make foreground objects too far to interpolate. Kling does a great job but on many frames it decides to move things rather than treat each frame as a snap shot.
might have to just rebuild these. Not sure yet. Or try and make these problem frames a Gaussian splat. No idea how to do that yet.
I did a shoot like that many years ago. I had to create a slower pan by morphing every 2 frames into an inbetween back when we had the original warper. It was misery, but it got nominated for a sports emmy.
I had thought about suggesting a Gaussian splat or photogrammetry approach for this depending by in how much camera data you have. You always need to do a fair degree of cleanup but with that many cameras I think you might get a decent result. Then you can control the camera movement and everything. Not sure if you have enough images to get a lot of detail for a closeup but for a mid to a wide you’d think it was feasible.
I suppose if all fails, you can manually rebuild the scene, by cutting out element and doing it as if you were animating a still. As allklier suggested. Megaplate might be a way stitch them all into one thing. Bring it maybe into Photoshop or something similar. Retouch any anomalies in needed before and after and rebuild the camera virtually and slow it down.
I’m with @AdamArcher. Solve the cameras make splats and then rerender the move in 3d. Might be the cleanest approach. Could potentially stitch together a bunch of solved SHARP outputs—one from each of the cameras. Never tried it but theoretically it could also work.
