[Return] [Bottom]

Posting mode: Reply

BBCode
(for deletion)
  • Allowed file types are: gif, jpg, jpeg, png, webp, webm, mp4
  • Maximum file size allowed is 25000 KB.
  • Images greater than 255 * 255 pixels will be thumbnailed.
  • 190 unique users in the last 30 minutes (including lurkers)




    First
    [1]
    Last

    File: falling_noose.mp4
    (1.05 MB, 1664x1248) ▶
    1103924
    This isn't a final product, just a piece of the puzzle.

    I wanted to get a noose to drop down from a ceiling, where it had been concealed. And to enter a scene without the camera view changing to point at the ceiling or any imaginary mechanism in the ceiling.

    No matter how I tried to prompt this, I get anything but what I want when trying to add this idea to a real scene. I'd get a noose slowly lowering into the scene at best, wrong place, wrong scale, always with the camera view tilting up to the ceiling, etc.

    So I decide to see if, in isolation, I could at least get the physics to work. And that, I can say, works great.

    I simple used the first frame of the attached video, created as a still image, and this prompt with MMH3:

    Noose drops under the influence of gravity, swinging from black pivot point. Rope straightens out under influence of gravity. Maintain solid green background throughout clip.
    >>
    File: noose-overlay-sample.mp4
    (1.29 MB, 960x720) ▶
    1353431
    It took at least an hour, if not more, to generate the first video, even on an RTX 5090, because I set the resolution up to 2M.

    The idea is to overlay a zoomed-in portion of this over the main video, using DaVinci Resolve to do the overlay.

    Here's a simplified example of what I'm going for, using a safely non-realistic still image for what will be realistic video.
    >>
    File: obvpm_h3_result_noaudio.mp4
    (1.71 MB, 1024x576) ▶
    1793978
    >>30045

    Honestly, you may be over-complicating your idea by a lot. MiniMax will easily take two input images and can handle green screen to combine them together properly.

    This is just
    <Subject 1> stands on the stool, observed by <Subject 2>
    Noose drops under the influence of gravity, swinging from the pivot point, swings between <Subject 1> and the camera. Rope straightens out under influence of gravity. <Subject 1> stands on the chair, frozen.


    Video input is slow and so this took 4m to generate (image only is about 120s) and it needs to be upscaled, but even that should not take over an hour on your HW.
    >>
    >>30052
    > Video input is slow and so this took 4m to generate (image only is about 120s) and it needs to be upscaled, but even that should not take over an hour on your HW.

    But your sample is a simple overlay. I needed the actual pivot point from which the noose swings to be way off screen. Trying to explain an unseen, off-screen pivot point is what totally confused MMH3.

    Maybe somebody with a much better prompt-fu could have managed it, but making a separate high-res overlay and doing the composition in DaVinci worked out a lot better for me.

    The hour-long processing time was simply from generating anything at 2 megapixel resolution.
    >>
    >>30052
    Α closer-to-final composite: https://files.catbox.moe/1j5ym5.mp4

    I also needed DaVinci because, no matter how I worded my prompt, even with AI prompt rewriting assistance, MMH3 was bound and determined that the girl would look upward when I didn't want her to do that, when I instead wanted her to keep her head down.

    So I fixed that by video editing rather than prompt editing.
    >>
    >>30119
    Is there any reason behind teacher's unbuttoned shit? Kind of dominance over undeveloped one?
    >>
    >>30120
    > Is there any reason behind teacher's unbuttoned shi[r]t? Kind of dominance over undeveloped one?

    Because leading up to this she made the girl suck her tits.

    [Top]

    Delete post: []
    First
    [1]
    Last