[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.
  • 108 unique users in the last 30 minutes (including lurkers)




    First
    [1]
    Last

    File: image.png
    (336 KB, 832x640)
    344881
    I'm having some fun finding out what I can do with MiniMax H3 running on my own home ComfyUI setup.

    https://files.catbox.moe/budgxa.mp4

    I'm still refining this, and will probably post an update later, but I'm pretty happy with where I've gotten so far.

    I first worked with an image-to-video (I2V) workflow, but had some problems like a diseased-looking penis when the video zoomed in. (I'm going to post more detail in follow-up posts to make it easier to add extra pictures.) I switched to a reference-to-video flow to fix that problem, but that opened room for a lot of other improvements as well.

    This is the prompt used to create this video:

    Use <Picture 1> as is for first frame. Use <Picture 2> as refernce.
    
    Mother kneels behind her daughter to make sure her daughter keeps her mouth open, eager to swallow as soon as the man in front of her begins to piss.
    
    [0-3 seconds] The woman says, as per <Audio 1>, \"Use my little girl as your urinal\". The camera begins to zoom in on both the girl's face and the mother's face, orbiting to get a more frontal view. In close up, the man's cock looks like <Picture 2>.
    
    [3-5 seconds] The girl says, as per <Audio 2>, \"Please piss in my mouth, sir.\" The man's cock does not touche the girl's mouth, he keeps the tip of cock a few inches away from her mouth and lips.
    
    [5-12 seconds] Urine begins to stream forth from the tip of the man's cock into the girl's open and eager mouth, arcing through the air. Use <Video 1> as a reference for what a stream of urine should look like. The girl lets piss puddle inside her mouth, swallows the piss, then opens her mouth again for more, two cycles of collecting piss and swallowing it. Her mother watches the girl swallow with admiration. The girl swallows most of the man's piss, not letting very much escape her mouth except when she closes her lips to swallow. The man continues to keep his cock a few inches from the girl's mouth.
    
    [12-15 seconds] The girl now takes the man's cock into her mouth and sucks on it, letting piss dribble out of her mouth as she sucks.
    
    Camera: No jump cuts. Keep girl's face in view at all times.
    >>
    File: image.png
    (1.35 MB, 1440x1080)
    1416887
    I created this start frame for another video a while back, and don't remember all of the details of how I made it.

    https://files.catbox.moe/grl9s4.png

    I used Qwen Image Edit on a crude image first created using https://app.justsketch.me/, which provides a digital version of the old-fashioned posable wooden mannequins artists often use for human figure drawing.

    This was much better that trying to verbally describe this pose, especially because AI models in general have an awful time understanding this particular kind of upright kneeling -- tell an AI image or video model you want a character to be kneeling, and it will almost always come out as kneeling and squatting at the same time, hardly ever understanding the idea of the upper part of the legs being vertical.
    >>
    File: image.png
    (467 KB, 640x816)
    479180
    I'd heard that MiniMax H3 wasn't great at doing genitals. Boy howdy, the stories were true.

    https://files.catbox.moe/ewmvbq.png

    The Reference version of MiniMax H3 provides a great way to fix this diseased looking member. I used Qwen Image Edit to zoom in on the man's cock, which it did much more appealingly, and then used that image as a reference image for MiniMax H3. With that little bit of help a much better quality video was generated.
    >>
    Once I had a healthy looking cock, and an R2V workflow to get that, I decided I might as well take advantage of the fact that you can supply audio references as well.

    Why would I want audio references? Because like many AI video systems, the artificial voices have all of the emotional range of a brick. My replacement voices won't be earning any Academy Awards either, but they're better than what MMH3 produced on its own.

    The voice samples I provided as reference inputs were created using ElevenLabs, specifically "Serafina" for the mother, and "Lulu Lollipop" for the daughter.
    >>
    327769
    One more thing I wanted to try was to see if I could make the urine stream look more realistic. It might be hard to tell from the directly-posted image, but here's a sample of an earlier version of the video: https://files.catbox.moe/0bqvma.mp4

    So I extracted 4 seconds worth of real-life pissing into a mouth (I love both lolis and milfs, and had a good milf sample available) and used that as a video reference.

    It really helped make the piss look more real. The downside is the already slow rendering time for the video (around half an hour for a 15 second clip using my RTX 3090) more than tripled.

    I'm definitely starting to feel the urge to splurge on a 5090. It's incredibly annoying, however, how the price of those cards has more than doubled because of AI industry demand.
    >>
    Amazing work. My only complaint is that I don't like the oversized penis. You try to make everything else realistic but then add a fake porn style penis.
    >>
    >>24587
    > Amazing work. My only complaint is that I don't like the oversized penis. You try to make everything else realistic but then add a fake porn style penis.

    Maybe porn has desensitized me too much, but the cock didn't strike me as that big. Definitely on the large size, but not a fantasy world monster... not too far outside two standard deviations. 😄
    >>
    A new run fixed the problem where mother wasn't properly lip synced, unlike the girl.

    https://files.catbox.moe/4ee2j5.mp4

    I'm not sure exactly how I fixed the problem, because I changed a couple of different things, and there's always the fact that there are random factors involved unless you lock down the seed (which I didn't).

    When I added only "Make sure Mother's lips sync with <Audio 1>" to my prompt, the mother was then lip synched, but the daughter failed to speak at all. So I added "Make sure girl's lips sync with <Audio 2>" as well. But at the same time I additionally slightly trimmed both reference audio tracks too, each of which had been just a small fraction of a second over the allotted time for each. It might have been a timing issue that had thrown off the audio handling.

    Since it seems there's always some unexpected rendering weirdness that has to happen, however, this run of the video didn't start with my specified start frame. The overall result is so good I didn't care too much. I'll run a few more iterations any way, and after that, I'd like to try to lengthen the video a bit beyond the 15 seconds where a single run maxes out.
    >>
    >>24590
    >I'm not sure exactly how I fixed the problem
    How i feel most of the time doing anything with AI.
    ref2va seems really powerful but yeah it slows things down a lot
    >>
    After watching the first video I was tempted to shout: "FAKE! There is a baby milk bottle hidden in his dick!" ;)
    The second thought was: 'Is he for her sake pissing well tasting lemonade instead of inedible urine?' :D

    I'm glad I watched all versions before complaining. The third is indeed the best, though the zoom is missing at the start so we can't see the whole nudity. But this gives us the thrill of waiting for more of your great work. :)
    >>
    >>24598
    > ...though the zoom is missing at the start so we can't see the whole nudity.

    The next iteration fixed that problem... but of course botched a new thing.

    The guy started pissing before the girl speaks, but just barely.

    His urine stream pauses midair about a centimeter out from his cock, and doesn't resume flow until the girl finishes talking. 😄
    >>
    I upped my resolution to 992x736 (I had been doing 832x640), but that slowed rending of this video down to 148 minutes, just two minutes shy of 2.5 hours.

    I had already decided to give into getting an RTX 5090, reading that it would do this kind of work 2-3 times faster.

    But what about FOUR times faster?

    Yep! I ran a test on Runpod.io using a 5090, same workflow, same prompt and inputs, and got the same rendering job done in 37 minutes.
    >>
    >>24572
    Amazing work, especially when she gently takes the penis with her lips. Don't listen to anyone, go on, you're doing great
    >>
    File: image.png
    (641 KB, 992x736)
    657261
    Here's an improved and extended version of the video. The merging of two clips isn't quite as seamless as I'd like, but not too bad.

    https://files.catbox.moe/hzitsw.mp4
    >>
    >>24739
    😍 Cool

    [Top]

    Delete post: []
    First
    [1]
    Last