TADA — Oral Swallow
"I'm finished. This is a result that you gave me. I'm showing I swallowed it"
She takes the finish, catches all of it, tips her head back, swallows it in one gulp — then opens wide and puts her tongue out to show you it's gone. The ta-da. No spilling, no mess: she catches it, she keeps it, she proves it.
This is a finale specialist: it performs the catch, the swallow, and the reveal, and it's built to be stacked on an action LoRA that carries the scene up to that moment (strengths and stack details are in each version's notes). Trained on 500 hand-narrated clips from 40+ different couples — every beat has dedicated footage.
Trigger: TADAA. — first word of your prompt.
The six beats
Buildup — she works him, head moving
The pause — he warns, she opens wide
The finish — spurts, aimed
The catch — it lands on the tongue and stays in her mouth
The swallow — head tips back, one gulp, eyes come back up
The ta-da — mouth opens wide, tongue out, bare and clean
Each beat needs about 3 seconds. Count your beats and size your render — the full sequence wants ~20 seconds. Don't cram six beats into ten seconds.
The start image is half the result
For perfect containment: POV looking DOWN into her open mouth. Her below — on her back or kneeling — camera above, mouth wide open, tongue out, tip resting on it. This is the trained angle: full pooling, zero spill.
Side angles make great story shots, but expect a dribble.
Always start mid-act with the mouth open — this LoRA continues an action far better than it starts one. A closed-mouth start frame gives you a chewing jaw.
Prompt craft — the words matter
These trained sentences are load-bearing; use them verbatim:
"landing across the waiting tongue below" — the aim
"The ejaculate flows into her mouth and she holds all of it in." — the containment
"The ejaculate fills under her tongue." / "...fills the back of her throat." — where it goes
"Her head tips back and the cum all goes down in one swallow, then the eyes come back up." — the swallow
"The mouth opens wide again and the tongue comes right out, bare and wet and pink." — the ta-da
Dials that work:
Amount: "a small / modest / large amount" — all contained at the money angle; modest is the sweet spot
Texture: "watery", "mostly clear with white streaks" — say the consistency you want. Avoid plain "clear" (renders amber) and avoid "jets" (renders a geyser)
A spoken warning ("I am cumming, get ready") makes a great transition trigger before the finish
Rules learned the hard way:
Name the head or mouth as the mover ("her head moves back and forth", "her mouth glides") — never "her lips sliding" (draws a stacked triple-mouth)
Give idle hands a job ("her hands resting on her knees") — parked hands stay out of the shot
Never negate ("not thick" summons thick) and never describe silence — name sounds that exist
End every prompt with a second of cleanup — she licks the tip clean. He just came; there will be drippage, and the cleanup sells it
Audio
Sound is trained in — wet sounds, breathing, grunts, and spoken lines all respond to a SOUND section in your prompt. See each version's notes for the sampler settings that keep the audio clean on that platform.
Credits
The example stack uses HMPenis by HearmemanAI and the cum physics LoRA by Opossum_ — stacked, not merged, per its license. Use their originals so you get their updates. Thanks to both.
Description
v1.0 — MiniMax H3
First release. Trained on 500 hand-narrated clips (40+ couples) — every beat of the sequence has dedicated footage: the buildup, the pause, the finish, the catch, the swallow, and the ta-da reveal.
Quick settings
settingvalue Strength0.45 (0.4–0.5; higher kills the finish) Stack withyour action LoRA @ ~0.4 · HMPenis @ 0.6 · cum physics LoRA @ 0.4–0.7 (texture dial) Samplereuler / simple — other samplers cause crunchy audio on H3 Steps6–8 fast · 12 for best audio (8-step turbo LoRA @ 1.0) Length20 s for the full sequence (~3 s per beat) Resolution1.0 MP, match your start image's aspect
The one rule that matters: for perfect containment use a POV-down start image — her below (on her back or kneeling), camera above, mouth wide open, tongue out, tip resting on it. Side angles work for story shots but expect a dribble.
Prompt with the trained sentences (see the main description for the full word guide): "landing across the waiting tongue below" · "The ejaculate flows into her mouth and she holds all of it in." · "The ejaculate fills under her tongue." End every prompt with a second of cleanup — she licks the tip clean. It sells it.
Trigger: TADAA. first word of the prompt. Audio is trained in — grunts, wet sounds, breathing all respond to a SOUND line.
FAQ
Comments (17)
Funniest trigger word in a long time! 😂 Thanks, made my day ❤️
Short and sweet!! 😁
Lmao
Seconded.
This is amazing! The results look solid, and the topic is somehow rarely tackled.
Did you try training without a trigger word? Or using it without one? Sometimes prompts pick up on the word and say it aloud, especially when you start lowering the strength. Sometimes you just need a touch of a lora to get what you want, so focusing on legit action phrasing, and phrasing that other loras may share, can help with weighting. Avoids frying the video with high weights.
Thanks for that. I appreciate it. I am trying to fill in places where there’s weaknesses. This was actually almost kind of a commission nonpaid commission. Somebody asked me to make this for them and I was like you know what they’re really isn’t anything that fits that so a lot of my Laura I think they’re gonna be turning more finishers which are not as popular as just say the main act thanks a lot for the positivity. I have not tried this yet without a trigger word. I just uploaded this morning about five minutes before I went to work thanks again.
@LORAGEEK What kind of hardware did you train this on? Or was it done on the cloud? thx
@Dav8890 All my own rig:
AMD 7900X cpu
128 GB DDR 5 RAM
RTX 5090
Fedora Linux OS
@LORAGEEK How long did that take you? I have about the same setup except a 3090, which is a 33% difference in VRAM... Oh, and Windows 11, which probably takes up more VRAM than Fedora.
@Jellai i’ll be straight up, honest with you. There was a lot of failures. This particular one I think it took me about a week. The isn’t that bad it’s probably only a couple hours maybe 3-4 hours but the failure rate was so high
@LORAGEEK Ah, so I suspect it'll take about 8 hours for me. After all those failures, did you learn some good lessons that would affect all loras you train from here on? Did you use AI Toolkit or Musubi? I heard AI Toolkit has a bad implementation, but that the musibi tuner nightly has the good shit.
@Jellai Absolutely absolutely learned a ton of stuff and I did use the AI I use AI to assist me. I use Claude, which is funny because any other place you would probably have people yell at you but if you’re not an AI around here then you’re in the wrong spot so I have Claude kind of assisted me and I’ve built a lot of tools for myself to use and Claude built like a whole framework.
@LORAGEEK Oh wow. So you're not using Musubi, Kohya, or AI Toolkit? You just rolled your own? Impressive. You really should check out the sophisticated stuff Musubi added, in case that provides more opportunities. Here's a recent (maybe even outdated) checklist:
https://github.com/kohya-ss/musubi-tuner/issues/1029
@Jellai Sorry I was at work talking and replying at the same time. I did use AI toolkit but did develop some custom workflows with Claude
@LORAGEEK ddr5 - arrrrrggggggg im stuck on ddr4, but at least i have a rtx6k btu still. arrrrggg
@The_Last_Goblin_King I got it as soon as it was released my kit cost me $220, and I thought it was a lot.
@LORAGEEK omg you know what's crazy, before rampocolypse I bought a 256gb kit on ebay for 250. Then, like a month later, RAM went crazy, and I sold 128 GB of it for 600$ and kept the 96 GB. i honestly don't know if I'm missing anything by doing DDR4 vs dd5 with my card.