<h1 id="he-finishes-in-her-mouth-and-you-actually-see-his-cock-spasm-and-contract-while-it-happens-over-and-over-not-one-twitch.-then-it-runs-out-the-side-of-her-mouth-and-down-the-shaft.">He finishes in her mouth, and you actually see his cock spasm and contract while it happens — over and over, not one twitch. Then it runs out the side of her mouth and down the shaft.</h1>This is the oral creampie / CIM finish specifically — not the money shot where she tilts her head back with her tongue out and he finishes onto her from outside. Here it happens inside, with him still in her mouth, and the contractions are the whole point.Different thing, different tool. If the money shot is what you're after, there are good LoRAs for that and this isn't one of them.Video + audio. It generates its own sound.<hr /><h1 id="two-versions-pick-the-right-tab-at-the-top">⚠ Two versions — pick the right tab at the top</h1>version base model trigger v1.0 — MiniMax H3 MiniMax H3 <code>CUMOUF</code> v1.0 LTX 2.3 / 10Eros v1.5 <code>CUMOUF</code>Same footage, trained separately for each base. They are not interchangeable — a LoRA only works on the model it was trained on.The two models want completely different prompts. Civitai shows one description for the whole page, so both guides are below. Scroll to the one matching the version you downloaded.<hr /><h1 id="minimax-h3-version">▶ MINIMAX H3 VERSION</h1><h2 id="quick-start">Quick start</h2>Base model MiniMax H3 Trigger <code>CUMOUF</code> — first word of SHOT 1 Strength 0.5 Guidance 1.0 — H3 is guidance-distilled, raising it breaks the image Length 12 seconds Start image one already mid-actionTested side by side at a locked seed: 0.5 and 0.7 gave identical motion, but the cum went unnaturally thick at 0.7. Nothing is gained above 0.5.Above about 1.0 it distorts badly. H3 LoRAs adapt the feedforward layers as well as attention, so extra strength redraws texture instead of strengthening motion — at 1.5 you get twisted anatomy.<h2 id="read-this-first:-write-the-prompt-to-match-your-picture">⭐ Read this first: write the prompt to match your picture</h2>Look at your start image. Describe what is actually in it — the camera angle, where her eyes are, whose hands are in shot, what is behind them, his colouring. Then say what should change.Anything you describe differently from the picture is an instruction to redraw it. And when H3 redraws, it re-derives the people too. That is when the face changes, the skin tone shifts, or the whole shot flips to another angle.Four ways to break it without noticing:you wrote picture actually shows what happens "close side view" a three-quarter view the camera flips and everything is re-derived "looking up at his face" her eyes are closed her face gets rebuilt "he puts his hand on her head" no hands in shot a hand has to be invented, and the people change with it "a bedroom, plain bedding" a couch the background is rebuilt, and it drags the rest alongCopying someone else's prompt is exactly how this goes wrong. It was written for their picture. Change the look line to yours and keep the action.<h2 id="structure">Structure</h2>A look line ending in "The environment is constant throughout.", then <code>SHOT 1:</code> with the action, then <code>Audio:</code> last.<pre><code>Close three-quarter view in a bedroom, warm lamp light, plain bedding behind them, tight framing, shallow depth of field. The man and woman from &lt;Picture 1&gt; in their original scene: &lt;describe them&gt;. Her face, her hair and his skin tone stay exactly as they are in &lt;Picture 1&gt; for the whole clip. The camera holds this one angle throughout. The environment is constant throughout.

SHOT 1: CUMOUF. He drives his hips forward and fucks her mouth, thrusting in and out over and over, hard and fast and deep, setting the rhythm himself. She stays where she is and takes it. He thrusts in deep and she gags hard around him, then he draws back and she breathes. He cums, still thrusting, and you see his cock spasm and contract over and over as it pumps into her mouth. The spasm is hard and repeats as he continues to cum. A light dribble of white cum spills from the corner of her mouth and runs down his shaft.

Audio: fast wet slapping, the bed creaking under them, his heavy breathing and low groans, wet muffled sucking from her stuffed mouth. Quiet room tone, no music.
</code></pre><code>&lt;Picture 1&gt;</code> is H3's own reference tag. It works with plain image-to-video, not only the reference model.<h2 id="what-one-small-change-does">What one small change does</h2>change result <code>He holds still</code> anywhere in the prompt Nothing happens. She sits motionless for the whole shot. Never write stillness <code>The spasm is hard and repeats as he continues to cum</code> Much stronger contractions. This is the intensity dial. Leave it out for something subtler <code>Thick white cum runs out and spills over her lips, sliding down the shaft</code> Gallons of it. Use deliberately if you want it over the top <code>A light dribble of white cum appears on the side of her mouth</code> Realistic texture and volume <code>a light dribble of clear cum</code> Renders clear. Most men are closer to clear than white — this is the realistic option Drop the colour word On darker skin tones it drifts toward caramel <code>She keeps her lips sealed and swallows</code> on its own No cum at all — and it kills the spasm too the same line plus a dribble line Works fine. It is the swallow alone that suppresses it <code>She moves her head back and forth along his cock</code> She does all the work, he stands there <code>He drives his hips forward and fucks her mouth</code> He drives it. Make him the subject of the action sentences Gagging in the <code>Audio:</code> block She gags continuously, out of sync with everything Gagging in the prose, tied to a thrust It happens at that moment <code>She coughs</code> Can loop <code>She coughs once</code> Happens once<h2 id="three-rules-worth-memorising">Three rules worth memorising</h2>1. <code>Audio:</code> is continuous. Prose is timed. Anything in the audio block plays across the whole clip — that is where ambience belongs. A sound that should happen at a moment goes in the prose, next to the action that causes it.2. Events happen in the order you write them. Speech, the finish, the aftermath — put them in sequence and H3 times them that way.This is the best thing about H3. A cabin chime written into the same sentence as the finish landed on the spasm. On LTX I have re-rendered the same clip twenty or thirty times trying to get a line to land before the action it describes instead of after. Here it is one render.<pre><code>A loud electronic cabin chime sounds twice as he cums, still thrusting, and you see his cock spasm and contract over and over as it pumps into her mouth.
</code></pre>Avoid spelling out noises — write "a loud electronic chime sounds twice", not the noise itself, or it may get spoken aloud.3. Say whose sound it is. An unattributed "heavy breathing" gets given to whoever is most visible — which produced a woman breathing heavily with her mouth full. Write "his heavy breathing and low groans" and "wet muffled sucking from her stuffed mouth".Never negate. "Less cum", "no spilling", "without going deep" all produce the opposite. Ask for the small version positively instead.<h2 id="setting-and-ambience">Setting and ambience</h2>Only the look line and the <code>Audio:</code> block change. The action stays identical.<pre><code>Audio: ocean waves rolling in and breaking on the sand, steady sea breeze, distant seagulls calling, fast wet slapping, his heavy breathing, wet muffled sucking from her stuffed mouth. No music.
</code></pre>If ambience buries the wet sounds, drop the wind first — it masks everything else.<h2 id="speech">Speech</h2>In quotes, in the prose, at the point it should happen. It lip-syncs on its own.<pre><code>He says, "swallow my load babe"
</code></pre><h2 id="known-behaviours">Known behaviours</h2>Don't ask for something that isn't in the start image. If the prompt says he puts his hand on the back of her head and there is no hand in frame, H3 has to introduce one — and introducing an element is itself a transition, so it re-derives the whole picture to make room. It usually strikes a second or two in, exactly when the new element would have to appear.Anything it has to redraw, it may re-derive. A camera move, a hand coming off, a pull-out. Anchor with <code>&lt;Picture 1&gt;</code> and hold the camera:<pre><code>The camera holds this one angle throughout.
</code></pre>If the first second or two is a different person and the rest is good, just trim the front off. Not worth re-rolling.Audio can push the picture. H3 runs audio and video through the same attention. A loud rhythmic sound cued to a moment can drive motion at that moment — an engine rev on the finish made the whole thing vibrate. Useful or not depending on what you want.Foley grunts are weak. Breathing and speech come through well. That is the source footage, not the prompt. It is not a sound LoRA.<h2 id="h3-training">H3 training</h2>Base MiniMax H3, pruned int8 Clips 1,734 × 39 frames @ 24 fps Rank / alpha 16 / 16 Steps 4,000 — about 2 epochs Trainer ai-toolkit Hardware one RTX 5090H3 reaches the behaviour in far fewer passes than LTX did — a third of the training. Rungs above 4,000 got worse, not better.<hr /><h1 id="ltx-2.3-eros-1.5-version">▶ LTX 2.3 / EROS 1.5 VERSION</h1><h2 id="quick-start">Quick start</h2>Base model 10Eros v1.5 Trigger <code>CUMOUF</code> — first word Strength 1.0 CFG 1.2 – 1.3 Length 8 – 15 seconds Distilled LoRA <code>ltx-2.3-22b-distilled-lora-384-1.1</code> Sigmas the longer schedule below — this mattersFirst-pass sigmas. Worth using: switching to this schedule improved the spasm, the detail and the audio noticeably over a shorter one.<pre><code>1.000, 0.955, 0.893, 0.812, 0.715, 0.603, 0.482, 0.241, 0.121, 0.0
</code></pre>The prompt used for the example renders:<pre><code>Performance: [CUMOUF. Close side view of his cock in her mouth, her lips wrapped
tight around the shaft. He pumps slowly into her mouth, then holds deep as he starts
to cum. You see his cock twitch and spasm over and over, arching and pulsing as it
pumps into her. Thick white cum fills her mouth and spills out over her lips, sliding
down the shaft.]

Sounds: [Man grunts loudly as he cums. Woman coughs and sputters trying to swallow it all]

Dialog: [(off camera) Man says: Take my load in your mouth. Swallow my cum.]
</code></pre>Note the three separate blocks — that matters, see below.<h2 id="three-things-that-will-waste-your-time-if-you-don't-know-them">Three things that will waste your time if you don't know them</h2>1. Don't ask for a man's voice without saying he's off-camera.LTX-2.3 lip-syncs. Ask for a line and it needs a face to put it on — so it drags one into your shot, and once there's a second mouth near a cock it does the obvious thing with it. Cost me an hour to work out.<pre><code>Dialog: [(out of frame) Man says: swallow it]
</code></pre>2. Sound goes in its own block, or it gets spoken out loud.I wrote "he grunts loudly" into the visual description and got a render where the man said the words "he grunts loudly". A grunt is a facial action, so describing it in the picture asks for a face doing it.<pre><code>Performance: [ the action — no sounds, no speech ]
Sounds: [ wet mouth sounds, grunting from the man, moans from the woman ]
Dialog: [ (out of frame) Man says: ... ]
</code></pre>3. <code>pumps</code> means ejaculating here, not thrusting.It's in 82% of the training captions as "as it pumps" — the cock pumping cum. So "he pumps fast in and out" asks for the climax while you're trying to ask for the build-up. Use <code>thrusts</code> for motion.<h2 id="words-it-knows">Words it knows</h2>Measured across all 1,525 captions:word in cock 100% spasm 79% contract 79% hard (intensity) 48%Zero occurrences — these fall through to the base model: penis, dick, load, spasms (with an -s), twitches, jerks, arch.That last group matters. <code>spasm</code> and <code>spasms</code> are different tokens. The trained phrase is "spasm and contract over and over" — use it as written.<h2 id="it-moves-in-for-the-finish-that's-the-point">It moves in for the finish — that's the point</h2>Start from a normal, wider shot. When the finish begins, the camera pushes in on its own.That is the behaviour, not a side effect. It's a finisher, and this is what a finisher should do — the same move an editor would make. You stay in the scene through the act, and when it matters the shot tightens to show you the thing you came for.There's a physical reason it has to. A spasm is a few millimetres of movement. From any distance it's sub-pixel and simply doesn't exist on screen — you can't see one from across the room. Every clip I trained on was cut tight for exactly that reason, so the model composes the same way: it frames for its subject.To get more of the move: start wider, and state the spasm phrase twice. To get less: start closer, and state it once. A start image already cropped tight gives it nowhere to go, and puts more pixels on the spasm from frame one.<h2 id="your-start-frame-decides-what-his-cock-looks-like">Your start frame decides what his cock looks like</h2>Raised in the comments, and it's a good catch — thank you.The model only knows the shape of something from what it can see in frame one. If it's already halfway into her mouth, the head was never visible, so when she comes off him it has to invent one — and it invents an ordinary one. If yours is distinctive at all, that's where the detail goes.So start with it near her mouth, not yet in it. One clean look at the shaft and head is enough and it holds for the rest of the clip.The trade-off: a frame already mid-act starts moving sooner, because this continues an action far better than it starts one. About-to-go-in is the sweet spot — clear anatomy, and something already happening.It will pull back at the end. That's trained in, not a fault: a lot of the dataset is the finish continuing after she comes off him, cum running out of her mouth and down the shaft. To push against it, keep the render at 8 seconds so it ends during the finish, write that she holds him deep and swallows, and say nothing about it spilling down the shaft. None of that is reliable — showing it clearly in frame one is.To get it out of her mouth: start from a frame where he's already out, or say she is looking at it and holding still.<h2 id="strength-1.0-and-the-range-is-narrow">Strength — 1.0, and the range is narrow</h2>Alpha equals rank, so 1.0 is genuinely the LoRA as trained.0.7 clean but flat 1.0 use this 1.5 stronger, but body morphing appearsIf the spasm is weak at 1.0, more strength won't fix it — that's a data limit.<h2 id="known-limits">Known limits</h2>The audio is trained, not the base model's. It generates its own sound — wet mouth sounds, breathing, and grunts that came from the source footage rather than the gritty default. Use the longer sigma schedule; it makes a real difference to the audio too.It's a finisher. 1,169 of 1,525 clips are the climax itself; only 93 are in-and-out motion. It won't drive face-fucking or a long build-up — use another LoRA for the act and bring this one in for the finish.Anatomy holds up but isn't perfect. Nine continuous seconds in her mouth and it comes back recognisably the same — that used to render as chewed-up meat. It still stretches sometimes on the way out, and it can only keep what your start frame actually showed it.Long renders drift. Trained on 2-second clips. 8–15 seconds is the tested band.<h2 id="ltx-training">LTX training</h2>Base 10Eros v1.5 bf16 Clips 1,525 × 49 frames @ 25 fps Sources ~14 couples plus solo footage, every window timed by hand Bucket 640×384×49, rank 8 / alpha 8 Branches video + audio, trained jointly Hardware one RTX 5090 Steps trained to 12,000; released checkpoint is step 11,000Later is not automatically better. I tested the whole ladder and 11,000 was the best trade between spasm strength and stable anatomy — the last checkpoint was not the one I shipped.<hr /><h1 id="how-both-were-made">How both were made</h1>Every clip was cut to hand-written timestamps, and only to seconds where the spasm is genuinely visible. Where I couldn't see it — a hand in the way, a body blocking the shot — that footage contributed zero spasm clips even though the event was happening.Every clip in a group carries one identical sentence for the event. Variety comes from the scenes, not the wording. That's the opposite of the usual advice, and it's what made this version work after eleven that didn't.Full method notes, traps and measurements are published alongside this — please copy any of it. I learned a lot from other people's write-ups while building this, so everything I found is written down and free to take.

CUMOUF — Oral Creampie / CIM with Spasms

Quick startUpdate: Verified works well with LTX 2.5Base model 10Eros v1.5 Trigger <code>CUMOUF</code> — first word Strength 1.0 CFG 1.2 – 1.3 Length 8 – 15 seconds Distilled LoRA <code>ltx-2.3-22b-distilled-lora-384-1.1</code> Sigmas the longer schedule below — this mattersFirst-pass sigmas. Worth using: switching to this schedule improved the spasm, the detail and the audio noticeably over a shorter one.<pre><code>1.000, 0.955, 0.893, 0.812, 0.715, 0.603, 0.482, 0.241, 0.121, 0.0
</code></pre>The prompt used for the example renders:<pre><code>Performance: [CUMOUF. Close side view of his cock in her mouth, her lips wrapped
tight around the shaft. He pumps slowly into her mouth, then holds deep as he starts
to cum. You see his cock twitch and spasm over and over, arching and pulsing as it
pumps into her. Thick white cum fills her mouth and spills out over her lips, sliding
down the shaft.]

Sounds: [Man grunts loudly as he cums. Woman coughs and sputters trying to swallow it all]

Dialog: [(off camera) Man says: Take my load in your mouth. Swallow my cum.]
</code></pre>Note the three separate blocks — that matters, see below.Three things that will waste your time if you don't know them1. Don't ask for a man's voice without saying he's off-camera.LTX-2.3 lip-syncs. Ask for a line and it needs a face to put it on — so it drags one into your shot, and once there's a second mouth near a cock it does the obvious thing with it. Cost me an hour to work out.<pre><code>Dialog: [(out of frame) Man says: swallow it]
</code></pre>2. Sound goes in its own block, or it gets spoken out loud.I wrote "he grunts loudly" into the visual description and got a render where the man said the words "he grunts loudly". A grunt is a facial action, so describing it in the picture asks for a face doing it.<pre><code>Performance: [ the action — no sounds, no speech ]
Sounds: [ wet mouth sounds, grunting from the man, moans from the woman ]
Dialog: [ (out of frame) Man says: ... ]
</code></pre>3. <code>pumps</code> means ejaculating here, not thrusting.It's in 82% of the training captions as "as it pumps" — the cock pumping cum. So "he pumps fast in and out" asks for the climax while you're trying to ask for the build-up. Use <code>thrusts</code> for motion.Words it knowsMeasured across all 1,525 captions:word in cock 100% spasm 79% contract 79% hard (intensity) 48%Zero occurrences — these fall through to the base model: penis, dick, load, spasms (with an -s), twitches, jerks, arch.That last group matters. <code>spasm</code> and <code>spasms</code> are different tokens. The trained phrase is "spasm and contract over and over" — use it as written.It moves in for the finish — that's the pointStart from a normal, wider shot. When the finish begins, the camera pushes in on its own.That is the behaviour, not a side effect. It's a finisher, and this is what a finisher should do — the same move an editor would make. You stay in the scene through the act, and when it matters the shot tightens to show you the thing you came for.There's a physical reason it has to. A spasm is a few millimetres of movement. From any distance it's sub-pixel and simply doesn't exist on screen — you can't see one from across the room. Every clip I trained on was cut tight for exactly that reason, so the model composes the same way: it frames for its subject.To get more of the move: start wider, and state the spasm phrase twice. To get less: start closer, and state it once. A start image already cropped tight gives it nowhere to go, and puts more pixels on the spasm from frame one.Your start frame decides what his cock looks likeRaised in the comments, and it's a good catch — thank you.The model only knows the shape of something from what it can see in frame one. If it's already halfway into her mouth, the head was never visible, so when she comes off him it has to invent one — and it invents an ordinary one. If yours is distinctive at all, that's where the detail goes.So start with it near her mouth, not yet in it. One clean look at the shaft and head is enough and it holds for the rest of the clip.The trade-off: a frame already mid-act starts moving sooner, because this continues an action far better than it starts one. About-to-go-in is the sweet spot — clear anatomy, and something already happening.It will pull back at the end. That's trained in, not a fault: a lot of the dataset is the finish continuing after she comes off him, cum running out of her mouth and down the shaft. To push against it, keep the render at 8 seconds so it ends during the finish, write that she holds him deep and swallows, and say nothing about it spilling down the shaft. None of that is reliable — showing it clearly in frame one is.To get it out of her mouth: start from a frame where he's already out, or say she is looking at it and holding still.Strength — 1.0, and the range is narrowAlpha equals rank, so 1.0 is genuinely the LoRA as trained.0.7 clean but flat 1.0 use this 1.5 stronger, but body morphing appearsIf the spasm is weak at 1.0, more strength won't fix it — that's a data limit.Known limitsThe audio is trained, not the base model's. It generates its own sound — wet mouth sounds, breathing, and grunts that came from the source footage rather than the gritty default. Use the longer sigma schedule; it makes a real difference to the audio too.It's a finisher. 1,169 of 1,525 clips are the climax itself; only 93 are in-and-out motion. It won't drive face-fucking or a long build-up — use another LoRA for the act and bring this one in for the finish.Anatomy holds up but isn't perfect. Nine continuous seconds in her mouth and it comes back recognisably the same — that used to render as chewed-up meat. It still stretches sometimes on the way out, and it can only keep what your start frame actually showed it.Long renders drift. Trained on 2-second clips. 8–15 seconds is the tested band.LTX trainingBase 10Eros v1.5 bf16 Clips 1,525 × 49 frames @ 25 fps Sources ~14 couples plus solo footage, every window timed by hand Bucket 640×384×49, rank 8 / alpha 8 Branches video + audio, trained jointly Hardware one RTX 5090 Steps trained to 12,000; released checkpoint is step 11,000Later is not automatically better. I tested the whole ladder and 11,000 was the best trade between spasm strength and stable anatomy — the last checkpoint was not the one I shipped.

Files with SHA256: 918a19314dfe0530ce3a36d9f28d4d37256cf10944b451f51178978e2a874b32