What you need loaded

base model PinkCherry (MiniMax H3 fine-tune) — this is what it was tested on this LoRA strength 0.5 second LoRA HMPenis by HearmemanAI, strength 0.60 guidance 1.0 — H3 is guidance-distilled. Raising it breaks the image duration 14–15 seconds

The HMPenis LoRA is close to required, not optional. H3 draws poor male anatomy — odd shapes, a stray split or line down the shaft. PinkCherry improves it a little on its own; HMPenis is what makes it look like an actual penis. This LoRA lives or dies on him being visible and in frame, so the two go together.

You do not need its trigger word in your prompt — just load it. Chain it after your turbo/speed LoRA. If your workflow keeps LoRAs inside a subgraph, you have to go into the subgraph to add it.

Thanks to HearmemanAI for that one.

⭐ Read this first: write the prompt to match your picture

Look at your start image. Describe what is actually in it — the camera angle, the pose, whose hands are in shot, what is behind them. Then say what should change.

Anything you describe differently from the picture is an instruction to redraw it. When H3 redraws, it re-derives the people too. That is when the face changes, the skin tone shifts, or the shot flips to another angle.

Copying 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.

⭐⭐ Getting a start image is the hard part. Budget for it.

Be warned: this is by far the most difficult and slowest step, and it is not close. Generating the video is easy once you have the frame. Getting the frame can take a morning.

The reason is that image models overwhelmingly want to draw penetration. Ask for a woman on top and you get full sex, near enough every time. There is no widely available image LoRA for this position, so you are fighting the base model's prior with tags alone.

What works:

• The Danbooru tag is sumata. A Pony checkpoint plus a cowgirl-pussyjob LoRA gets you in the area.

• Ask for the position positively — sumata, shaft pressed flat between her labia, riding along the length of it. Do not write not inside her; negation produces the opposite every time.

• Put the wrong poses in the negative prompt: penetration, vaginal, reverse cowgirl, from behind, facing away. The negative field is a different mechanism from the prompt and it does work here.

• Scene words flip the pose. Adding night, outside or cityscape will turn her around into plain or reverse cowgirl, because in the training data those settings sit next to those poses far more often than they sit next to sumata. Get night from lighting words instead — dim blue light, single lamp, deep shadows — and let the video prompt supply the setting.

• If it still flips, generate the pose somewhere it works and inpaint the background afterwards. The pose stays bit-identical because you never re-roll it.

She has to be facing him. Facing away is not in the training data at all, and the model has nowhere to attach the anatomy — it comes out bent into shapes that are not anatomy. Front-facing only.

The clearer the start frame reads as outside, the better every render after it.

The prompt

Structure: a look line ending in "The environment is constant throughout.", then SHOT 1: with the action, then Audio: last.

Put PJSLIDE at the start of the action. The HMPenis LoRA needs no trigger word — just have it loaded.

Full example — the whole arc in 14 seconds

Slide, a kiss, dialogue, and the finish, all inside one shot. This is the prompt the demo clip was made with.

Close side view in a bedroom, soft natural morning light, tight framing, his cock and her hips large in frame, shallow depth of field, plain bedding behind them. The environment is constant throughout. The camera holds one steady close side view throughout. The woman is 25, riding her husband as they wake up in the morning.

SHOT 1: PJSLIDE. She rocks her hips and slides her pussy along the length of his cock, the shaft pressed flat against her lips, closing her eyes and smiling down at him.

She leans down and kisses him on the mouth, then sits back up and keeps grinding along his length. She whispers, "good morning sweetie, did you sleep well". He says, "yes Abby, this is a lovely way to wake up". He rubs her breasts squeezing them tightly.

She rides him faster, sliding along him, and he cums, and a light dribble of thin clear cum spurts out over his stomach. He breathes hard as it spurts out.

Audio: wet skin sliding, their breathing, quiet room tone, no music.

One shot, not three. Events fire in the order you write them, so you do not need cuts to sequence them — and every cut spends seconds on something that is not the act. The paragraph breaks are just for your own eyes; the ordering is what does the work.

14 seconds, not 12. The whole arc needs the run-up. At 12 the finish gets crushed into the last moment.

The three trained sentences

Keep these close to verbatim and everything around them can vary.

what sentence her doing the work She rocks her hips and slides her pussy along the length of his cock, the shaft pressed flat against her lips. him doing the work Her hips are lifted up off him and he rubs the head of his cock up and down along her pussy lips. the finish He cums, and thick white cum runs out and spills over his stomach. He breathes hard as it spurts out.

What one small change does

change result He holds still / she pauses anywhere Nothing happens. Never write stillness He cums, and thick white cum runs out and spills over his stomach. Maximum volume — it comes out like a soft-serve machine. This is the trained sentence word for word, so it fires the concept at full strength. Use it deliberately if you want it over the top He cums, and a light dribble of thin clear cum runs out over his stomach. Realistic texture and volume. All three words matter — light dribble, thin, clear. This is the one to use thick white on its own opaque and heavy mostly clear with thick white in it the in-between, closest to real over her pussy and stomach twice as much. The model fills whatever area you name onto her stomach one destination, sensible amount he pushes hard against her at the finish force reads as volume. More cum She rides him faster before the finish builds the pace into it instead of jumping a spoken line lip-syncs on its own, and does not stall the motion she moans in the Audio block continuous, out of sync with the action she moans in the prose happens at that point

The three volume dials, in order of power

The first two rows above are the same sentence with the volume words swapped. You do not need to touch the strength slider to move between them — this is a prompt problem, not a bake problem.

1. Quantity words. a light dribble of thin in front of it. Drop them and there is nothing limiting the amount, whatever else you write.

2. One destination, not two. Name one place for it to land.

3. No push words at the finish. Pushes hard, slams, drives all read as force, and force reads as volume.

Say where it comes from, and say whose it is

Without an origin the cum appears out of nowhere in mid-air:

a light dribble of thin white semitranslucent liquid runs from the tip of his cock down onto her stomach

And watch the subject of the sentence. This one put it out of his mouth:

He groans and a light dribble of thin white semitranslucent liquid runs onto her stomach.

He is the subject, and the last thing he did was groan — so it came from his mouth. Split them, and make his cock the subject of the sentence that emits:

He groans loud.

His cock twitches against her and a light dribble of thin white semitranslucent liquid runs from the tip of his cock down onto her stomach.

Two rules worth memorising

Audio block = continuous ambience. Prose = timed sounds. Anything in Audio: plays across the whole clip. A sound that should happen at a moment goes in the prose, next to the action that causes it. This is precise enough to land a sound on a specific beat.

Never negate. "Less cum", "no penetration", "without going in" all produce the opposite. Ask for the small version positively instead.

A spoken line needs a beat after it. Speech takes real time to deliver, so a line written directly before an event lands after it — "oh my god, I'm gonna cum" arriving five seconds late. Put one short action sentence between the line and the thing it announces and it comes out ahead of it, which is what you wanted:

He says, "oh my god, I'm gonna cum".

She rides him harder, slamming down on him, and he grips her hips.

He groans loud and pushes up hard against her, and a light dribble of thin white semitranslucent liquid spurts out over his stomach.

What it does well

The camera moves on its own, and it looks right. It will drift to a side angle to show the grind from the side, then ease back to frontal. Nobody asked it to; it reads as deliberate camerawork rather than drift.

It takes direction between the beats. A kiss, a look, a line of dialogue in the middle of the slide all land where you put them, and it goes back to grinding after.

Sound. Wet sliding and breathing come through cleanly.

It holds the position with her on top. She stays outside for the whole clip.

Known behaviours, and how to prompt around them

⭐ Don't ask for something that isn't in the start image. The big one. If the prompt introduces a hand, an object, a second person or a change of clothing that is not in your start frame, H3 has to invent it — and introducing an element is itself a transition, so it re-derives the whole picture to make room. That is when the face changes.

Either pick a start image that already has what you are describing, or don't mention it.

The face or skin tone changes when the camera moves. Same cause. Fix, in the look line:

The man and woman from <Picture 1> in their original scene: <describe the scene>. Her face, her hair and his skin tone stay exactly as they are in <Picture 1> for the whole clip. The camera holds one steady close side view throughout.

<Picture 1> is H3's own reference tag and works with plain image-to-video.

The first second or two is a different person, then it settles. Trim the front off. The rest of the clip is unaffected — not worth re-rolling.

Too much cum. That is the finish sentence, not the strength. Swap it for the light dribble line before you touch the slider.

The action loops or repeats. Add a count — she kisses him once rather than she kisses him.

Him standing over her is the harder half. She-on-top is 1391 clips; him-active is 168. It works, and it works well, but it needs more holding:

• Give his hands a job. both hands on her hips keeps them somewhere. Left loose they drift back onto himself.

• Don't write holding himself in his hand — it renders as stroking, not grinding. Put the motion in his hips: he rocks his own hips, grinding the length of his cock up and down along her pussy lips.

• Repeat the placement every paragraph. the shaft pressed flat against her lips, riding along the outside of her. Say it once and it drifts into penetration around the halfway mark — the base model's prior for a man standing over a woman is very strong.

Don't write he lets go if a hand is on him in the start image. That hand is what anchors the anatomy to his body; remove it mid-clip and it can detach and follow her. Either keep the hand, or start from a frame that has no hand on him at all.

How it was made

1908 clips from 13 sources, all hand-narrated — where the slide is actually visible, where he finishes, and who is doing the work. Trained on MiniMax H3 with audio, rank 16.

The file here is step 4000, which is about 2.1 passes over the set. That sounds far too low and it is not — H3 reaches the behaviour in a fraction of the passes a 2.x-generation model needs. 5000 also works and is slightly stronger; above that it started losing the texture, and below it the slide gets vague.

The full checkpoint ladder is mirrored on Hugging Face if you want to try the other rungs yourself.