Can be used with LTX2.3 at higher strength but the audio quality will be worsened. Try using custom loaders and don't add audio keys.

This a low rank helper/additional nsfw Lora. It's a nsfw helper lora for nsfw helpers. It is not captioned or deeply trained enough for consistent solo use outside of a few cases, and it's designed around shortcomings with other nsfw helpers that lack enough influence to make the man thrust towards a receiving partner in a sexual scenario. Use preprocess compression strength of 34-38 to give it enough freedom to draw faster motions and body movements. It took me longer to test and figure out how to use it than it did to caption and train. All the previews use it at 0.8-1.0 strength with other loras around 0.2-0.4 strength, and that's generally a good ratio for use. Tbh the other loras don't do much other than slightly reinforce some motions, but mostly they just stabilize this one and temper some of it's issues.

If you don't have this one, it's useful for POV images with this.

The other one I usually use with this is the better nsfw motion, but I can't attribute it.

Those both will take over and induce riding motion if set too high, so you only use them at a very low strength to counter the wild overfitment of this.

The vocals are also incomplete until I caption and train the voice lora I plan to do, and that will work alongside it to fix the bad LTX2 default voices which are incompatible in these scenarios. Right now all you get is occasionally some good slapping sounds and breathing and moaning. I use my oral suite and test version of sexy moves loras just to change the female voices.

Workflow I used. Right click and save as a .json in your workflows folder. I dumbed it down to make it simple. The key to that workflow is that you use the dev model with distilled loras to overmanipulate the first pass to have good audio and a lot of motion. The output video from that first pass is for motion previewing only, saving you time from letting it go all the way. If the motion looks bad and the audio is bad, cancel it and refine. The second pass takes the full sized image with no compression and animates it with the upscaled latent from that first pass with very little input, giving you a high quality visually restored version with the audio from the first pass forced and tracked into it. You can't pass the audio latent into a half-strength distillation, and you can't full strength distill a res2s sampler on the second pass without it getting distorted; so that's the compromise. That's why my LTX2 videos look very different than others.

With the new LTX2 Guiders, the main parameters you'd want to mess with are increasing CFG between 1-1.5 to see if you can get more out of prompt. STG is the main increase, you can safely leave it at 1.5, or got to 1.8. in layman's terms STG strengthens the sensibility of the model and will reduce weird things happening (heads twisting around, mutations, weird stuff) so you'll want to try it upwards. For audio use 1.5-2.0 CFG like I leave it by default, and use the negative prompt mainly for audio terms.

Because it's LTX2 and still early on usability really starts more about issues that can happen. Both with this Lora and in general:

Not enough motion or figures a locked in place stiff and then it's cartoonish/weird and/or they just burst out laughing: This happens when the model gives up on trying to animate an image i2v, or if preprocess it too low. Increase preprocess strength to 34-38, change resolution and aspect ratio, reroll seed. Stack more loras, or turn the strength up.

The penis is like 1-2 feet long: Caption it's length. You usually always must do this if turning preprocess strength up. For some reason, that works. The man's 7-inch penis, the man's 9-inch large penis. Metric probably works too. LTX2 is actually more sensitive to prompt than it seems.

The motion is crazy and they're flying around moving around like mad: Lower lora strength to 0.75-0.8, lower distilled lora to 0.8-0.9 if using that, lower CFG, raise STG slightly if using that, lower Preprocess node strength, remove captioning that would inspire the model to make crazy poses and movements.

The part of the penis from the image just ends up terraced with a some cylinder extension coming out of it: Increase preprocess strength, change resolutions and format, reroll seed.

Other problems and quirks that I haven't fixed and have no idea why it has them:

The camera will suddenly shift, the pose will suddenly change: When you train to overfit but still use Differential guidance in overfits unevenly. I think the camera thing is just from a few POV head-mounted camera shakes, and the pose change is uneven training over fitment just rigidly conforming the figure into a pose it prefers more.

The woman's eyes always close: It's kind of realistic, but it is also a skew in the data I didn't realize was there. Try prompting for looking at the camera and eye contact.

You can prompt with this these dataset prompting structures, or just ignore it and prompt based on other higher ranking lora captioning. It doesn't really matter. A normal prompt would be:

(camera angle), (scene description), (figure pose description), The man fucks her. His penis (goes into, slides inside, penetrates, goes in and out of) her (vagina/anus). (other evolving actions) (dialogue) (sound description).

The more robust and the more descriptions you stack the better the outcome. Also mix in other lora tagging.

I also try this follow-up moodboard prompt at the bottom of some prompts, just to pad out tokens and make the model behave more on-point by giving it some kind of prompt-enhancement. It's not Lora-specific but the entire thing pretty much activates all the keywords so you can only prompt for camera, setting, audio, and speeds before it:

Overall tone: NSFW video centered around sex, extremely rude and explicit, uncensored, raw sensual erotic pornography, full nudity, detailed genitals always visible, hardcore sexual themes, adult content only.

Voice tone: Breathless, sensually overwhelmed, high-pitched cute female moans turning into desperate gasps and broken whimpers, orgasmic cries, erotic panting, heavy exerted breathing, mixed with soft moans, male deep guttural groans and grunts of pleasure, voices filled with lust and love.

Audio queues: When bodies contact there is a soft slapping sound, objects the enter inside body-holes are slippery and make wet sounds.

Motion themes: Ultra fluid-smooth realistic motion, hypnotic male pelvic thrusts causing testicles to jiggle, natural female body jiggle and bounce, realistic flesh physics, correct anatomical soft and hard penetration, detailed skin-on-skin contact, love-filled passionate scenario, poses stay and keep posed around motions.

You can just slap that at the end of the prompt and see if it helps or edit it with less/more.

Angle descriptions:

POV. The camera captures the man's point-of-view during sex.

Seen from a low angle.

Seen from behind.

Seen from the side in side profile.

Seen from underneath the woman.

Seen from a side angle. Angled side view. (3/4 angle)

Seen from above.

Pose descriptions:

doggystyle

missionary position

cowgirl

from behind

POV missionary

Lying on her back.

Crouched with her legs open and a man lying down below her.

Lying down with a man next to her.

Speed descriptions:

Fast rapid thrusts

Deep and Slow

Average pace fucking

Then, the speed and pace increases

Then, the sex gets slower.

Movement descriptions:

Sex.

Anal sex.

The man fucks her.

Fucking.

Her breasts jiggle.

Her butt jiggles.

Her floppy breasts jiggle excessively.

The man's testicles bounce.

The penis goes into the vagina with penetration.

The penis slides inside the vagina.

Voice descriptions:

The woman says during sex:

The man says during sex:

with quick line delivery

orgasmic voice tone

shrieking voice tone

cute voice

whisper

heavy breathing

erotic moaning

"..." followed by moaning

"..." followed by a naughty giggle

Sounds:

Repetitive slapping sounds from skin contact.

Wet repetitive sex sounds.

The woman is breathing quickly and loudly.

She is moaning erotically.

Exerted breaths.

This is only trained on only male thrusts. Something in the base model influences females to squat up and down already, and when most of the nsfw trained on riding action takes over anything else and made the man's actions harder to get right. That's all this tries one was trying to improve. It kind of accomplishes it, but in the same usual inconsistent LTX2 way. I think after this one and seeing some other results, LTX2 training is actually quantity over quality. It really needs to see multiple hundreds of very similiar data examples at low repeats to grab just that and not latch on to other data.

The partner is also gender neutral and this is also trained on anal with male anatomy nearby, so futa and male on male is also enabled with some right prompts/image/and lora mix. All this does is zero-in on the man thrusting in multiple scenarios, speeds, and angles. They're not all fully covered, but if you load this alongside other pose loras and nsfw helpers it will certainly go farther towards that effect with the same amount of tweaking and adjustments you can get it right. It has audio, but the different training settings I tried here seem to be extremely weak on training audio robustly so it has very low interference, and also very low audio effect. The right sound is usually brought out with in combination with other similar loras.

Just like Wan the real decent nsfw results in LTX2 come from Lora stacking, they just make it more stable and consistent by over-influencing the base model's nonsensical off-prompt inputs. But, because I gave differential guidance another try (won't be using it for LTX2 anymore) this lora will add it's own nonsense sometimes. DG overtrains and gets the concept in, but it also overfits literally everything else in the data. Tattoos, weird pose changes, random hands. It still is usable because other loras will tone that stuff down and fix it.

For T2V this is motion-only influence; this is useless until there are well trained nsfw finetunes, and they might not need this. There is currently one being trained that may work.