LTX2 i2v - Sexy Move

THIS IS FOR I2V.

Only for use on generated and fictional adult aged characters. It does not contain undressing data.

I haven't even tried T2V at all, doesn't seem like it's going well. This dataset is specifically designed for i2v only. T2V requires too much effort for me. But, this test kind of worked I guess.

Some people can't make i2v work at all; this is the modified workflow I used on most previews. It uses the new latent normalization-will never distort an output contrast, won't give you that frozen motion zoom-in, and usually won't return buggy messy outputs unless there's some wild prompting going on. It has a lot of custom tweaks, and reasoning described on that page.

Don't expect porn or nudity. Nudity it still only cheated from the start image. This lora has no penis data, only some dildo riding that aids the hips movement and torso grinding. There are nsfw loras that are coming out that seem to work good with it, but the vocal data in this one needs improvement for 1.0, essentially this has to also be a 'sexy talk' lora at the same time. Dancing and most body motions still seem weak and need more data.

Strength: 0.2-0.8

It can interfere heavily with music and dialogue a 1.0, 0.8 seems good enough for all the motions. This is actually overtrained, and on purpose, just to test what data had the most weight. Results were: voice and self touching, dancing and other motion not as good.

Prompting:

Just put this at the start of prompt-

Style: realistic - sexy.

Style: realistic - explicit.

sexy: dances and general motion with breast rubbing.

explicit: focus on genitals, low angle, genital touching, lots of moaning.

That's there because afaik that's an actual LTX2-style dataset tag with Style: style or video in the style of which even works without the lora. This is meant to be a low level sexy boost to cut through garbage outputs and hone the model along with other action loras.

Other concepts specific to this that seem to work are listed below. The lora is strong enough that you can actually get an output with just the style prompt if the image lines up well enough, it doesn't shortcut as much as the wan 2.2 version yet though it will be slow motion with just some subtle moves so it still requires a thought-out prompt and iterations.

1.0 will remove the audio portion since it will most likely interfere when used with other loras. This is supposed to be a small motion support used with other targeted concepts that establishes an overall movement baseline and tone. All the other little bonus data with the voice, pussy, and other stuff is getting pruned and given to it's own JOI/ASMR style lora. Should be able to then give more weight towards the hip bumps and gyration, and more bouncing support, along with some support for a female looking back over her shoulder when facing away since the base model doesn't actually seem too good with that.

Training tips:

Diversify your audio as much as video - I only trained 3 different female speakers, and it shows. Most outputs are hijacked with the same vocal inflection and tone of voice unless sent to a different language. The training mixed them together into one Australian sounding girl, not sure what that's about. I also trained only a few repeated background music tracks to test. Don't train music with vocals unless it matters. The music data will show up, but with this lora it is at random ignoring the tags. Not sure if they became related to other concepts, or that the tagging needs to be better, or caption dropout just kicked them out first. But, you definitely can train songs and music.

LTX2 learns i2v flow well at 512, 32 rank is fine - I think I could've probably added even more concepts to this, and will. The torso and belly dancing motions need to be reinforced, but for the actual v1.0 I'll attempt to add female masturbation as well as more behind motions, and a lot more vocal data. Also I need to hunt down the exact phrasing in LTX2 base model to describe these kinds of lower body motions (grinding, twerking, gyrating) because sometimes it doesn't seem it's clicking for it, even though there was tons of data for that.

Abliterated doesn't matter for this one- I didn't train with it. Using the normal v.s. uncensored gemma produces completely different results. Abliterated version ruins dialogue, but also seems to make it way slower and erotic if you want that effect. The hardcore porn output was done with normal gemma, so yeah. It seems to only hold significant impact with actual nudity introduction or establishment T2V side, which still isn't possible even with this. This is not a nudity or undressing lora of course. LTX2 isn't really censored just ignorant of the concepts, this lora establishes a lot of them.

Prompting keywords:

Visual:

Fully Nude.

she is doing a sexy move with her body

bouncy dance

sexy dance

slow sexy hip motions and hip gyration

moving her hips side to side

Hip wiggle

Feeling her body with her hands.

Touching

Dipping her hips backwards towards the camera while she is posed on all fours.

Turns around to face away from the camera

Turns side to side

Rubbing her breasts with her hands.

grabs her huge breasts and squeezes them together

She is bouncing and riding up and down

Facing the camera.

Static camera.

Seen from behind.

Perfect body

Curvy

Skinny

Huge natural breasts

Huge firm round breasts

Large breasts

Large round breasts

Average-sized breasts

Small breasts

Large nipples

Small nipples

Saggy

Firm

pussy

innie pussy

anus

ass

butt

woman (older)

fair-skinned girl

black girl

Asian girl

latina

Audio:

sexy voice

talking dirty

sexual moaning and erotic breathing

sexual gasps

erotic inhale

Background Music:

happy pop song

Hip-hop

Hard Techno

Light Techno

Dubstep-style

Rap-style instrumental

Pop song

Background sound is an empty room tone.

Music is playing from a speaker in the room.

---

Music ignores the prompt usually. It's prevalence will probably be toned down. Literally just a test. If you don't want audio at all, set the Audio latent to 1 frame.

I only had to tag these visual descriptors because of the way the i2v trains, you barely need to describe the subject unless it matters for following motion.

The more consistent results from the Wan2.2 version, even though this dataset was bigger and longer, stem from a lot of evolution on pre-existing data in that model-especially for the dances which the torso motions are lacking in this version so far. LTX2 team needs to train some vertical tiktok girls or something, might be more useful than the Mr.Bean animated show credit roll.