Version 2.0 - LTX 2.3/I2V/T2V
Version 2.0, LTX 2.3 ready, I2V and T2V
I run about 417 frames when I generate videos, the training set assumes video lengths of about 10-15 seconds to fit the motions trained.
If you make a good video with the lora all I ask is you post it to the gallery, share it as that is what helps me improve the lora in the next training session. I hope everyone enjoys this lora, any input you give me or suggestions ill take for the next version
I feel LTX 2.3 does better with a strong starting image (I2V) but both work. This version 2.0 was trained to do both T2V and I2V modes at the same time.
Version 2.0 was trained on 105 high res nude female videos in various positions and lighting conditions focused on female anatomy (breast jiggle, vulva, nipples, natural ass movement , walking, showering, masturbating etc). it should be a step up over version 1. further training and refinement on more diverse datasets will come next as this lora is improved. trigger word is 'LTXNUDES'.
Bouncing Tits:
"young nude woman jogging on the spot outdoors in public park, high-energy motion causing large natural breasts to bounce and jiggle wildly up and down with each step, heavy sag and rebound, erect pink nipples swaying prominently, wide pale areolas textured and bumpy"
Trigger word: LTXNUDES
Choose a workflow:
Work Flow 1: https://civarchive.com/models/2354193/ltx-23-all-in-one-workflow-for-rtx-3060-with-12-gb-vram-32-gb-ram
Workflow 2: https://civarchive.com/models/2445871/ltx-23-i2v-fp8bf16
Training captions:
close-up low-angle view of young nude woman on bed
smoothed shaved pussy
fingers parting slick outer labia and circling swollen clitoris
baby oil all over her toned body
actively thrusts her hups up and down in steady rhythm
duldo fully inserted into her shaved pussy
she energetically bounces up and down
breasts jiggling
striking large breasts
massages her large natural breasts with both hands
fingers circling nipples
she teases the camera by slowly running her hands all over her body
close-up of a beautiful woman's shaved pussy as she masturbates
small areolas and perky breasts
realistic dildo mounted on the floor
she firmly squeezes her large breasts together
stands in steamy shower under warm running water
intense pleasure expression
intimate erotic self-pleasure scene
both hands gently spreading vulva wide open, parting her labia fully to reveal pink inner folds
intimate risky public flash
using one hand to gently spread her firm butt cheeks wide apart
she slowly turns her body to the side in a sensual sway, hips twisting gracefully
one finger inserted deeply into her tight anus
fully exposing tight pink anus and wet shaved vulva from behind
large pale natural breasts with puffy pink areolas and small erect nipple
one hand sliding up along her side and cupping her breast teasingly
upward and over her large augmented breasts to fully expose them
nude woman on all fours outside
soft belly and wide hips
labia spreading motion, clitoris detail, arousal wetness, and wide-spread legs pose
thick outer labia majora glistening with arousal fluids
reaching back with both hands to spread large plump buttocks wide apart
close-up low-angle view of young woman nude from waist down
Description
v1.0 basic female nudity
FAQ
Comments (25)
"abliterated gemma" plz noooo
try it with the non abliterated gemma please, im curious.
Is abliterated Gemma not good?
@candid80 I havent had any issues with it, I use it for all my LTX-2 nsfw videos, I just meant if someone doesnt want to use the albliterated gemma they are welcome to try it too
@candid80 The abliterated Gemma isn't exactly bad, but it's not really necessary. The standard Gemma doesn't censor inputs anyway when used as a text encoder.
@g1263495582 interesting and does the abliterated gemma effect the video quality vs using the stock gemma?
@g1263495582 Well, that's not really true.
@sexgod1979 No, not really
@sexgod1979 Honestly, in terms of quality, there isn't much difference. Or maybe I just don't see a huge difference yet.
@kossan explain ?
only usefull if you use the prompt enhancer with a nsfw image attached, surely you can goon your own prompt with standard gemma.
@Protagonist_NL And that enhancer sucks anyway. It takes a year to process.
what about the heretic instead of abliterated?
abliterated gemma does nothing other than ruin speech accuracy and probably other things, it won't magically make the model output NSFW, standard gemma isn't censoring your prompts as a text encoder.
@rogerstone442 That's not true. I have multiple seeds trying my lora where the nudity output is completely different between them and much more close to normal with the abliterated. The normal TE tries to reduce erotic themes, movements, sexual voices, and nudity for sure.
@tenstrip your normal text encoder doesn't(won't) understand nsfw slang, just prompt it with accurate medical terms
That looks great! Could you please send more example videos of T2V/I2V?
will do!
When it comes to ltx do you use this for model 1 and model 2? If so what strengths?
you need the cameras, nsfw loras in both the model1 and model2 path, the detailer lora goes only in the model2 path
@sexgod1979 thank you. how about the camera loras? i assume only in stage 1 right? or does it need it in stage 2 so details can be generated correctly as the camera moves?
@Jdoe666 No, static camera should be in both stages. so stage1 -> distilled lora -> nsfw lora -> static camera, then for stage 2 -> nsfw lora -> static camera -> detailer lora
@sexgod1979 Any chance you could share an example workflow for this? I'm not sure if the workflow I'm using just doesn't use different loras at different stages or I'm misunderstanding what needs to be done here
@sexgod1979 Curious, the official workflow seems to use the distilled lora for stage 2 and not for stage 1. Going to try your proposed lora setup.
@MrTwister242 Yes thats done on purpose. STage 1 is where the heavy temporal motion coherence happens, GGUF Q8_0 doesnt need that push in stage 1. IN stage 2 the distilled lora provides the better convergence at low steps. Several community members on Kijai tested this as well and found this worked better. Its beasically using it stage 2 gives better rate dev motion/coherence in stage 1 and then the distilled sharpness/polish in stage 2