DR34ML4Y by DR34MSC4PE
This LoRA, incredibly, manages to handle ALL of our specialty LoRA concepts - Double/Single Blowjobs, Missionary, Reverse (and regular) Cowgirl, Doggy, wrapped in one model! It obeys very well and is largely extremely accurate to the given keyword/position (see samples below)! Don't be shy; give it a try! =)
===========================================================================
8/7 - H3 training in progress...training
6/27 - v3 LTXXX - Out Now! Come 'n get it!
TRAINING IS STILL ONGOING - You may experience some concept imbalance as we make changes to the dataset. If you notice bad audio, we recommend NOT using the abliterated text encoder as it seems to cause this effect to be extremely noticeably during speech!
Please note - read the model description and understand that the trigger words STILL WORK but your surrounding prompts will need to be much more verbose in the style LTX2 requires!. Performs BEST WITHOUT DISTILLATION. The distillation lora and checkpoint actively fight the nsfw training and account for body horror. YMMV - we recommend using the full dev checkpoint with between 0.25-0.35 distill strength. Adjust steps and CFG accordingly for both passes.
EXAMPLE WORKFLOWS HERE! Shout out to RuneXX!
AS ALWAYS: Enjoy, generate responsibly and share what prompting tricks you learn along the way!
==========================================================================
Sample Prompts (best for WAN, LTX needs updates):
Note: prompts w/curly braces are dynamic prompts for use with Wildcard nodes, remove them for regular use like the other examples.
Note - Missionary and cowgirl can have some mix-up if you don't specify the woman is on her back in the prompt.
m15510n4ry, A shot of missionary sex and a detailed view of a woman's vagina and a man's erect penis which is penetrating her. The woman's skin is light and the woman having sex has natural pubic hair above her vagina. The penis is erect as it thrusts in and out of her vagina. There is a slight sheen of moisture on both the penis and the surrounding skin. The background is a fancy bedroom where she is at the edge of the white bed laying on her back and he is between her spread legs. The lighting is even highlighting the natural skin tones and textures.
bl0wj0b, A close-up view of a single {tiny|emasciated|young|18-year old|old|cheerleader|pale goth|petite|rail-thin|plump teen|fat|full-body} woman with performing a blowjob on a man's erect penis. The girl is nearly naked except for {choker|leggings|a slutty belt|crotchless panties|a microkini} barely covering her {white|black|olive|fair|tan-lined|thai} skin and {small|huge|a-cup|DDD} breasts and {glittery|pink|black|rainbow} hair. Her nails are manicured with a {glittery|pink|black|rainbow|natural} design as she grasps the base of the penis and uses it to guide it in and around her mouth as she sucks on the mans {penis and testicles|shaft and balls|erect penis|penis tip as it dribbles white-clear semen from the tip and onto her tongue and lips}
d0ubl3_bj, A close-up view of two fat girls with performing a double blowjob on a man's erect penis. The girls are nude save for slutty choker collars, the girl on the left has pale skin blonde russian slutty pigtail hair. Her nails are manicured with a glittery design as she grasps the base of the penis. The other woman has light thai tan-lined skin and skin black hair with bows and her eyes are closed. Both women are focused on the man's penis which is circumcised and erect. They both lick and run their lips along the shaft of the man's penis. One womans hand moves slightly up and down the base stroking his penis as the suck and lick and share it. The lighting is cinematic yet voyeuristic.
c0wg1rl, a {4k UHD footage|iphone video from 2000s|vintage 1970s|futuristic 8k} {pov|side-shot|top-down xxx pov|4k video} {tiny|emasciated|young|18-year old|cheerleader|pale goth|petite|rail-thin|plump teen|fat|full-body} girl with her arms behind her back and legs spread apart in the reverse cowgirl position. She has {white|black|olive|fair|tan-lined|thai} skin {small|huge|a-cup|DDD} breasts and {long straight|wet|pink|black|rainbow} hair. Her eyes are closed and she has a slight smile of {pain|mischeif|embarassment|lust|sleepiness|possesed|evil} on her face as she {opens|fixes|rolls|focuses} her eyes at the viewer. Her vagina is fully exposed and her legs are slightly bent at the knees. At the bottom of the frame a man's erect penis is thrusting into her vagina.
d0gg1e, A footage|iphone video from 2000s|vintage 1970s|futuristic 8k} {pov|side-shot|top-down xxx pov|4k video} of a {tiny|emasciated|young|18-year old|cheerleader|pale goth|petite|rail-thin|plump teen|fat|full-body} woman {bent over her bed at the waist with her stomach and chest flat on the bed|on her hands and knees|bent over presenting her pussy and anus|bent at the waist reaching toward the ground with ass up|face down, ass up}. She has {white|black|olive|fair|tan-lined|thai} skin {small|huge|a-cup|DDD} breasts and {long straight|wet|pink|black|rainbow} hair and is having her {tiny pussy|butthole|anus|vagina|puffy little vagina|hairless pussy|hairy vagina|4K pussyhole} penetrated in and out repeatedly by one man's modest sized erect penis, slowly but totally in and out as she looks back toward the camera and POV man fucking her. The background is {other young ladies gather around to gain the experience of watching her suck the penis, watching in lab coats that are open exposing their naked bodies in the background|in a church confession booth}
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (trainer/captions/tooling code/data/qa)
Be sure to check out our FLUX model - C4PACITOR!
Please generate responsibly!
Description
Fine tuning continues - likely to be the second to last as we near concept maturity. Doggy and tongue interactions still need some balancing and we're trying some other stuff to bring the final release ASAP.
FAQ
Comments (43)
this looks good but tbh im honestly tired of fighting LTX every second just to get a decent output.
These are legit regular ass T2V prompts. Temporally done, like 0-5s: describe the action 5-10s: now describe this action, etc etc
@c0ur4ge Oh? thats not bad at all. but how dose it fair on i2v?
@Ragamuffin20 That is a popular source of the images below, tbh. So... very well I'd say! :)
@c0ur4ge its good, but no wan
i respect the work you guys do; you've made some of my fav loras of all time, but honestly, I have to agree with the sentiment that ltx is basically trash for most things. The only halfway decent stuff I've ever managed to get out of it was v2v extend type things. I'm not even talking about nsfw/complex content. Just basic sfw stuff. It's a real hassle :(
@mjh02111964582 I agree the hill is steep and, as usual, we look at this stuff as a challenge. So no worries, if it's not your thing, the WAN stuff still works great and we've got some other things baking too. But a lot of this is evolving the dataset and practice!
I have to back this sentiment too. I certainly appreciate the hard work that is put into creating these LORAs, however I just think the model is trash. One has to generate too many samples to get a somewhat decent output, it takes the fun out of it. Hopefully someone in the future comes out with a better model that incorporates audio.
@toty21 yeah what it is is that ever since they started making laws to crack down on AI regulation especailly deepfakes companies like lightrix and the guys that made flux make sure there models are difficult to use for NSFW purposes in order to avoid backlash from a certain group of people. in short its politics. the same reason why grok, who was once unrestricted, now censors bikini images on the spot while chatgpt and gemini semingly have no problem taking and editing images like that.
@toty21 I wouldn't look at my output as representative of "doing it right either", tbh. The lora is capable, but the people in the comments below are the ones getting the good stuff out of it and it frankly requires some different workflows and prompting techniques. With that, and again, sticking to the dev model vs. distilled makes a world of difference.
@Ragamuffin20 Which means, don't delete your local models as they might be the last you'll get until the puritanical fascists behavior comes to an end.
@DaddyWolfgang i dont see it coming to an end tbh. these are the same people screaming for their genders and pronouns to be acknowledged
i find it extremely difficult to get talking out of this lora. is there a way? i use i2v and nobody speaks a word
I'm going to bet there's alot of conditioning on your image that's doing this more than anything else, tbh. I have no such issue, ever. Works almost every time. They will occasionally slur or say the "wrong" part of the prompt.
The ltx audio is cooked/overtrained. Do you just train them together?
Initially yes, and it's further than I wanted. They've had seperate LR for most of training. For now, do not use abliterated TENC and I'd account for echo or any artifacts in negatives which seems to mostly work. Upcoming release contains an interpolated version that is better but trying to iron out the last of the doggy/bj tongue morph stuff (if even possible).
@c0ur4ge just a random thought, might be silly or already done, did you ever consider using Qwen TTS 12Hz 1.7B? I'm not a professional creator like you so forgive my ignorance, if so.
@freestuffpl0x42069 In what capacity? I personally don't think they really do a ton more than move you further from the latent space the stuff was trained on. As always, YMMV.
@c0ur4ge It seems that my comment didn’t post. After researching I realized that was a dumb thought. I’m glad that you figured it out because it’s 1 of 2 things:
Abliterated model
Quantization issues
I'm using Wan 2.2 I2V with this LoRA (DR34ML4Y), along with the Lightx2v 4-step acceleration LoRA. The result is that the video is in slow motion, and the penis is floating and detached from the body. My starting image only has a woman and no penis at all. Is it required to have a penis in the reference image for this LoRA to work properly?
When I removed the Lightx2v LoRA and generated again, the motion became normal and the blowjob action + penis shape looked good. The only issue was that the penis was attached to the man's chest instead of the correct place. Has anyone run into this problem before? Any solutions? thank you so much！
It's absolutely not required to have a penis already present.
You need to adjust the weight of the loras you're using and I'd recommend using the best quality light 4step you can get away with. Look for the 1030 high and 1022 low Lightx2v Loras.
The model you're using matters a lot too, DaiSWA's models are basically the platinum standard for WAN2.2, you don't use the lightx2v loras with his lightspeed versions, only use those loras with the "TrueVision" checkpoints as those don't have speedup loras build into them.
It sounds like you're trying to use Lightx2v loras with models that might already have them build in.
If you want to render a man entering the frame with a penis use:
"A man enters from the (location) with an erect penis and inserts it into the woman's pussy/anus/mouth/nostril/ear."
Describe the penis, the testicles, and if you only see the man from the waist down. It helps to describe things as you want to see them.
I'd recommend using some sorta of uncensored language model for creating prompts for you.
Try the 256 rank distilled speed-up loras for both high and low.
Yea the audio is cooked, even using the non-abliterated text encoder the audio is broken. The sex action is great though.
I've.... not had a whole ton of issues personally but I can see about interpolating the audio components with an earlier version to see if it cleans up artifacts. I generally put "slowmo audio, flange, reverb" in the negatives, fyi.
so for i2v i don't want to have the penis visible just like the girl straddling on the man without having additional penis showing up. what's the workaround? i'm on gguf cfg 1
will this work when its mixed with 10eros LTX checkpoint model ?
Yup!
Can just having this on without trigger words help with NSFW? does it pick up on other general NSFW terms?
Yes. And yes.
A fisting lora would be cool
I'm trying to use it with ltx i2v reverse cowgirl and can't get anything good out of it, she barely moves or she moves completely off and just looks horrible as in body horror. Probably works better as t2i?
does handjob work?
I can confirm at least for the Wan version that it does work very well for handjobs!
@zym0x and also for i2v?
@dominik1996d926 yes, i only do i2v
@zym0x can you give me a universal prompt that i can use with a prompt enhancer?
I like it but the lora keeps changing the faces of the anime girls into some real/cartoon hybrid lots of morphing? Is there a way to fix this so it's more in line what i see here?
hmm, why does the dr34mly doggy's POV vagina from top down look like the front pussy. Looks like it has a labia.. not that great aesthetically. Using Wan version.
You need to be very specific about the camera angle - our dataset is so if you use contradicting terminology, this can happen. Especially true in with LTX that really struggles with rear ends.
ok I'll try to be more specific.
On the LTX2.3 10EROS model, the understanding and effectiveness still aren't great overall. While it does a good job at improving prompt comprehension for the 1.4 version of 10EROS, it feels like the training for poses and body positions isn't sufficient yet.
For example, the missionary position(腕掴み正常位)where the hands are pulled straight and pressing against the chest is very hard to get right with prompts. Breast groping also doesn't work properly — the girl often ends up touching herself instead.
Even though it's not perfect yet, I'm still very grateful for the LoRA you provided. It has significantly increased the playability of LTX2.3 10EROS. The original model had terrible prompt understanding, so your LoRA really saved my experience. Thank you, and I hope it continues to improve in the future!
The above text was translated using translation software. If there are any inaccuracies or unclear parts, please understand it with similar meaning. Thank you for your understanding.
yeah, it’s always gonna perform weird on models that are also derivative of the base model. glad it works well at all there!
i compared 10eros and this lora. this lora is better at retaining the original face (in i2v workflow) than 10eros. but then i tried base ltx2.3, the face is 100% retained. is there a way to train this lora with blank faces so that the original face is always retained?
Details
Files
DR34ML4Y_LT3X_V3.safetensors
Mirrors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y - All-In-One NSFW v3.0 - LTX2.3 - m15510n4ry,bl0wj0b,d0ubl3_bj,d0gg1e,c0wg1rl,r3v3rs3_c0wg1rl,0.25-0.35str.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y - All-In-One NSFW v3.0 - LTX2.3 - m15510n4ry,bl0wj0b,d0ubl3_bj,d0gg1e,c0wg1rl,r3v3rs3_c0wg1rl,0.25-0.35str.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y - All-In-One NSFW v3.0 - LTX2.3 - m15510n4ry,bl0wj0b,d0ubl3_bj,d0gg1e,c0wg1rl,r3v3rs3_c0wg1rl,0.25-0.35str.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y_LT3X_V3.safetensors
DR34ML4Y - All-In-One NSFW v3.0 - LTX2.3 - m15510n4ry,bl0wj0b,d0ubl3_bj,d0gg1e,c0wg1rl,r3v3rs3_c0wg1rl,0.25-0.35str.safetensors