CivArchive
    LTX-2.3 Deepthroat - v0.1
    NSFW

    LTXdeepthroat — LoRA for LTX-2.3

    A LoRA for generating oral/deepthroat content with LTX-2.3 video models. Test workflows in training data.

    Trigger Word

    LTXdeepthroat

    • LoRA strength (Stage 1) 1.0

    • LoRA strength (Stage 2) 0.85

    • Distilled LoRA (Stage 2) 0.6

    Prompting Tips

    This LoRA responds best to literal, descriptive prompts. Describe what you want to see as if you're directing a camera operator. Avoid poetic or abstract language.

    Do: "She slides her lips forward along the shaft toward the base" Don't: "A mesmerizing rhythm of passion and desire"

    Vocabulary

    Use these terms for best results:

    • glans — head of the penis (avoid "tip")

    • shaft — the length

    • base — where the shaft meets the man's body

    • slides / glides — when the woman controls the motion

    • thrusts — when the man controls the motion

    Getting Better Results

    • Describe the male body if visible.

    • Specify the camera angle — POV, profile, low angle, etc.

    • For deepthroat scenes add "lips sealed around shaft" to prevent unwanted tongue.

    • Describe skin — tone, undertone (warm/cool), texture. The more detail, the better the render.

    Example Prompt

    LTXdeepthroat, a first-person POV looking down at a woman with long blonde
    hair and fair skin. A man's bare torso is visible with natural skin texture
    and light body hair. Her lips are sealed around the shaft, sliding slowly
    forward toward the base. She pulls back, the glistening shaft reappearing
    as the glans emerges between her lips. Static camera, soft warm lighting.

    Known Quirks

    • Male torso needs explicit description or it gets amorphous.

    I2V system prompt i use.

    You are a prompt writer for an AI video generation model. You will be given a reference image. Extract the visual details and write a generation prompt that would produce a video with a similar look and feel, but with motion added.
    
    You are NOT captioning the image. You are writing a CINEMATIC DIRECTION that borrows the image's visual DNA — the specific colors, textures, materials, lighting mood, and character details — and adds motion to bring it to life.
    
    Always begin with "LTXdeepthroat,"
    
    EXTRACT WITH SPECIFICITY — every noun needs a visual adjective:
    - NOT "brown hair" → "shoulder-length chestnut hair with copper highlights, slightly messy"
    - NOT "fair skin" → "porcelain skin with a cool pink undertone, faint freckles across the bridge of her nose"
    - NOT "a bed" → "rumpled dove-gray linen sheets on a low platform bed"
    - NOT "soft lighting" → "warm amber sidelight from the left, casting a soft shadow along her jawline"
    - NOT "a man" → "a broad-shouldered man with a deep olive tan, dark stubble, and a faded tattoo visible on his forearm"
    
    PULL THESE FROM THE IMAGE:
    - Hair: color with a modifier (ash-blonde, honey-brown, jet-black), length, texture (tousled, slicked, wavy), any accessories
    - Skin: tone + undertone + surface quality (dewy, matte, flushed, sun-kissed, glistening with sweat)
    - Eyes: color, makeup (smudged liner, clean lashes, heavy shadow)
    - Body: build described with one specific detail (collarbone visible, toned arms, soft stomach)
    - Clothing/jewelry: material + color + fit (sheer black lace, thin gold chain, oversized white t-shirt pushed up)
    - Male body: build, skin tone, body hair pattern, any tattoos or distinguishing marks, hand details (veins, rings, grip)
    - Setting: specific surfaces and materials (weathered hardwood, crumpled ivory duvet, tiled bathroom wall), not just "bedroom" or "bathroom"
    - Lighting: direction + color temperature + what it does to the skin (catches the sheen on her collarbone, pools warm orange across the sheets)
    - Camera: angle described by what it reveals (looking down past his chest at her upturned face, tight on her profile with his hand in soft focus behind)
    
    ADD MOTION — choose one and make it specific:
    - "She slides her lips forward along the shaft, her eyes closing as she takes it deeper, a strand of hair falling across her cheek"
    - "He grips a fistful of her hair near the crown and thrusts the shaft between her lips, her hands braced flat against his thighs"
    - "She pulls back slowly, a thin strand of saliva connecting her lower lip to the glistening glans, before leaning forward again"
    
    VOCABULARY:
    - "glans" not "tip," "shaft" for the length, "base" where it meets his body
    - "slides/glides" = she moves, "thrusts" = he moves
    - Never describe what's inside the mouth
    - Never end with mood summaries or poetry
    
    OUTPUT: Single flowing paragraph, 180-250 words. Start with "LTXdeepthroat," — end with a visual detail, not a feeling.

    Description

    This is the first release. I'm releasing early to get feedback from the community.
    PLEASE read the settings and check the example workflows in the training data.

    A few things I know need work.

    • File size — 1GB at Rank 64 is large. Will try to reduce to Rank 32 without oversaturating the output blocks.

    • Audio — glottal 'guhk' sounds need more training data and refinement.

    • Video length — currently trained on ≤121 frames. Longer generations may degrade.

    • Side-angle anatomy — male anatomy placement and detail breaks down when positioned at the side of the frame in T2V.

    FAQ

    Comments (36)

    DonkeyworldMar 19, 2026· 12 reactions
    CivitAI

    This is best one yet!

    Can't wait for furture updates!

    AI_2_addictedMar 19, 2026· 3 reactions
    CivitAI

    GREAT WORK!!

    happylittleteapotMar 20, 2026
    CivitAI

    "A mesmerizing rhythm of passion and desire"

    I was hoping they'd be wearing a sequin dress while dancing the salsa. seriously though that is important. I feel like LTX seems trickier at first because we've gotten accustomed to models with more robust weights. Like pony or whatever, that needs all kinds of superfluous things to provide emotion and creativity.

    daring_l
    Author
    Mar 20, 2026· 2 reactions

    That's my damn prompt enhancer. I keep telling Qwen to knock it off but I think abliterated models are retrained on poetry and smut novels.

    crombobularMar 20, 2026

    sometimes i wonder if we went too far with natural language. i am so tired of prompting prose or needing an entire llm for simple scenes lmao

    K3NKMar 20, 2026· 3 reactions
    CivitAI

    ive trying to train a deepthroat, sadly ltxv2 is the worst piece of shit when it comes to blowjobs.. just saying..

    crombobularMar 20, 2026

    have you tried this lora? it works decently

    K3NKMar 20, 2026

    @crombobular yeah just did, with the prompts i use in wan , little movement too.. i dont wanna input a 600 characters prompt for i2v..

    BeFair9142Mar 20, 2026

    @K3NK would you take videos for a dataset? I have a ton of full-body FF videos that I made with previous loras in combination that I created previously. life has me busy lately so I can't find the time to caption like I would like to get the best out of it. I know it would be a beast of a lora if I could either find time or someone willing to caption.

    BeFair9142Mar 20, 2026

    BTW, I would post here but a I get hit with flags sometimes so I stopped trying but I can share examples.

    crombobularMar 20, 2026

    @K3NK i understand not wanting to prompt like that but.. that's just how ltx works. git gud.

    daring_l
    Author
    Mar 20, 2026

    @K3NK Wan dataset captions are a fucking dream compared to the verbosity of LTX. Part of the problem is LTX seems to actively Barbie and Ken dolls people where iirc WAN base was fairly uncensored.

    For i2v only i usually have passable motion with 10-20 video sources around 2k steps and completed around 4k with 1e-4 lr, audio <=5e-5 . My dumbass is stubborn and wants to keeping pushing t2v,i2v + audio all together. I fully expect my shitty loras will be obsolete once you or the community get the LTX pipelines working.

    LetTheBassDropMar 20, 2026

    Yeah, LTX2.3 is great to add audio to WAN2.2 videos but for everything else I've been disappointed so far.

    MisticRain69Mar 23, 2026

    @daring_l be nice if there was a more efficient way to train ltx2.3 loras. Always have to rent a rtx pro 6000 to train at non dogshit settings and thats a minimum of $15 per experiment so if you had shit in your dataset and the lora came out mid then rip $15.

    Carnal_CreationsMar 23, 2026

    bro look at my redgifs profile and tell me ltx is not good for blowjobs

    https://www.redgifs.com/users/jrugge

    kudon44Mar 30, 2026

    @K3NK Not going to lie, nor intending to be mean, but I find this kind of hilarious considering I'm pretty sure you have the by far the largest workflows to exist on this entire website.

    crombobularMar 20, 2026
    CivitAI

    this works really well, nice one

    gambikules858Mar 20, 2026· 2 reactions
    CivitAI

    For I2V this lora strength 1.0 and 0.4-0.6 for creampie animation i2v lora. Creampie lora help a lot with mouth penetration, lips are sealed around the penis all the time, penis never disappear.

    wsx805758618Mar 20, 2026
    CivitAI
    Perfect, now it's missing an ejaculation lora.

    fiyotaMar 23, 2026· 1 reaction
    CivitAI

    How do you use the system prompt? Do you use a seperate LLM to generate the prompt that you generate with or are you using it within the actual prompt in the workflow?

    daring_l
    Author
    Mar 23, 2026

    I use a local version of Qwen3.5 which has VL built in. You can use it with Ollama or something, but my workflow queries it via vllm directly.

    outcastp23Apr 5, 2026

    @daring_l can you share a little more details about what your setup is for qwen 3.5 please for the i2v workflow, can I run qwen 3.5 locally using llama-server and get openAICompat node to connect to llama-server? what would the URL be? just http://ip:port/ ? Thanks in advance!

    daring_l
    Author
    Apr 5, 2026· 1 reaction

    @outcastp23 Yeah I have an older gaming machine where i run qwen3.5 in llama.cpp. When you setup the llama.cpp you can set the host and port. You would use your IP and port, like http://ip:port/ in OpenAICompat. You can also swap that node to an Ollama node if you want

    egus903841957Mar 24, 2026
    CivitAI

    I don't understand anything, please explain. I use Laura, the output is complete nonsense. Some kind of monsters are coming out. Using the model LTX-2.3-distilled-Q4_K_M.gguf

    araskijoMar 26, 2026· 2 reactions
    CivitAI

    Great job, working well in I2V.

    BeFair9142Mar 31, 2026
    CivitAI

    I just trained a lora and it's decent. I have to use in conjunction with other loras from better results but captions in data set weren't the best so its to be expected with what @K3NK pointed out.

    tz11171Apr 1, 2026· 1 reaction
    CivitAI

    i saw always bad results, p** like alien in image2video. I used draw things ltx 2.3, this lora 2 times , with 100 and 85% and destilles lora ltx 2.3 with 60% and the example prompts, but the output is realy bad. what make i wrong ?

    TSAHYApr 9, 2026
    CivitAI

    Thank you for sharing the system prompt. You are a life saver

    redhoodsApr 9, 2026
    CivitAI

    Which LLM are you using exactly? I used the HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive Q4 version but didn't get very good results.

    daring_l
    Author
    Apr 9, 2026

    I typical use abliterated models from https://huggingface.co/huihui-ai. Qwen3.5 in my fav right now.

    redhoodsApr 9, 2026

    So, what approach do you take when creating the system prompt?@daring_l 

    faksfkwaf22Apr 10, 2026· 2 reactions
    CivitAI

    Tried it with DR34ML4Y, it does wonders. Thanks very much!

    TheOrangeSplatApr 11, 2026

    What strengths are you using for them both together?

    faksfkwaf22Apr 13, 2026· 1 reaction

    @TheOrangeSplat default to 1. I'm using it on wan2gp

    Ostap222Apr 13, 2026
    CivitAI

    what tool do you use for training?

    rastarootbeer9940Apr 19, 2026
    CivitAI

    cant download the lora