CivArchive
    Ultimate DeepThroat - MiniMax H3 LoRa - K3NK - MiniMax H3 - v1.0
    NSFW

    • ᴛʜɪꜱ ʟᴏʀᴀ ɪꜱ ᴘʀᴏᴠɪᴅᴇᴅ ᴀꜱ ɪꜱ ᴀɴᴅ ɪɴᴛᴇɴᴅᴇᴅ ꜰᴏʀ ʟᴀᴡꜰᴜʟ ᴀɴᴅ ʀᴇꜱᴘᴏɴꜱɪʙʟᴇ ᴜꜱᴇ.

    • ɪᴛ ᴀᴄᴛꜱ ᴀꜱ ᴀ ɢᴇɴᴇʀᴀʟ-ᴘᴜʀᴘᴏꜱᴇ ʀᴇꜰɪɴᴇᴍᴇɴᴛ ʟᴀʏᴇʀ, ɪɴꜰʟᴜᴇɴᴄɪɴɢ ʜᴏᴡ ᴛʜᴇ ʙᴀꜱᴇ ᴍᴏᴅᴇʟ ᴇxᴘʀᴇꜱꜱᴇꜱ ᴘʀᴏᴍᴘᴛꜱ ᴀɴᴅ ᴄᴏɴᴄᴇᴘᴛꜱ ʀᴀᴛʜᴇʀ ᴛʜᴀɴ ᴘᴇʀꜰᴏʀᴍɪɴɢ ᴅɪʀᴇᴄᴛ ᴛʀᴀɴꜱꜰᴏʀᴍᴀᴛɪᴏɴꜱ. ᴛʜᴇ ʀᴇꜱᴜʟᴛɪɴɢ ᴏᴜᴛᴘᴜᴛ ɪꜱ ᴀʟᴡᴀʏꜱ ᴅʀɪᴠᴇɴ ʙʏ ᴛʜᴇ ᴘʀᴏᴍᴘᴛ ᴀɴᴅ ᴛʜᴇ ᴄᴀᴘᴀʙɪʟɪᴛɪᴇꜱ ᴀʟʀᴇᴀᴅʏ ᴘʀᴇꜱᴇɴᴛ ɪɴ ᴛʜᴇ ᴜɴᴅᴇʀʟʏɪɴɢ ʙᴀꜱᴇ ᴍᴏᴅᴇʟ.

    • ᴛʜᴇ ʟᴏʀᴀ ᴅᴏᴇꜱ ɴᴏᴛ ʀᴇʟʏ ᴏɴ ᴘᴀɪʀᴇᴅ ɪɴᴘᴜᴛꜱ, ʙᴇꜰᴏʀᴇ/ᴀꜰᴛᴇʀ ᴇxᴀᴍᴘʟᴇꜱ, ᴏʀ ᴛʀᴀɴꜱꜰᴏʀᴍᴀᴛɪᴏɴ-ꜱᴘᴇᴄɪꜰɪᴄ ᴛʀᴀɪɴɪɴɢ. ɪᴛ ꜰᴏᴄᴜꜱᴇꜱ ᴏɴ ɪᴍᴘʀᴏᴠɪɴɢ ᴄᴏʜᴇʀᴇɴᴄᴇ, ᴅᴇᴛᴀɪʟ, ᴀɴᴅ ꜱᴛᴀʙɪʟɪᴛʏ ᴀᴄʀᴏꜱꜱ ɢᴇɴᴇʀᴀᴛɪᴏɴꜱ, ɪɴᴄʟᴜᴅɪɴɢ ᴍᴀᴛᴜʀᴇ-ᴛʜᴇᴍᴇᴅ ᴏᴜᴛᴘᴜᴛꜱ ᴀʟʀᴇᴀᴅʏ ꜱᴜᴘᴘᴏʀᴛᴇᴅ ʙʏ ᴛʜᴇ ʙᴀꜱᴇ ᴍᴏᴅᴇʟ.

    • ɴᴏ ꜱᴘᴇᴄɪᴀʟɪᴢᴇᴅ ᴅᴀᴛᴀꜱᴇᴛꜱ ᴛᴀʀɢᴇᴛɪɴɢ ɪᴅᴇɴᴛɪᴛʏ ᴇxᴘᴏꜱᴜʀᴇ ᴏʀ ᴄᴏɴᴛᴇɴᴛ ʀᴇᴄᴏɴꜱᴛʀᴜᴄᴛɪᴏɴ ᴡᴇʀᴇ ᴜꜱᴇᴅ ᴅᴜʀɪɴɢ ᴛʀᴀɪɴɪɴɢ. ᴀʟʟ ꜰᴀᴄᴇꜱ ᴀɴᴅ ᴘᴏᴛᴇɴᴛɪᴀʟʟʏ ɪᴅᴇɴᴛɪꜰɪᴀʙʟᴇ ꜰᴇᴀᴛᴜʀᴇꜱ ᴡᴇʀᴇ ᴍᴀꜱᴋᴇᴅ ᴏʀ ᴀɴᴏɴʏᴍɪᴢᴇᴅ ᴛʜʀᴏᴜɢʜᴏᴜᴛ ᴛʜᴇ ᴛʀᴀɪɴɪɴɢ ᴘʀᴏᴄᴇꜱꜱ.

    • ʙʏ ᴅᴏᴡɴʟᴏᴀᴅɪɴɢ ᴏʀ ᴜꜱɪɴɢ ᴛʜɪꜱ ʟᴏʀᴀ, ʏᴏᴜ ᴀᴄᴄᴇᴘᴛ ꜰᴜʟʟ ʀᴇꜱᴘᴏɴꜱɪʙɪʟɪᴛʏ ꜰᴏʀ ɪᴛꜱ ᴜꜱᴇ ᴀɴᴅ ᴀɢʀᴇᴇ ᴛᴏ ᴄᴏᴍᴘʟʏ ᴡɪᴛʜ ᴀʟʟ ᴀᴘᴘʟɪᴄᴀʙʟᴇ ʟᴀᴡꜱ ᴀɴᴅ ᴘʟᴀᴛꜰᴏʀᴍ ʀᴜʟᴇꜱ.


    Ultimate DeepThroat - MiniMax H3 LoRa - K3NK

    • Trained on Musubi tuner AkaneTendo25 fork with 964 clips (x468 Gag dataset, x378 69Pov dataset, x118 PovBJ dataset)

    • Trained at 512,512

    • Rank 32, Automagic3 Optmizer, h3_lora_targets "attention:24-49;mlp:24-49"

    • Trained video and audio

    • Trained with Pruned model (will affect fluid dynamics)

    • All the example videos are made with 4-6 steps, quacking audio in some of the example videos is from the gag5-step00007400-QUACKING.safetensors, using it along side this nodes)

    The workflow im using is embeded inside the videos, drag 'n drop into comfyui

    Sideview (Ref)

    How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video.
    
    subject_definitions:
    <Subject 1> is the young woman matching Picture 1 — young woman, kneeling, mouth closed.
    <Subject 2> is the large erect fair-skinned penis — visible veins and round testicles clearly visible.
    
    integrated_multimodal_description:
    At 00:00.000, the video starts matching Picture 1 —  She says <d>[English] "Let me take care of that cock of yours!"</d> with a confident expression. At 00:02.000, she finishes speaking and smiles slightly, opening her mouth wider in anticipation. At 00:03.000, a large erect fair-skinned penis matching <Subject 2> enters from the right side of the frame. The penis slides in from the side, visible with veins and testicles clearly visible. At 00:04.000, <Subject 1> opens her mouth wide and takes the penis into her mouth. She has her lips stretching around the shaft, taking him deep into her throat. She moves her head back and forth in a deepthroating motion, swallowing the penis from tip to base. The man moves his pelvis back and forth vigorously and energetically, thrusting his hips as she deepthroats him. Her lips stretch around the shaft as she takes him deep. The man's pubic area and lower body are visible in the frame. She continues the back and forth head movement throughout, her lips smashing against his pelvis with each thrust. The video ends with her still in the deepthroating position.
    
    overall_soundscape:
    her dialogue "Let me take care of that cock of yours!" at 0.0s, wet oral sucking sounds starting at 4.0s, heavy breathing from him, wet tongue movements, lip smacking sounds, head movement sounds, hip thrusting sounds, saliva dripping sounds, gagging sounds, wet throat sounds, side view intimate audio
    
    non_diegetic_music:
    N/A
    

    POV (Ref)

    How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video.
    
    subject_definitions:
    <Subject 1> is the young woman matching Picture 1 — young woman, kneeling, looking up, mouth closed.
    <Subject 2> is the large erect fair-skinned penis — visible veins and round testicles clearly visible.
    
    integrated_multimodal_description:
    POV deepthroat. At 00:00.000, the video starts matching Picture 1 — <Subject 1> is kneeling, looking up at the viewer, mouth slightly open. She says <d>[British english] "¿What would you like to do with me today?"</d> with a playful expression. At 00:02.000, she finishes speaking and smiles, opening her mouth wider, exhaling softly. At 00:03.000, a large erect fair-skinned penis matching <Subject 2> slides up from the BOTTOM of the frame. The penis appears from below, visible with veins and testicles clearly visible. At 00:04.000, <Subject 1> opens her mouth wide and takes the penis into her mouth. She has her lips stretching around the shaft, taking him deep into her throat. She moves her head up and down in a deepthroating motion, swallowing the penis from tip to base. The man moves his pelvis up and down vigorously and energetically, thrusting his hips as she deepthroats him. Her lips stretch around the shaft as she takes him deep. The man's pubic area and lower body are visible in the frame. She continues the up and down head movement throughout, her lips smashing against his pelvis with each thrust.
    
    overall_soundscape:
    her dialogue <d>[British english] "¿What would you like to do with me today?"</d> at 0.0s, wet oral sucking sounds starting at 4.0s, heavy breathing from him, wet tongue movements, lip smacking sounds, head movement sounds, hip thrusting sounds, saliva dripping sounds, gagging sounds, wet throat sounds, POV angle intimate audio
    
    non_diegetic_music:
    N/A

    69 (Ref)

    How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video.
    
    subject_definitions:
    <Subject 1> is the young woman matching Picture 1 — young woman, kneeling on top, looking forward, mouth closed.
    <Subject 2> is the large erect fair-skinned penis — visible veins and round testicles clearly visible.
    
    integrated_multimodal_description:
    69 deepthroat. At 00:00.000, the video starts matching Picture 1 — <Subject 1> is positioned on top, kneeling, looking forward, mouth slightly open. She says <d>[English] "Im gonna make you happy, daddy..."</d> with a confident smile. At 00:02.000, she finishes speaking and smiles wider, opening her mouth in anticipation. At 00:03.000, a large erect fair-skinned penis matching <Subject 2> slides up from the BOTTOM of the frame. The penis appears from below with testicles clearly visible, veins visible. The man is positioned below her, lying on his back, his face below hers for 69 position. At 00:04.000, <Subject 1> opens her mouth wide and takes the penis into her mouth while simultaneously the man takes her face into his mouth below. She is on top of the man, both heads aligned for mutual 69 deepthroat. She moves her head up and down in a deepthroating motion, swallowing the penis from tip to base. The man moves his pelvis up and down vigorously and energetically, thrusting his hips as she deepthroats him. Her lips stretch around the shaft as she takes him deep. Both heads remain aligned in perfect 69 position throughout. The man's pubic area and lower body are visible in the frame, testicles clearly visible. She continues the up and down head movement throughout, her lips smashing against his pelvis with each thrust. The video ends with her still in the deepthroating position, both heads still aligned in 69 position.
    
    overall_soundscape:
    her dialogue "Im gonna make you happy, daddy..." at 0.0s, wet oral sucking sounds starting at 4.0s, heavy breathing from him, wet tongue movements, lip smacking sounds, head movement sounds, hip thrusting sounds, saliva dripping sounds, gagging sounds, wet throat sounds, mutual deepthroat sounds from both mouths, 69 position sounds
    
    non_diegetic_music:
    N/A

    Have fun! 😜

    Description

    FAQ

    Comments (16)

    WaifuAIDegenPrompterSep 2, 2026· 1 reaction
    CivitAI

    I am getting some inconsistent results.

    When I run without a turbo LORA I2V, weight has to be 0.35-0.4 to get something that isn't artifacting or blurry. I have to generate a couple videos to get a goode one.

    Running turbo LORA 8 steps I2V, results are also varying when weight is at 0.7 to 1.0

    I've tried importing the workflow from the videos above but it crashes my ComfyUI

    K3NK
    Author
    Sep 2, 2026· 1 reaction

    there must be something wrong with your wf.. i dont have any issues and ive tested plenty models and the lora works in fl2va and ref2va, what str are you using turbo loras at?

    @K3NK 
    I'm using a turbo workflow and a the MiniMaxEasy Workflow.

    Turbo Lora at 1.0 to 0.9 strength. Other things run fine. Not sure what I'm doing wrong if it's me.

    DaddyWolfgangSep 15, 2026

    What are you prompts like and are the images you're starting with good quality?

    @K3NK You're my hero

    anth0nxSep 16, 2026

    perhaps your resolution, if aspect/pixels deviate far from training might cause this? just a thought

    K3NK
    Author
    Sep 16, 2026

    @anth0nx probably is the turbo loras, use the v0.1 from kijai, im using the 3step lora + v0.1 fl2va from kijai's hf

    AlbanianDoorknobSep 16, 2026

    I can confirm that at a weight of 1 using the default non-workflow method the result is a blobby mess. Some results were worse than others but they were unusable.

    Weight of 0.3 resulted in a much better (albeit somewhat squiggly line-y) generation. The penis merged a bit with the tongue/lips at this weight.

    Changing the resolution had no noticable effect on quality of generations (864x512, 512x512, etc...) Number of steps showed no major difference (tested at 15, 20, 30)

    I wonder if this was trained on a specific model other than some other models? I'm only guessing since I do not have the hardware to train myself.

    K3NK
    Author
    Sep 16, 2026

    @AlbanianDoorknob trained with pruned model, but shudnt be an issue like that, i only use pruned for generating, try that one.. weight 0.3 is probably doing nothing..

    AlbanianDoorknobSep 16, 2026

    @K3NK Yeah, I'm using a pruned one: minimax_h3_ref2va_pruned-Q4_K.gguf.

    I'm wondering that since this is a fork of Musubi if it is that conversion issue that has plagued that repo.

    So regardless it looks like I need to take extra steps to get this to work and I know it isn't just a "TRAIN BAD >:(" issue because you're one of the best trainers out there.

    @AlbanianDoorknob @K3NK I apologize, I wasn't trying to say training bad. I was more trying to figure out why my results were inconsistent. I love K3NK's training. most of the stuff is so good.

    AlbanianDoorknobSep 19, 2026

    @WaifuAIDegenPrompter No I meant like from myself, not a dig at you.

    MiraStudioSep 4, 2026· 2 reactions
    CivitAI

    Your embedded workflow is a ComfyUI nuke, other than that good lora training 👍🏻

    lt_warfSep 15, 2026
    CivitAI

    This works great for reference to vid. Great work!

    hboxgames132Sep 16, 2026· 2 reactions
    CivitAI

    HAHA ! I noticed you're spanish speaking person, you put that inverted question mark. Thanks btw

    nightcallSep 18, 2026· 1 reaction

    ¿Which question mark are you referring to? I don't see it.

    hboxgames132Sep 19, 2026

    @nightcall  can't tell, my mistake

    LORA
    MiniMax H3
    by K3NK

    Details

    Downloads
    2,255
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/31/2026
    Updated
    9/22/2026
    Deleted
    -