CivArchive
    Prone Bone - Minimax H3
    NSFW

    Triggers

    • a man and woman are having sex in the prone bone position.

    • viewed from behind

    • viewed from the front

    • viewed from the side

    • a pov shot

    • viewed from a three quarter angle

    • another man is in front of her. he is spitroasting her

    Description

    FAQ

    Comments (20)

    ODBsSep 28, 2026· 1 reaction
    CivitAI

    Knew it!

    brnfd24434343dSep 28, 2026
    CivitAI

    Not usually, this is one seems a little plastic.

    teeay25225Sep 28, 2026

    Bro this does not come from this lora, H3 is to blame. But you can do something about it: Lower your speed lora strength from 1 to 0.85. Set videoshift to 6.

    ss9999Sep 29, 2026

    Turn off turbo. Use full 16- bit model, not the very inferior INT8 convrot. Increase steps to 44 minimum. Use a double-work per step sampler like SA Solver pece. Some NSFW loras improve naturalness of skin so try using one or more of them in conjunction. The v2 "unlocker" lora might help, along with one of the cinematic ones. I have only just begun testing this model. It doesn't seem to be as tolerant of lora stacking as Wan 2.2 but does seem to do some things better than Wan, even with a heap of loras.

    ss9999Sep 29, 2026

    Some prompting might help, like "low contrast" and "washed out lighting". I have barely started testing this model but so far I've already had some good success combating the deep fried look.

    ODBsOct 7, 2026

    @ss9999 or just, you know, use a good reference image.

    ss9999Oct 7, 2026

    @ODBs Does that require the ref model, though?

    ODBsOct 7, 2026· 1 reaction

    @ss9999 Not really, no? You do know what FL2VA is... right? Beside, it's been a month since everyone moved on to hybrids.

    See, you're pretty much completely clueless, yet go out of your way recommending the wildest fucking shit that you've evidently never run yourself. 44 steps minimum... like, why not 45 or 46? And as if 44 steps was not enough, you the recommend a two-step sampler that H3 has never been tuned for?

    And 'very inferior INT8 convrot' - just, please stop.

    ss9999Oct 7, 2026

    @ODBs 1) Turning off turbo is always the first thing to see if it will reduce artificiality. Turbo is a lossy process that reduces quality. There is no free lunch, although with a specific set of settings sometimes turbo is good enough for one's particular purposes.

    2) Quantization like INT8 is not a free lunch. There is always a quality reduction because it's lossy. I have been very disappointed with it in my testing so far, when compared with the full 16-bit generations.

    3) The devs themselves recommend around 50 steps with no turbo. If they had believed that INT8 convrot is all anyone needs they would have released the model in that form, plus one in FP32 for training. They did not, nor did ComfyUI devs decide to only provide INT8 convrot to the community instead of also offering a 16-bit version.

    4) I haven't done a lot of testing because I work full-time and have other responsibilities but I have had better results so far with two-step samplers like HeunPP2.

    5) You and some others may use hybrids but that does not mean everyone does or should.

    Differences of opinion don't warrant abusive posting. I was simply offering suggestions in good faith that people are free to try and free to ignore. I have been working with AI since Stable Diffusion 1.2. I am not an expert with this particular model because I am very busy with work and taking care of a mother with dementia. However, I do have a great deal of experience with image generation models as well as quite a lot with Wan 2.2.

    ODBsOct 7, 2026· 1 reaction

    @ss9999 I wasn't talking about turbo. 20-32 steps + 5 steps latent upscale is more than enough, anything else you're just wasting inference time. Sampler - you won't get better results than with res-multistep, euler as a backup. Minimax was built with linear in mind has been tuned on res_multistep.
    Recommending BF16 as if everybody is running 32+ gigs of VRAM is just... perplexing. And no, INT8 convrot does not cause plastic skin. BF16-->INT8 Convrot is not equivalent to fp16-->fp8. Int8 convrot can only be disappointing if your prompts are poorly written. If anything, you'll get more worth out of BF16 text encoder.

    'You and some others may use hybrids but that does not mean everyone does or should.'
    There's virtually no difference between fl2va and ref2va. Hence they have been merged.

    ss9999Oct 8, 2026

    @ODBs Int8 convrot has been clearly worse in Wan 2.2, when compared with Q8_0, from what little time I have had to test it. And Q8_0 is worse than 16-bit. Prompts can't put data back that no longer exists because of lossy compression. One creator here was absolutely adamant and certain that FP8 is good enough. People can have strong opinions but data shows that lossy is lossy. Some lossy schemes are, of course, worse. I am satisfied with Q8_0 on my 12 GB system for Wan 2.2. It's "good enough" because it's the best that can be had with such limited VRAM. For my 32 GB system I run the 16-bit model. I find it droll that I was being lectured before about how FP8 is so much superior to Q8_0 and now that INT8 convrot is the new fad — FP8 suddenly is no longer so great but it is.

    I have two systems, one with 12 GB of VRAM (hence the quantization testing) and one with 32 GB of VRAM. Plastic skin is one aspect of artificiality/overall quality but not the only one. You may not have been speaking about turbo but I was in my original advice which you nastily dismissed as cluelessness.

    I tested res multistep (especially as it is ComfyUI's workflow default) and euler and had better results with samplers such as Heun PP2, Seeds2, Seeds3, and SA Solver Pece. If others have better results with single-step samplers that's fine. I am only reporting my experience. I always try Heun PP2 after trying Euler for comparison, with any model, as Heun PP2 is a Euler derivative that has additional correction. So it usually does a better job than Euler, at the cost of slowness. Unfortunately, it's not always better than other samplers for each seed.

    You're claiming that the devs released two separate models that aren't substantively different from one another to justify being two separate models. I find that hard to believe. I have only barely begun testing the ref model but what little testing I've done resulted in worse realism. Reddit comments suggest that it tends to have worse sound. People made hybrid Wan models, too — and there was a lot of evangelism about some of them. I have yet to find any custom checkpoint that performs better than the original separate models.

    "20-32 steps + 5 steps latent upscale is more than enough, anything else you're just wasting inference time." Where is the proof of this claim? Plenty of people think 8 steps with a turbo model is enough, no matter which model we're talking about. Enough for what purpose and for who? There is also a rather big difference between recommending 20 steps and recommending 32. Why 32 and not 42?

    Don't look at the new model that just dropped that needs an 80 GB card to run the recommended 50 steps. I think the MiniMax devs should have released a MoE model like Wan 2.2 because companies like Nvidia and AMD were sandbagging on VRAM even before the current bitbarn feeding frenzy.

    ODBsOct 8, 2026

    @ss9999 
    'Int8 convrot has been clearly worse in Wan 2.2, when compared with Q8_0,'
    How is Wan 2.2 even relevant here? Minimax's architecture has little to do with how Wan MoE worked apart from them both being diffusion models.

    'I tested res multistep (especially as it is ComfyUI's workflow default) and euler and had better results with samplers such as Heun PP2, Seeds2, Seeds3, and SA Solver Pece. '
    Without example footage, the above doesn't really have any weight/meaning to it. I also tried seeds 2, for example - and guess what, it was crap - stiff motion, completely changed texture and lighting from the first_frame input - whereas res_multistep stayed consistent over multiple gens.

    'You're claiming that the devs released two separate models that aren't substantively different from one another to justify being two separate models. I find that hard to believe.'

    You're pretty much just admitting to me that you don't know what you're talking about here. ref2va-fl2va situation has been discussed plenty of times and is only a short search query away to verify.
    What *I*, in turn, find hard to believe are your claims of 'having years of experience' - especially given the fact that you haven't posted a single example of of your work here.

    'Where is the proof of this claim? Enough for what purpose and for who? There is also a rather big difference between recommending 20 steps and recommending 32. Why 32 and not 42'
    On my page, smartass.

    'I think the MiniMax devs should have released a MoE model like Wan 2.2'
    Yeah uhhh, no thanks.

    While I have sympathy to your mother's ailment and appreciate that you're busy with work (if what you're saying is genuine) - again, I've yet to see a single generation of yours, only longwinded discourse with little substance. The AI segment online is filled to the brim with such cases (especially on reddit and youtube) where there's a ton of talk about this and that workflow/setup and 'oh my lawd this new model is a gamechanger' yet when you look at examples being posted - there are either none to be found or they are your typical generic undinspired dross. I have the same laugh when I see people online discussing shiny new Alexa 35 or Red Komodo as if it's going to make their mediocre camera/storytelling skill any better. Jeez.

    Truly 'democratizing' technology.

    chrmaroudas625Oct 10, 2026

    @ss9999 Just wanted to wish you the best about taking care of a parent like that, been there, done that, it sucks

    NugusSep 28, 2026· 1 reaction
    CivitAI

    aixartsSep 28, 2026· 3 reactions
    CivitAI

    Great work!

    Swoosh333111Sep 29, 2026· 1 reaction
    CivitAI

    can it do side missionary view

    kermitfrog1202
    Author
    Sep 29, 2026

    Well... It's a prone bone position lora... so...

    artyclawSep 29, 2026
    CivitAI

    juicedrinkingratOct 5, 2026
    CivitAI

    Dumb question how can I get the workflow using this Lora ?

    kermitfrog1202
    Author
    Oct 5, 2026

    Save one of the examples and open it in comfyui.

    LORA
    MiniMax H3

    Details

    Downloads
    4,930
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/28/2026
    Updated
    10/23/2026
    Deleted
    -