A mirror of the MiniMax-H3 LoRAs published on
CivitAI. The files here are byte-identical to the
CivitAI downloads (SHA-256 listed below). CivitAI carries the sample videos, the version
history and the comment threads; this repository exists so the weights can be pulled with a
plain resolve URL, without an account.
One anatomy adapter teaches H3 what a body part looks like. One action adapter teaches
it what a body does. They are single-concept and they stack, so put them under whatever
character or scene LoRA you are already running.
Every LoRA name in the table links to its CivitAI page, deep-linked to the exact version this
file came from.
Where a strength is blank, no single number was published — start at 1.0 and pull it back.
How to prompt each one
HMNSFW V2.5 — the current motion adapter
Trigger hmmotion. Strength 0.5 – 0.9. Trained on close to 1000 videos, roughly 15x the V2
dataset, over 40 hours of training. I2V is very good here, with coherent anatomy and motion, and
it handles most T2V prompts with the occasional bad seed.
Your prompt is everything. A short or vague prompt gives you a bad result. Use the example prompts
on the CivitAI page as anchors and match their length and register.
This was going to be V3. It isn't finished, and more versions are coming.
HMNSFW V2 — the previous motion adapter
Trigger hmmotion. Strength 0.5 or below; above that it collapses coherence. Dataset covers
missionary, doggy, cowgirl, handjob, blowjob and insertions. I2V works well across most positions,
with the occasional deformed genitalia; T2V is hit or miss, which is what the anatomy adapters
below are for.
Long descriptive prompts beat short ones by a wide margin. The training captions run 165-269
words with a median of 225, so a keyword list is off-distribution for this checkpoint. Write
one flowing paragraph of plain anatomical prose, not tags. The full prompt-writing system
prompt is on the CivitAI page.
Generated with the full bf16 model, trained on bf16. dpmpp_2m + Beta scheduler, 20 steps.
HMCumshot — the action adapter
Trigger cumshot. Strength 0.9. Captioned in plain, blunt language, so prompt the same way.
A thorough description of the cum texture helps a lot. Stacks on top of HMPussy and HMPenis.
Optionally add the ExtendIntermediateSteps node with 2 extra steps starting at sigma 0.85.
V0.5 expanded the dataset to 56 videos with more angles.
HMSquirt — squirting
Trigger hmsquirt. Experimental, trained on a small dataset, and it does work. It is not the
best thing here and a better version is in progress. Run HMPussy alongside it at 0.4.
Trigger pussy. Strength 1.0. Rebuilt from the ground up on 4x the dataset, covering innies,
bushes, shaved, anuses and most everything else between a woman's legs. One file, smaller than the
old pair, and no video partner needed unless you want the motion adapter under it.
HMPussy v0.5 — two files, both required
vagassist_e40 is stills-trained: run it at 1.0. It restores legible structure in female
genital anatomy, which the base model renders soft and vague. hmpussy_v6_epoch30 is
video-trained: run it at 0.35, sitting under the stills file rather than carrying a
generation alone. Skip the video file if you do not care about motion.
Trained on fingering, spreading and plain show-off. LoKr adapters on the MiniMax-H3 fl2va
bf16 base. V0.5 improved anuses considerably; they are still not all the way there.
HMInnie — shape control
Trigger inniepussy, and it replaces the word "pussy" in your prompt rather than sitting
in front of it. Every training caption opens with a camera clause, so include one:
inniepussy shown from the front / from behind / from below.
Four axes, written as plain English inside the sentence, not as tags:
Mound — flat · soft · puffy · very puffy
Cleft — closing to a smooth seam · with a defined cleft · parted around a visible opening
Sits alongside HMPussy for general genital detail, and under HMNSFW when the anatomy needs to
survive motion. Trained with AI-Toolkit on 117 images at rank 32 on the full bf16 base.
HMBreasts V2 — chest control, current version
Versatile across most sizes and shapes, and the prompts can stay short — write the chest in plain
prose the way the CivitAI examples do, no leading trigger token. Use the seeds2 sampler with the
ddim_uniform scheduler and the 8-step Turbo LoRA.
The four axes from V1 still apply and still read as plain English inside the sentence:
tiny tits with brown areoles
tiny sized tits and pale areoles
HMBreasts V1.0 — chest control, previous version
Trigger HMBreasts, leading. Strength 1.0. Base H3 renders a chest as an approximate shape and
ignores anything you say about it; this makes size, areole size and areole colour things you type.
Four axes, again as plain English:
Breast size — tiny · small · medium · large · very large
Areole size — tiny · small · medium · large · very large
Areole colour — pale · ghost · brown · dark
Nipples — erect · hard · pierced
Spelling matters: it learnt areoles, so write it that way. Example clauses:
HMBreasts, large, natural breasts, medium sized brown areoles and erect nipples
HMBreasts, small, perky breasts, tiny sized pale areoles and hard nipples
HMBreasts, very large, round breasts, large sized ghost areoles and erect nipples
The axes were trained independently, so tiny breasts with large dark areoles is a combination
it can actually give you.
Honest limitation:medium is the weakest of the five sizes. The training material
clustered at both ends, so tiny/small and large/very large steer hard while the middle is
softer than either. A shape word usually pulls it back.
It owns the chest and nothing below it — genitals are HMPussy, and the two stack. It does not
touch faces, bodies or identity. Trained on stills, so it teaches H3 what a chest looks like;
how it behaves through motion still comes from the base model.
HMPenis — anatomy
Trigger HMPenis, leading. Works best with medium-to-large sizes. Specify the camera
direction in the prompt: front (POV-like), back, or side. Useful adjectives: large,
circumcised, glans with a colour adjective (pink / pale / brown). The dataset is thin.
Stacking
Anatomy adapters (HMPussy, HMInnie, HMBreasts, HMPenis) are stills-trained and single-concept.
HMPussy V1 is stills only; hmpussy_v6_epoch30 from v0.5 is the video partner if you want it.
Action adapters (HMNSFW, HMCumshot, HMSquirt) are video-trained. They compose: put the anatomy
adapters underneath, the action adapter on top, and your character or scene LoRA above both.
Pick one HMNSFW, one HMBreasts and one HMPussy stills file. The two versions of each are the same line, not stackable
partners.
Verification
Every file is byte-identical to its CivitAI download. Each hash below resolves through the
CivitAI model-versions/by-hash endpoint to the version linked in the table above.
These load with a stock LoraLoaderModelOnly in ComfyUI. For wiring references into
Reference-to-Video without rebuilding the graph every time, see
ComfyUI-MiniMaxRefPack.
Licence
The adapter weights are released as-is. Use of the MiniMax-H3 base model is governed by the
MiniMax-H3 Community License. You are
responsible for what you generate and for complying with the law where you are.
