TL;DR: This LoRA helps with nipples and areolas, but also with skin texture. It's the final LoRA model and might already be overtrained, so use it at lower weights. Just give it a try! (Also in combination with other LoRAs — it's a little helper, not a master.)
Hello, dear party people,
here's another essential I've retrained for Minimax H3: the good old Nipple Fix^^ that we so badly needed for the censored FLUX model, and which isn't quite as critical anymore … current models are reasonably well acquainted with these "racy details" on their own. ;)
I tested it solo — and the main benefit right now is that it overhauls the skin texture along the way, much like my main LoRA does: Perfect Skin/Eyes/Hands. (But it's better used in combination with other boob models.)
Ultimately it slots into a whole squad of my own and third-party LoRAs that I used for these sample videos, to pull off a complete studio setup in a stylish brick-walled loft at golden hour. And the model herself is fully in frame — hair, eyes at a distance, hands and feet. That's not so easy to do … but yes. It works. Compared to WAN 2.2, Minimax H3 is far more receptive to LoRAs and, as far as I can tell, does a very good job of actually acting on "the new input".
You'll find the new LoRAs I used either online or here on Civitai — or over on my Patreon page, where you get the workflows plus additional training steps, in case this model strikes you as too "overtrained".
More generally, I have my own ComfyUI node pack that's equipped for Minimax H3 as well, and it includes a custom sampler with live video preview:
https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack
With that, you can run the entire package on a 5060 Ti with 16GB and 32GB of RAM; it does take over 10 minutes without a turbo LoRA, but the quality is good, especially thanks to my new upscaler node running a solid old ESRGAN model like 4x_UniversalUpscalerV2-Neutral.
Heads-up: if you run the nipple LoRA at full strength (0.8–1.0), you may end up with a close-up even though you wanted a full scene. It can also interfere with motion or animation … though that applies to every LoRA trained on stills alone … movement simply isn't baked in. :)
https://www.patreon.com/c/polyhedron_ai
Here's the overview:
Minimax_H3_Perfect_Eyes_and_Skin_and_Hands_perfe8ct (Polyhedron, Civitai)
Minimax_H3_Breasts_nipples_nippl8es (Polyhedron, Civitai)
Minimax_H3_Glamgirls_glamgir8ls (Polyhedron, Patreon)
Minimax_H3_Oxygen_oxyg8en (Polyhedron, Patreon)
Minimax_H3_perfect_Feet_fee8t (Polyhedron, Patreon)
Minimax_H3_Golden_Hour_gold8en (Polyhedron, Patreon)
Minimax_H3_Studio_studi8o (Polyhedron, Patreon)
Minimax_H3_Brick_Wall_bric8k (Polyhedron, Patreon)
Minimax_H3_Realism People (fal, Hugging Face) [https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA]
Minimax_H3_Tiddies and Realism (Plaguekind, Civitai) [https://civarchive.com/models/2858760/minimax-h3-tiddies-and-realism-slider?modelVersionId=3229050]
Minimax_H3_Photorealistic Image Generator (JonXL, Civitai) [https://civarchive.com/models/2884206/minimax-h3-photorealistic-image-generator-lora-workflow?modelVersionId=3260165]
And yes: I'll be making all the LoRAs available on Civitai bit by bit … It would still be amazing if you could actively support me, because all of this is running up quite a bill right now.
Did I forget anything? Quite possibly! :D I'm handing it over regardless … fare you well.
Experiment, experiment. Experiment!
Take care,
Frank
Description
V1.0
FAQ
Comments (7)
I'm not clear on which LoRa types should be expected to work with which H3 models. I see you specified T2V, any idea if it should work with i2v/r2v?
nips look really good btw
Trained it on T2V — that one starts from pure noise, so that's where it definitely works. Haven't tried i2v/r2v. Different checkpoint, and the more the thing is conditioned on an input image, the less room an appearance LoRA has to do anything — with i2v the start frame already decides what the skin looks like. R2v would be the better bet. Lower weight, see what happens.
Cheers! :)
T2V == FLV (First-Last)
I2V may or may not imply REF. Sometimes people say I2V when they mean REF, sometimes they say I2V when they mean FLV.
Essentially:
- T2V should mean it was trained on the FLV model, and should produce what it was trained on without a starting image.
- I2V by default implies it was trained on the FLV model. but it needs a starting image to access what it was trained on
- REF should mean it was trained on the REF model, and what that entails, meaning it shouldn't work on the FLV model.
You are right that people don't always say what they did. This is just the guide.
Great work. But please, would it be possible to create a version focused on overall realism—scenes of everyday life, realistic clothing, and natural, unfiltered skin—like an amateur video shot on an ordinary smartphone, without focusing on breasts or nipples? That would be amazing. Thanks!
Loras are a product of the data they are trained on. Sometimes if you don't have data from consistent sources, it can be more impactful to sort and train them on a different axis. Combining loras is pretty possible, and if they're from the same creator they should overlap nicely (if they prompted similarly).
There are a few "cellphone" loras out there at the moment, and each suffer fairly heavily from hurting adherence at too high of weights. I do share your want of convenience, but H3 is already so good at this style it almost seems detrimental to the overall output without very careful prompting and data.
The several examples you put up look like you trained on only one clip in your data? Or did you just use the same seed and prompt, and change a small part of it?
Same setup and same prompt with the seeds I always use for Civitai: 2094697661, 2094697662, 2094697663, 2094697664, 2094697665, 2094697666. :)