MiniMax H3 Penis LoRA
MiniMax H3 may be the best open-source video model available right now—so, naturally, I had to be the first to bring some D to it.
This LoRA is designed for:
Text-to-video penis generation
Revealing or extending a penis in image-to-video workflows
Adding explicit anatomy to reference-video workflows
No character LoRA required
The main showcase video was created using only:
Two headshots
Three full-body shots featuring the outfit
One penis reference image
That was enough to establish the character and create the video—no character LoRA was required.
I created the original character reference sheet with Krea 2, using a LoRA trained on my likeness. However, you don’t need an existing likeness LoRA: Flux 2 Klein 9B can create strong character reference sheets from images you provide, and GPT Image 2 is another good option.
Recommended settings
Euler has produced better results than the res_multistep sampler used in many default workflows. I’m currently testing LoRA strengths between 0.5 and 1.0. The Turbo lora at 6 steps seems to be good also.
If you discover settings that work especially well, let me know—and please share your results using Add Post below!
Check out my other H3 LoRAs.
Development log
August 5, 2026 — Version 1 / 8,350 steps
This is where the magic really started happening. Adding my image dataset to the existing video dataset made a major difference. I continued training to 8,350 steps, and the results look excellent so far. This checkpoint will become Version 1.
August 5, 2026 — 6,000-step update
Released an updated checkpoint trained for a total of 6,000 steps—more than six hours of GPU training, on top of all the troubleshooting required to keep the early runs from blowing up before 200 steps.
August 4, 2026 — 3,000-step preview
It finally trains! This first preview uses the same dataset as my latest penis video LoRAs and was trained for 3,000 steps. There are still occasional anatomy issues—backward glans included—but it’s already a lot of fun to experiment with.
Description
Version 1, 8350 steps
FAQ
Comments (15)
Thanks for this. Excited to give this a shot. Appreciate the alpha. (Edit: Hater's gonna hate I guess.)
First impressions: Vast improvement on the sausages. It seems to be "penetration resistent". My tests with it had ladies going to ride on top, and then having it just rest in front of them - where as the sausages have a kind of believable effect.
Can get skin color confused with black skin, especially with edits.
Thanks. I'm still training it but the second release at 6000 steps is what all my showcase videos used.
@coachbate Well I appreciate the work you've been doing. It was underconsidered for a long time, and you seem suitably motivated. Hopefully the technique to minimize warping H3's general behavior while filling in it's "missing data" becomes more clear in time. But at this point it's still a welcome attempt. Looking forward to trying the next version.
this one has no penetration in the training data, but it should be able to do doggy with prompting. I have more loras in the queue coming for action
@coachbate I certainly don't know enough to know, but it seems like it needs very little help. I can do decent doggy now without anything, it's just that penises/vaginas/anuses look weird. It struggles with more assertive motion, but that seems easy enough to address.
So far almost every T2V generation hit the nail on the head. Great Lora already!
Can't wait to make all the weird scenarios happen lol
so glad to hear stuff like this. please do share with the Add Post so others can see what is possible
As your fan, I am very happy that you were able to complete this Lora so quickly. Thank you so much. So I don't hesitate to spend 2000 buzz, not to use this Lora for the first time, but to encourage you to create more perfect Lora products in the future. I love you very much
very sweet, thank you. and just think of it as $2, knowing I only get $1 of that. This is like a second full-time job to learn and deliver these so i appreciate it. You often just hear all the haters
I tried a few things. This LORA works pretty well for shots that are just close-ups of the penis. As the penis gets smaller and smaller, the distortion in the video gets worse and worse.
i did not train it on any far away penises. Point me to a video that has this the way you like and I can incorporate it into the dataset
This works great for adding the missing detail from the model. Do you know if there's a specific way for me to prompt for normal sized penises that are either soft or small?
who wants that? kidding. I definitely have small and soft in my dataset, so did prompting not help? Using a reference image is the best way. You can use my Krea 2, Qwen 2512, or Z Image Base to create your best images to use for reference images for H3.
I've added a post to the gallery with a bracketed set of tests of the same prompt and seed with this LoRA at various strengths: 1.00, 0.75, 0.50, 0.30, 0.00 (control). Turbo "Larry" 600 step LoRA, 8 steps, full BF16 models.
So who has the best dick out there?