For more LoRAs and updates, Join my Discord:
https://discord.gg/ZVWVhT43GW
You can deploy MiniMax on Runpod with my template:
👉 https://get.runpod.io/minimax-template
👉 Deployment tutorial
Trigger is HMPenis. Lead with it.
Dataset is pretty thin because there's a limit to what I am willing to do for this gig.
It works well with medium-large sized cocks.
Specify the direction in the prompt:
Front (Like a POV angle)
Back
Side
Use words like:
Large, circumcised, glans with color adjectives (pink/pale/brown)
Description
A slight larger (no pun intended) dataset and lower file size, renders better penises.
FAQ
Comments (24)
Does it work properly with a small penis in I2V mode without turning it into a giant monster cock?
Giant monster cock?
I thought these were completely average
@HearmemanAI They're too slender.
@HearmemanAI Not average at all. Typical Pooorn Movie Digggs. Never saw someone with such a thing in real life. :-) This might be enough for normal vids, but it would be cool if we could generate believable different sizes for fantasy movies, fairytale scenes, etc.
Much more better. Thanks.
Any chance for example prompts to be added to the example videos?
tried your example videos prompts, getting similar results like yours but not anatomically detailed like your?
i did t2i with fl2va int8 and nvfp4 textencoder . Also using turbo lora 4setp_1.0_768p_comfyui_bf16 with str=0.76 and step=8 res_multistep simple.
Should i try using your workflow sampler and scheduler?
I use bf16 with ddim uniform and euler
@HearmemanAI then what about the workflow you published, can you share your current generations workflow via pastebin or something?
@HearmemanAI This right here is the biggest thing for quality. And to be honest i don't see that much difference in gen time using smaller "faster" models and the full model. So I just use the full model.
@DaddyWolfgang The model side only effects performance in so far as your graphics cards and reasonably move data to and frome it. H3 is giant, so really only commercial cards cand use it efficiently.
@boobkake22 BF16 doing fine on my 3090. Difference of a few minutes only over 10 second video generation at 768 x 1344.
it still has an issue with knowing the front and the back of a penis, can you please fix that?
I tested all penis loras and most of them have the same issue.
@CyclopsGER yeah it is really upsetting, h3 is really bad at this, and almost great at just about a lot of things.. we can do this guys fix the loras!!
@ytyuiu545 I guess its hard to train. In my clips the camera is rotating.
It's a hit or miss.
I hand labelled the dataset several times (and lost tons of confidence doing it), the details are too tiny to be properly captured without influencing the whole video.
@HearmemanAI this is gonna sound nuts.. maybe mimimax h3 is just not good for porno? we need someone to invest in a true NSFW model that doesn't need 12 loras to fix lol
Thank you for your hard work anyways!
@ytyuiu545 MiniMax is awesome for this stuff!
@CyclopsGER for strait stuff... maybe, but not so much for the gays... it made some guys look like the have vaginas in the back, even with the ass and balls lora haha
Does anyone know how to structure the H3 prompt with the inclusion of the trigger word? Where is H3 expecting it to be?
could you please make a girl asshole helper lora
There's a pussy and anus lora on my profile
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, 0.90, 0.75, 0.50, 0.30, 0.00 (control). Turbo "Larry" 600 step LoRA, 8 steps, full BF16 models.
Penis LoRA Comparison Part 1
Penis LoRA Comparison Part 2