MiniMax H3 T2V/I2V
Took me a while to figure this one out, but I think this is a solid first version.
MiniMax is amazing, but a bitch to train for. Mainly because it learns too much and one or two items in your dataset can lead to weird behaviour.
For instance, too many closeups in this one makes it want to constanly zoom in or focus on the dick (which is kind of anoying when you don't want that). Ajusting the lora stregnth can help with this; you can get away with a as little as 0.6 strength if the subject is not in the forefront.
Keep in mind that the keyword is necessary for this one, otherwise you'll just get lumpy blobs.
I'll probably come back to this for a V2, but not until next month, Training is expensive as fuck, especially for video models.
If you want to help me out you can use my Runpod referal link (HERE) to get some free credits for you and for me. Some great templates to get up an running quickly and experiment.
If you enjoyed this Lora, consider kicking a few bucks my way!
https://ko-fi.com/jouhellerxxx
Description
FAQ
Comments (24)
How many seconds of video footage would be suitable for LoRa training on a MiniMaxh3, and what should the training speed and interval be?
In my experience, less is more, around 30 videos/images will give you better results that 60+(images train as if they were 1 frame long videos, so they still help).
Videos should be between 3-5 seconds long, at least 720p.
MiniMax lears really quick, so you can probably see if you are headed in the right direction around 1000 steps, but the good results will appear beyond 3000 and 6000 steps (For concept loras like this, styles might be a bit less.)
Avoid having similar images/videos, the model will latch on to that and asume thats all you want, keep it varied.
@jouhellerxxx391 How much VRAM does a graphics card need? How long does it take to train LoRa?
@evan39102561943 I used a RTX PRO 600 WK with around 140gigs of RAM on RunPod (If using ai-tookit, uncheck "Low VRAM" otherwise it'll choke and die). I'm not sure about the exact timing cause I trained so many versions of this one, but I'd say around 20 hours to get a good result? (Sampling every 250 steps to check your progress)
Is it only one size or can your prompt for big, small, thick, etc?
I think the model itself should do it from alone if u prompt it. Because the model needs just needs to know the concept of it. And h3 is very powerfull
@AndyZocker It is good at understanding and complying but shaping a penis is another thing entirely. It is like Krea-2 with just the bypass filter and no NSFW LORA. I guess we will see.
You can, but more often than not, I'll probably lean towards BIG. I'm thinking its it was caused because I used too many closeups in the data, so it didnt learn the relative size properly. Will try to fix that in future versions.
@jouhellerxxx391 I have tried with 0.6 and 1 strength, same prompt same seed. The penis was normal or at least what the LORA thinks it is a very big penis. LOL
I also noticed something else. My character had tattoos, it was a reference to video workflow, not I2V. At strength 1 the tattoos had the tendency to disappear, even if I prompted for that later, I also tried it with the turbo LORAs and without. Also ... and I will just shut up (LOL) the penis had different color than the skin, slightly different, lighter color.
Anyway, the penis is fine, thanks for your work. I am just giving you some feedback. :-)
@aferventu807 Haha, no worries, all feedback helps. The funny thing is, I REMOVED all tattoos from the dataset cause my first two versions were adding tattoos without being prompted. As I said, I think the model is TOO good at learning stuff and now It thinks it should do all clear skin. Hurdles for the next version, for sure.
(Also, I did not test Ref2Vid at all, this model does so much shit... just kinda fogot)
But hey, if the only problem we are having is that the dick does not come out exactly as we want it, it's progress 😅
@jouhellerxxx391 It is progress for sure, and the whole training of Minimax is a whole new territory, as for all new models. Today, I started experimenting with ref2vid, until today I was experimenting with i2v. The model is unbelievable and the greatest potential is in ref2v, since you have everything under control .... if you have the patience! hahaha
"focus on the dick (which is kind of anoying when you don't want that)"
Such a w e i r d feeling: I can see the words, understand them as individual words, but when I try to combine them and understand what they mean as a sentence, the result is something incomprehensible that doesn't make much sense.
It's a HILARIOUS problem to have in this particular instance, yes.
good
the king!
Thank you for sharing!
Awesome! Would be great if you can a 'cut' one for circumcised schlongs
Yes please! 🙏
love uncut haha thank u x
Straight up one of the best LoRAs I've used.
So far, this is the the only lora for H3 that actually works as described. Excellent!
It works perfect with T2V, but with I2V it can sometimes completely change a characters body/skin color or just completely override the ref image altogether unless you really turn down the strength (e.g 0.3), especially when it comes to toon and/or anthro characters. The effect is even stronger when using a turbo lora. Fortunately the motion still works well at a low strength.
Uh... yeah, I imagine that can happen. Never considered the anthro thing....
I'll keep it in mind for next version, and i'll probably just make a different ITV version, if Minimax already has the dick it needs waaay lesss nudging from the lora.
So fast so good! Very versatile and easy to use!