TL;DR: This LoRA helps with skin texture, the eyes and the hands for Minimax H3: t2va. It's the final LoRA model and might be already overtrained ... so use it with lower weights. Just experiment!
Okay, folks,
I don't really know what to say other than: MiniMax H3 is ridiculously good. <3 The hype is real. I've spent the last few days messing around with both models:
ref2va (reference/image to video) [Get ready to meet your maker!]
t2va (text to video) [here]
So here's my first LoRA — the bread-and-butter one: skin, hands and eyes. And I hit a wall almost immediately. We need to be a bit careful with LoRAs here, because it's the same story as with Wan 2.2: the base model already knows(!) what all of this is supposed to look like, and that can clash. Anyway. They do have an effect, but the whole thing is very prompt-sensitive. Keep that in mind.
The current videos were made with four LoRAs, two of mine, two external:
*polyhedron_minimax_h3_image_lora_v2__nippl8es // (currently only up on my Patreon — we'll be needing this one later, wink wink)
https://www.patreon.com/c/polyhedron_ai
*polyhedron_minimax_h3_image_lora_v2__perfe8ct // (this is the LoRA here in its final version; if it feels overtrained to you, all the training steps are on my Patreon as well with the workflows I used for the example videos)
And then two external ones. First:
That's what got me started on this LoRA in the first place — though the workflow is an absolute mess. So I'm putting up a clean, tidied-up one that by now runs entirely on Polyhedron Suite nodes. The quality speaks for itself. Go try it. I've also gone and pulled the whole MiniMax H3 universe into my own repo, which you can grab here — updated today — and it's available through the ComfyUI Manager too:
https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack
And here's the second LoRA:
https://civarchive.com/models/2858760/minimax-h3-tiddies-and-realism-slider?modelVersionId=3229050
Tiddies! Already in there, and we'll be needing that one later as well. :D
We've been hearing "this model is the real gamechanger" for years now. This time I actually think it's true. :)
Catch you later!
Frank
Description
V1.0
FAQ
Comments (14)
Your description is long and describing a bunch of different other things. Your videos just show someone crying. What exactly does this lora do?
realistic skin and better eyes probably what bothers me is there is no recommended weight ( lora strength ) and settings in the description
@JJoker375 just experiment!
and what was meant by ref2va (get ready to meet your maker)? I am confused about that, I use it all the time for extending videos and it is amazing (if you get the prompting right).
@olivetty is this the final version or you plan on mor realism loras cuz i think minimax right now needs more of these
She is crying because she is constipated.
@sarahbonestone I SEE
@JJoker375 MiniMax H3 Power LoRA Stack automatically puts the strength at 0.64....
this lora helps a lot on low resolution, less artefacts, big thanks
thanks
Hello, I have a couple of questions. Do this lora works for Ref2VA or only T2VA? Which is the suggested weight to use? Did you train it with Turbo Loras or also with the base Minimax Version?
Hi there: I guess, it's only trained with the T2VA-Model (with Ostris on runpod) because only still images were used - but maybe it has also an effect on Ref2VA? No Turbo LoRAs were involved ... I use a weight of 0.4-0.6, but test it for yourself ... :) Have fun! Frank
@Polyhedron_AI Thank you very much for your reply. I will test it right now. I hope everything comes fine
@mankogamerhn Yes. I am also testing, especially Sampler/Scheduler-combinations like "res_multistep" with "Simple" with no Turbo ... Keep it up. :)