text2vid img2vid H3 Minimax - Fully uncensored local AI video generator released today

no it doesn't, as in it generates the prompt using vram then the model unloads, no vram usage difference during "normal" workflow process.
 
I'm looking to do some video reference swapping, is there an easy way to feed a video to an LLM and have it produce the text description of the video for a prompt? I don't want to write out a detailed description of each video I try to swap.
 
Use grok for that. It's fully uncensored for generating prompts with all your references (image, audio, video). Feed it the official prompting guide for ref2va first so that it generates your idea draft into the correct format.
 
I'm mostly there with everything but anyone else getting weird artifacting when doing ref2vid? Like the whole thing's smeared with vaseline sometimes, Doesn't seem to be really resolved by just increasing the MP unless 1-1.5 is still super low.
 
Heyyy I'm really new into this AI stuff,is there a full tutorial on how to install and use (Lora and everything) ?? I'm looking for some video reference swap
 
these days it's probably easiest to download the desktop version of comfyui that installs everything nicely for you. AI Search on youtube has a tutorial for minimax h3, he always explains everything pretty well for beginners, including installation and stuff (he has a separate one for comfyui although you don't really need that if you download comfyui desktop). Ask chatgpt to help you install the spectrum node and connect that and the comfy kitchen attention thing from the modelattentionbackend node. just copy and give it my text and it'll know what you want. But use a reasoning model on chatgpt if possible
 
I would also suggest finding workflows. Pixaroma on yt has some good ones and it's basically what I've been using as a total beginning to generating locally. Once you get familiar with comfyui and how H3 works you can start generating your ideas better.
 
Now that I've tried this, lets just say I get the alleged RAM shortage an I've got 64gb.
I struggle to get sharp images and they often spas out so I'm a little struck on what to change.
 
do we need any special stuff to make the swap our character into porn vids lol
 
Besides feeding the ref2va prompting guide into an llm like grok and then getting all the formatting right for the prompt you really don't. You need a reference image of your subject and the reference video and that's it. It's not perfect but it works.

The whole experience is a big learning curve but it's totally worth it. Also you need decent hardware to run at it home
 
If you're running it through a lightning Lora this is probably the culprit. It makes things way faster but the image degrades. I'm still learning but my general workflow is generate something running it through a lightning Lora to make sure the prompt is decent, then turn that off and try 10 or 20 steps without. Has massively improved the quality of the overall result. After that obviously you can also increase MP
 
does anybody have a good prompt for a full nelson? kinda having problems getting it right
 
I'm having issues where I tell grok to make me a prompt using the guide that was posted here and every single time I get these weird circumstances happening in the video:

The man is riding the woman and she has a penis.

The man has a penis, but he also has a vagina.

The man is playing the role of the woman fucking the woman.

The woman has a penis coming out of her stomach.

So, what I did was, I told Grok to ensure that none of these things happen and to correctly portray the woman with a vagina getting fucked by a male with a penis, but even with that, while it gets it correctly, the man doesn't have a ballsack, but a vagina.

I don't know if it's because of the LORAS or Grok being bad prompting when I tell it to also be explicit, detailed and specific, but man this stuff is annoying as hell. Any tips?
 

Crude idea I'm working on. Facial and post shower clean scene will probably be added sometime down the line. The subject is an AI generated female
 
Sounds like it might be the LORA weights. If you're running for example, lightning lora > base xxx lora > specific xxx lora eg. detailed vaginas, try setting the weights to something like 0.8 > 1.0 > 0.35

You also want to look at how many steps it's running, if it's none lightning try 10 or 20, if it's lightning try 4 or 8 etc etc

I'm still learning a lot but the prompt only goes so far, start with a basic prompt and then generate a ton of very short clips with different settings to find out what works. then you can go from there
 
You must log in or register to reply here.
Back
Top