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

I tried Use 10eros turbo hybrid beta 5 in place of the turbo and mystic nodes and the dick always comes out deformed looking, sometimes upside down but always with a black ring around the head. I tried 0.8 strength and 1.0. Tried 4 steps and 8, both had the same issue. How long do your 15 sec generations take at 1 MP?
 
It's all in here... if you want to do this, you have to read and study this thread, no1 can give a simple walk through.
 
The formatting is straightforward and you just feed it to an uncensored llm that will then turn your natural language idea into a working prompt for the model. There's no reason to copy someone else's prompt considering all assets will be different from video to video. This isn't like grok where one prompt could be applied to many i2v's in order to circumvent moderation
 
I have these specs where should I start for genning local vids/using ref images of someone and swapping them into a ref vid.

Ultra 9 285k
192GB Corsair Vengeance (4x48) 6400mhz
MSI 5080
 
For ref2v the prompt details do matter a lot and what you are trying to do with a reference is very specific. And using LLM's does not work for many specific cases as the official docs don't include everything and people are continually figuring out its what is possible. For trying to achieve very specific actions that are not details in the official docs, being able to discuss prompt techniques can be very helpful.
 
Give the thread a good read. You'll find all the info here. Get familiar with comfyui, workflows, models, nodes etc.

Also ref vids are very taxing and even though you do have capable hardware it's going to take a long time generating what you are asking for.
 
Yes that is what this very thread is for. The full prompts themselves aren't that important. The tools are already here and the only limitation is really the creativity from users for their outputs.

The techniques are important though like the one someone posted about here for referencing pure motion in videos. That is something I have seen going around in other places and it works really well to avoid any bleeding from reference sources.
 
Which loras did you use for the pussy and the licking/fingering and what strength? Any turbo loras and what strength. Also number of steps?
 
Can yall give me some of the best lora you can find for anything ? (ass, boobs, pussy, anal) I'm stacking every good lora idk if its a good idea or its gonna duck
 
Stacking loras, especially unused ones ain't a good idea, even the lora strengths effect everything. Just use the 10eros model without loras like me but I'm lazy.
 
What is required to regenerate a video from a different camera angle after referencing a motion (such as a piston-like movement) using ref2va? While I suppose I could create and reference images of the same pose from a different angle using software like Blender, applying the exact same pose to two figures is far too much trouble. I wonder if this could be resolved using a prompt or a LoRA.

Also, on a separate note—if someone is dissatisfied with the low volume of ejaculate in the original video, what do you think is the best way to increase that amount?
 
Damn, apparently the AMD fork of comfyUI only supports RDNA3 or newer. First time in almost 20 yrs I decide to go with an AMD/ATI card, bah.
 
Detailed and specific prompting will give you exactly what you want in your scene. Articulate your thoughts and ideas as best as you can
 
A tip for when using ref2vid with a video reference:

Since the reference video is providing the motion and movements you can get away with only 8 steps if you are using the reference video's audio instead of generating new sounds. If you are adding your own audio or voices, around 12 steps gets it to sound decent. Turbo loras aren't needed in these cases and generally just make the results worse.
 
So a fun thing to do with the model is have it bring your favorite webcomics to life. You can use the panels themselves as a storyboard for each vid. Being able to reference voices is also godsend since the generic ones are really harsh on the ears after a while. Now for some reason the model (10Eros max turbo hybrid beta 5) simply refused to simply not generate nipples/areolas over the female's shirt. It was totally random. In some scene it would not do it and in other it was impossible to prompt out. That along with other issues made this a longer process than I feel it should've been. In some cases poor prompting was the issue. Anyways here's a rough draft and my first attempt at making a webcomic Chapter into a full vid.

 
How do you control the camera placement so well? With all those cuts and stuff. In my experience its so random, applies my prompt if it feels like it lol.
 
Kinda just guiding the model with detailed descriptions in the prompt. A lot of the angles were already from the webcomic panels so it just came down to describing where the camera was positioned at, which direction it was aiming at ect. Generated way too many videos in order to get things right especially since my prompts were lacking most of the time which made the model be really confused about the placement of things. Good thing this whole thing was made locally and not some money sucking api.
 
You must log in or register to reply here.
Back
Top