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

A few tips I found after experimenting with using a video as a reference for character replacement with added customizations :PepeHmm:

  • Turbo loras aren't worth it for how much quality and prompt coherence loss there is.
  • Generating longer than 15 second videos still often works but the prompt coherence starts degrading more so you'll have to generate more times to get the one that works.
  • There is a pretty specific prompt format that helps keep your character looking like your image. See my latest examples for what I've been using.
  • If you are simply copying the audio from the video reference you can get away with 8 steps. If you are generating your own audio or modifying the source audio, steps become more important, often with at least 12.

Ways to best speed up generation times with minimal quality loss:

  • Use the latest version of Comfy Kitchen Attention
  • Downsize your reference videos as much as possible (as long as the key details are clearly visible). I've been using 300p resolution videos with 20 fps as a base. The smaller your reference video the faster the generation times. If you are doing a cum facial for a reference video, you will likely need higher resolution and frame rate to pick up the needed details. A simple fucking motion video can be much smaller and less fps.
  • For character image references, don't use extremely large files. I've been using 1000 x 1000 pixel jpegs that stay under 1mb each.
  • For your video VAE use minimax_h3_video_vae_int8_convrot and start ComfyUI with --fast fp16_accumulation for even faster encoding and decoding.

For my latest examples with my latest settings and prompts being used, see my posts here:
https://simpcityopen.com/threads/235409/
 
Throwing multiple references into one 1000x1000 (or any size really) collage also helps as the model can keep track of everything. Frees up reference image slots among other things
 
None that can be shared here but it's my go-to method for everything I do. You can place a character, their outfit, the environment, and even their nude body parts in the collage and it works. Label each panel, describe them, and the model will handle them correctly.

Edit: The last video I shared here actually used one image for multiple panels from a webcomic that had different things going on and the model handled everything perfectly. You can have multiple subjects in one image so that's pretty much the idea
 
Can anyone recommend any good image2vid AI? I’m a novice at all this I literally only used Grok and that was in the middle by that point it had got nerfed then eventually it went to complete shit
 
everytime i generate video i get same person on video it doesnt change to other girl . i cant understand what am i doing wrong , tried different promts , only 1 from 20 tries maybe it changed girl to reference pic , all other times just same girl , maybe someone can explain what am i doing wrong pls
 
i tried to do same as posted here , took same videos and promts, also i noticed there is qwen promt node , do i need to disable it maybe ?
 
If you are getting any video at all its likely a prompt issue.

If this is from using one of my prompts I would check these first:
1) Is the character image you're using decent quality? If not, try turning your photo into a character reference sheet first and then use that as the input image. A Character Sheet with a front and back full body shot and a face close up typically works well.
2) Is the prompt correctly reflecting a detailed and accurate description of your character? You can have an LLM describe the image of the person to help create better descriptions for them.
 
Any good H3 loras for a natural looking penis? The dicks look crazy and I can’t find a good lora to make them look normal
 
my trick has been to make a refmod of my own dick (not going to get into explaining refmods) and it works on multiple levels!
 
Im trying to get a floor mounted dildo to work without adding loras, it can get the pose and dildo without issue, but it keeps animating the woman grinding against the shaft on the outside.

Oddly, if I use an animated source it works just fine, but any realistic image reference, messes it up.

Anyone have success?
 
Yes, and its pretty damned good even local. 15-30s of a good source will get you good inflection, emotion, even moans. You can compound one's you created to get even closer to the inflection youre after
 
Well it gives you a working penis that's not perfect. Not to mention the balls tend to get generated with a gap in the middle that resembles a vagina.

Get the turbo hybrid beta 5. If you want a better looking dick stack on a reference and you're good to go
 
Sorry if wrong thread for this , but has anyone used ltx 2.5 .I saw a video of someone making an (image to video) 5 second video (sfw) in 20 seconds locally .I could be wrong but this is really fast .First of all is it uncensored or does it have uncensored version of it .How does this compare to h3 minimax .Also are there other models faster than this .Is there any way to run it on an 8gb card .

this was the video btw :

its a month old with how fast ai is advancing maybe there are better models?
 
You must log in or register to reply here.
Back
Top