LTX Example Workflow
https://paaster.io/698559390a7a8c3988ee9e91#1NG7332gvq5L4A9yTnKw4INMZ00tgOFaCAEFqU8d-as
WTF?!
Why not ;)
I'm back again with a new experimental escapade that doesn't seem to be done commonly on here: a custom CogVideo Wan Video LTX-2 LoRA!
Training (LTX-2)
This LoRA was trained on a dataset of 143 clips with hand-revised captions.
These clips came from 54 unique source videos:
Real Life (42 clips => 74 clips, ~50%) mainly amateur clips from Reddit/RedGifs/Pornhub, and a couple studio-shot videos
Anime (7 clips => 14 clips, ~ 10%) drawn/3d animated clips
Furry (5 clips => 45 clips, ~40%) 3d animations
Around 80% of these clips had their own audio.
Trained for 4,000 steps (4.5 hours on a runpod H200 SXM) using the official training script: https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-trainer/docs/quick-start.md
Rank 32
Training (Wan2.1 I2V [OLD])
This LoRA was trained with around 110 or so clips of up to 11s, some of real life amateur/porn videos, some of animations
990 steps (90 epochs)
Trained on 1 x A100 SXM4 using https://github.com/tdrussell/diffusion-pipe
Training (CogVideo [OLD])
This LoRA was trained with 19 videos of solo male masturbation, mainly amateur vids.
4,000 steps
Trained using H100 for around 18 hours or so, using https://github.com/a-r-r-o-w/cogvideox-factory
Description
FAQ
Comments (4)
I have a question about the data set. Namely, about the quality of the video and the details of the prompt. Well, and I also wonder how active the actions are in the data set video (because with this lore sometimes there are white spots and the hand movements are not very active)
Hey, DM'ed you quite a bit of info.
But for everyone else, dataset is here: https://gist.github.com/AnonWaterVole/667998b2d3f733403871c83a1c9ad94e
It's pretty bad quality -- this was a mainly "is this even possible?" experiment. I plan to make this again, but with longer videos, higher-quality labels, variety of styles, etc.
> this lore sometimes there are white spots
I think it's trying to do cum, but I guess it doesn't have enough examples on actually *where* it should be xD
> the hand movements are not very active
Yea, I think this is because the training videos are like ~1s or so + I don't differentiate between "stroking speeds" ;) -- I will try to fix this in my re-make.
@Spat1984 yes please, i love you <3 🥹
@Spat1984 The repo no longer exists, can we find the dataset somewhere else for uh research purposes...