About this version (qwen v2)
A brushup of the dataset, with better result than V1 for qwen. I will use the same dataset again for zturbo when the base model is released.
Trigger word: PENISLORA
What can this lora do?
This lora can add erect penises to both men or women viewed from the front/side. Other angles such as POV may have a backwards penis head.
Other things it can now do:
Side view of the penis
Cumming / Cumshots
Blowjobs (its captioned for the words "blowjob" and "deepthroat" )
What can't it do?
No penetration in the training data. Also nothing from POV angle, though there is a few images from above and 1 POV video in the training data.
Sometimes blowjobs with cumming have the penis slip out the closed mouth.
Recommended Settings
It works pretty good with the new lightning dyno high model. I'll link to it in my example workflow. I like to use dyno high model (no lightning lora), then for low I use the lightning v2 lora on the regular 2.2 low base model.
Dataset
84 images at 512x resolution
43 videos at 256x resolution
(I let DP pick the aspect ratio automatically)
This is the same exact dataset as the 2.2 5B model. I made no changes.
Training
I used the default diffusion pipe settings.
[optimizer]
type = 'adamw_optimi'
lr = 2e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8
I was baffled why it was taking so long to train the high until I realized after over 60 hours of training that I had put my videos in the images directory which resulted in the high being trained ONLY only on videos and twice (once with a very high resolution). Once I fixed this, I went back and trained from 11K steps up to around 13K with the images in the training data. The high model was fine without to be honest.
For the low, I trained it properly with videos and images the whole way, around 6K steps in I upped the image resolution from 512 to 1024 actually and didn't get an OOM (it fit around 24GB exactly). I trained it to around 10.5K steps. Also I trained the low on the full timestep range (0 to 1 instead of 0 to 0.85) from some advice, it may switch better over from high to low on the speed up lora with low steps.
I think I might do another version with more angles such as POV and from the behind to make this work for any situation. In that case I don't think it needs 10K steps per training session, epochs around 5K steps looked fine.
The results
I think it was a combination of improved captioning and 2.2 base model being better. But this lora turned out really well.
Description
Trigger word: PENISLORA
See notes for the 14B version about prompting. The same principles should apply.
Warning:
This lora is still not fully trained, but I have trouble testing it, so I will release now for people to try, and will probably release an updated one in a day or two with more training.
I trained this for around 60 hours at 13,200 steps at 60 epochs locally on my 3090. 5 repeats for the video and images respectively.
The training data is 100% same as the 14B v0.7 lora except that one was trained by loading weights from 2-3 sessions of training while this one was resume from checkpoint (finally got it working). Also I tried to reduce the video resolution from 640x480 to 256x192 which is a big downgrade.
It was trained off the default 1.3B T2V model from Wan.
I couldn't really test it properly, I have no experience with this 1.3B model and no one in any discord would reply to me or even give a shit about me or my questions or problems regarding best settings. So I'll release it in this semi-broken state, and probably try to train for another 10-20 more hours/epochs and release a v0.7b before I give up and go back to improving the 14B lora. Please if you do have a workflow for 1.3B (native or kijai nodes) please share so I can properly test it. (I added my workflow example in the data section) Or please generate and share good results here so I can see it turned out ok. I think its really hit or miss in this current state, and maybe it will improve with another 10-20 epochs.
Currently it tends to give hanging erections rather than standing. The only thing I noticed is I had to adjust my CFG down to 4 from 6 to get it not to look all saturated and pixelated.
EDIT:
I got some help with 1.3B workflows and I have solid settings now to test. I will do another training run over night and update hopefully over the weekend with a really well trained version of this 1.3B model.