NaughtyTimes MM H3 Lora
Join our discord: https://discord.gg/GCrr4Cj3D
Use the right lora for your base model (pruned or unpruned), you may get "something" but if its not trained against that model it won't work right. Im adding pruned v2 loras now since people were asking why "It doesnt do anything" yesterday (this is why)
*** V2 = trained on unpruned model! ***
*** V1 = trained on pruned model! ***
V2.0 unpruned lora is rank 128, 256, and 512 extracted from PinkCherry base model. It retains about 24%, 34% and 48% respectively of the base knowledge in PinkCherry fine tune. It doesnt have the distilled issue of v1.0 so you can run it at full strength if you want. Try different ranks and strengths to see what works or doesnt for you.
I dont know which v2.0 lora rank 128,256 or 512 works best. Ive been testing them at str 1.0. Sort of a try different ranks and strengths and see which might be the better combo.
Don't mix it with PinkCherry base unless you want crazy, use the base model with this lora.
Pink Cherry Base Model: If you want to try PinkCherry Base model its at https://civarchive.com/models/2838593/pinkcherry-mm-h3
V2 Base Model: unpruned fl2va
V1 Base Model: pruned fl2va
(confusing I know, sorry!)
Lora Strength: 1.0 for v2
Lora Strength: 0.5 for v1
Training terms: vulgur captions, not medical terms. cock, pussy, fucking, cum, doggystyle, blowjob etc
Strength: 0.3 to 0.5 for V1
Strength: 1.0 for V2
my videos are just quick low step generations, which I used to test the lora with.
I didnt use any image conditioning (text to video training only), but it works on I2V as well it seems. Again this was a fairly quick "does it work" training session before I started a real model fine tune
Enhances wide variety of nsfw concepts was trained on about a large dataset of 4k videos, use at str 0.5 and on the pruned int8 model as that is what I trained it on. This version is alpha quality so temper expectations. Ill train this one up further after im done fine tuning the base model.
Description
v2.0 pruned lora
FAQ
Comments (24)
i am using the alpha-v2rank512 version, tested your PinkCherry model too, it is very good for an alpha. only real issue i am facing is slowmotion, and some "short" thrusts on penetration scenes, not going deep enough. the slowmotion gets better if i increase the output video fps. i create base image on krea2, then pass it to h3. 20steps on h3.
keep up the good work @sexgod1979 i really appreciate that you are doing all this for free. when you release v1.0-stable of the PinkCherry model and this lora, you could make them early access, then those who want to support you can donate buzz, and then it goes free after the early access phase like normal early access does. i will give some buzz if you do that.
it could be too high of a rank, thats a lot of capacity. Try like rank 128 and then 256. keep in mind its a lora, its not the same as a full fine tune PinkCherry base which retains all its learned behavior. The fine tune model will always beat the lora, every time.
@sexgod1979 is there a possibility to make a bf16 version of the v1.0-stable of PinkCherry? (i mean when you are at stable phase, we are at alpha now) i know that might be hard since even the comfyui version of bf16 is like 60gb, not many of us can run it but some can.
these images demostrates well what i mean by that the motion is not very deep: https://civitai.red/images/139140813
https://civitai.red/images/139135882
@movis27344361 THanks, I will review those
@movis27344361 one thing I would consider is download the PinkCherry base model and try your prompt I2V or T2V there, try both. the base model has substantially more learned behaviors than this lora can store (its only a low rank lora), the lora itself has millions of paramers but the base model has billions of params, so its considerable difference.
I might have missed something, but I'm not getting anything with this Lora.
pruned or unpruned model?
Same. Seems to be barely any difference A/B'ing with it on or off
pruned for me. Tested at various strengths. No real difference whether it is off or on.
@potatometer350 it was trained on unpruned for v2. ill correct descriptiom
likely because youre using a pruned base model. pruned for V1 lora, unpruned for V2. fixing model description and adding unpruned v2 loras. this model is a bit of a pain in the arse
uploading pruned lora v2 weights now. so there are unpruned and pruned loras now for v2. sorry about that all, my fault here.
Same here. SexGod-NaughtyTimes-v2-rank256 makes 0 difference. exaggerated big, thick and deformed diiggs. Same motion behavior as H3. bad thrusting and shoving in Dogggystyle and others.
@denolim465778 you are probably using it in T2V mode, try I2V and compare. This model is brand new, asking a low rank adapter to re-write a 33 billion param model with only a few hundred million parameters isnt an easy task. T2V will get better over time, stick to I2V and I think youll have a better result (at this stage of the game). Or stack other loras, or use a nsfw fine tune base model. But thank you for testing, appreciate it.
@sexgod1979 just to let you know, i was not complaining. i know you're hard at work. just thought it my help. Yes tried T2V and ref2VA mostly. Not sure if i tried I2V, but i will. thanks for the reminder.
using the "unpruned" r256 on ref2v int 8 convrot has no effect, though the base model is not pruned.
Ha-pruna matata! Thanks.
I am using the minimaxH3-int8-pruned model in the T2V workflow, but using your pruned lora and unpurned lora is not effective
try PinkCherry fine tune 0.5-alpha-C on huggingface without the lora
the best h3 lora so far =) thanks a lot!
Just got the unpruned version. Head of the wiener still changes with movement. This was i2v. In my case the camera is orbiting the subjects who are sitting on a couch. There is no real interaction with the wiener. The camera movement alone is enough to make it weird. Looking forward to the next version. Thanks for your efforts.
Actually with further testing it's still not doing anything.
I'm attempting to use in ComfyUI. I have the int8_convrot pruned version of H3. I'm using the pruned v2 lora, but hardly any change with the lora?