NaughtyTimes MM H3 Lora v3.0
Join our discord: https://discord.gg/KjWkmcdeE
NaughtyTimes v3.0 (fp32) is the same trained model as PinkCherry H3 v1 final. One is a lora, one is a full checkpoint (fp16/bf16), you decide. There will be no PinkCherry v1 int8 because the lora has higher precision and 98% weight retention.
ComfyUI/Lora Loading: Use the standard lora loader in comfyUI and latest comfyUI that has the updates to apply the adaln unpruned keys. otherwise youll need the pruned version with the adaln modules/keys removed. I dont think the power lora loader node has adaln h3 modules implemented yet?
There are unpruned and pruned versions of this lora, see below for reasoning and which one you should use. If you can run unpruned, you should. There are 13 billion params in the unpruned model that I removed in the pruned lora so they would load in ComfyUI on the pruned base.
Scheduler/Sampler: Euler/simple (not beta, not res_multistep etc)
Workflow: Standard comfyUI i2v or t2v workflow is fine, with euler/simple (not beta, res_multistep etc as it was trained with simple sigma schedule)
Steps: I normally run 30-40 steps
Strength: 1.0
Resolution: I use 0.7MP or higher, as this was trained at high res
Turbo: have not tested it, may work may not
Model Modes: Both I2V and T2V were trained
Pruned vs Unpruned
Most other loras dont train the unpruned modules that I do, which is why I have unpruned and pruned versions of this lora. There is a good reason I train the unpruned modules, because they are very important in Minimax H3. Long story short, run the unpruned lora.
Pruned vs Unpruned: The unpruned lora version contains significant alterations to the unpruned model, if you can run the unpruned base model you should. The pruned lora model has all these wonderful changes stripped out completely and is not ideal (it will not act the same way as the unpruned model). But I will leave that choice up to you, I only run the unpruned base model. The unpruned model has all the Adaln projections intact and are important for this lora since I trained those
Description
it was trained on distilled version so dont go over 0.5 strength and use the pruned int8 mode (not the ref model) first frame last frame model
FAQ
Comments (51)
I might as well ask here too since you didn't really clarify, is it still the 1500 step lora, or has it been changed since the first release?
Its the exact same lora, I havent trained it further. Using the GPUs to do fine tuning now. will swing back to retrain this lora later after im done
is this the same as "pinkfluffy" from your HF? just want to make sure I dont duplicate it... ;)
Yep, same SHA hash
yes
What is this lora specifically do?
It looks like its a general NSFW helper.
@KiraNugget Thanks.Have to try it.
it was just trained on varied sex clips, nothing specific
What's the point of this lora? If the initial image already contains everything you need, a simple, good prompt is enough.
I didnt find it was enough, but I guess everyone's milage varies.
also I mostly posted it because people were asking me on chat to post it as huggingface took it down. I wasnt really sure I was going to bother as im already working on a fine tune
It doesnt know how a penis interacts with a vagina very well or what blowjobs should sound like. Its thankfully very good out the box just needs a proper final lap to really unlock good nsfw imo
I feel like we'll be saying that about a lot of loras for this mode. We finally have a model than can do just about anything if you write your prompt correctly. I'm gonna be waiting for some stuff to drop that can definitively be proven to generate something impossible for the base model to generate before I go download crazy. This model is freeing up a lot of my hard drive space lol.
Please change the name to just 'H3' without company name. Thay have problems with that.
H3 is TRITIUM, and that should be the *"I don't know what you're talking about but this isn't porn trained off your model"* name.
Thank you! But I get the feeling that this is a model that will be very difficult to uncensor.
The vagina lora is decentish
Uncensor?! This ist the most uncensored video model i know?!
@Lora_Addict it truly is, isnt it wonderful?
@sexgod1979 it's fulfilling almost all wishes, indeed wonderful!
LMAO, you know that this is the most uncensored model we've ever gained?
I want to ask if this lora can let me generate more sex positions, I want to generate 69 position, but it didn't work either way
it wouldnt be significant enough in the dataset, likely need a specific 69 lora for that
The GOAT 🐐!!! 🔥
Thanks for this. Now that I know this does t2v good; im switching my 6000pro back to comfyui. ;)
Hi Bro how long 6000 pro take for 10s. Is it able to do 30s too?
@Apprehensive6311 I haven't pulled my 6000 from my llm box yet. But 5080 takes about 35s at 512x384 for 10s using the turbo lora with 6 steps.
what about REFRENCE TO VIDEO?
nope, wasnt trained on it. That model is different, different weights.
I think you can try with a lower strength like 0.3
You can do it but expect the quality to take a hit.
I did try reference to video with the model using two images like <Picture 1> etc and it did actually work even though this model isnt really designed for it. Was happy with the result
are you guys just using "load lora" node and hooking it between the model and the 2 model inputs in the default comfyui workflow? i can't seem to get it to work, it always seems to make the video output and audio quality worse, and i can't get a normal looking penis at all. Is there a special "minimax h3 load lora" i'm missing?
I'm just using a normal load lora node between the pruned h3 model and the basic guider and basic scheduler nodes. Try using only at 0.3 strength
yeah try 0.3 it was trained on distilled base which wasnt a great idea, so youre getting overcooked outputs
Or try 0.5 strength and 10-12 steps.
Just test a simple doggy sex scene in R2V mode with my own custom nodes and H3 workflow.
If my futanari with her big erect penis and physical consitency is perfect, the sex act itself is totally crap.
The scene was very simple
Ther futanari is on all four on her bed, and a man is behind her assfucking her.
My duta is on all four, the bedroom is perfect, but the man and the sexual act, are totally crap.
So, i think this will need more traningand maybe add triggerword for each positions.
not bad at all, the penis is still odd looking, mostly the head/tip area, but better than the flat model, thanks for doing this, looking forward to V2!!
the result is not good enough as I expected. Can you share the workflow you use, please?
I get better results with the convrot models without loras. Natural and fluid and it doesn't alter the faces. I'm sure we'll get great stuff moving forward but LoRAs need more time as it just released. I'm sure we'll se some exceptional stuff in the future!
Hi Good lora so far, just wanna report, not working with turbo lora. ^^ keep the work up thanks for sharing
"Is there demand for a similar ref model lora?" Most defiantly.
Anyone tested this and confirmed if this lora is a troll and needs banned or legit?
All his prompts are just troll prompts that have nothing to do with his example videos.
It was late and I had to get to bed and just didnt have time, but just load the video into your comfyUI and the prompt is there in the meta data. Also it would be nice if civit could take the extra 10 seconds and code in a reader for meta data so I dont have to do the work for them 1000s of times a week when I post models :)
@sexgod1979 I downloaded your workflow, thanks for sharing. The LoRA isnt actually in there? Is it buried in some weird node? I dont see it used
@SurrepticiousFabric try this image for workflow, I think the lora is in there buried behind the subgraph (doubleclick subgraph to see it). This is plaguekind's workflow I got off civit but v1.0 didnt have a lora loader so I added one
yeah sorry all ill try to make sure I give my prompts from now on, just tired lazy on my end
@sexgod1979 exactly... there is even an open pr on civitai thing on github... it's been years. They should have added video metadata years ago.
civit can't find your prompts in the metadata 90% of the time and will happily slap you with a "can't verify if this is AI please provide a prompt" Which makes it really annoying to upload anything.
90% of the time if you see a prompt that doesn't match, it's probably not done out of malice.
lmfao why don't you fucking test it.