SexGod PinkCherry LTX 2.3 NSFW Checkpoint
Model type: checkpoint
Training: full fine tune LTX 2.3 dev base
Supports: I2V, T2V, V2V
Distil lora + Workflow available at:
https://huggingface.co/SexGod1979/PinkCherry_NSFW_LTX23/tree/main/v1
ill continue to work on this model and add more niche content and styles/actions (I am taking requests!). Will continue training this model with more style/actions, let me know what you want!
Description
FAQ
Comments (74)
42gb? How can you run this? Is the minimum a 64gb-128gb strix halo/mac m4??
Because even a RTX5090 cant run this right?
true, I have an RTX 5090, ill create fp8 and nf4
an RTX 5090 can run the full 45GB bf16 yes (I have one and run it all the time even the LTX 2.3 dev directly). But generally yeah thats a heavy one for the average person, ill create a nice fp8 and lower quant version
around 42gb is the size for undestilled ltx23 models. with offloading you can run pretty much any model, assuming you have the total ram you need for everything and it supports the hardware you use.
You can run it with e.g. an RTX5080 with 16 GB VRAM/64 GB RAM using Dynamic VRAM.
You do not need to fit the model into VRAM, only for LLMs.
I'd recommend learning how to swap/chunk your models. You'll lose some speed but open yourself up to a whole world of things you might be missing out on now. You can do it with native drivers for nvidia in some cases, and nodes/modules with your front end of choice.
@sexgod1979 First of all, thanks for your efforts and sharing with the community. As a newbie, I asked some GPT's weeks ago and they said RTX5090 is not capable of running 42 GB, so i never tried. Since you said you use it all the time, i would like to try your checkpoint.
Question1: Is there anything i need to take care of beforehand, like special custom nodes, changing settings in different workflows, etc. or do i just use the 42GB as an usual model with a RTX5090?
Question 2: Is there really a noticeable difference between the 42 GB and fp8 version? If yes like what?
please do the Q5 quantize ver!
right, yes on it!
@sexgod1979 A Q8 version would be nice, too—if that's not too much to ask :)
I'll have to convert this to Int8 and play around with it. Thanks! Does it to about the same as usual for T2V or is there an improvement?
t2v works normally. I can create and upload an fp8 version. gimme a min
I posted an fp8 version
did anyone already test how this comares to sulphur/10eros? i don't have space for all these giga chunks 😭
this is probably just one of sulphur/10eros with some loras merged.
@MrFlex yea i guess, would be good to get some clarification from OP, what exactly have we here?
does it work well in wan2gp?
I havent tested any loras with it to be honest at this stage (its only v1), I have not compared it against sulphur/10eros either. In terms of training its a fine tune on a large dataset using custom training code
@sexgod1979 but a finetune of what? plain ltx 2.3?
@azra1l It has to be LTX 2.3 base model by definition or you couldnt even load it in an LTX 2.3 workflow. Look at the meta data in the safetensors file it has the lighttricks description right in it :)
@kronos1959777 I dont know, I havent tested it in wan2gp, perhaps give it a try and report back?
@sexgod1979 wtf... sulphur or 10eros? i am able to load these in a LTX2.3 workflow... the description doesn't help at all regarding my question. anyways, thanks for finally answering my question. tried your model, will have to experiment some more.
It would be so much better to just have LoRA's. This shift towards checkpoints means you and Sulphur/10eros are now just duplicating work that cannot be combined. (They should also be making LoRA's.) I get that people like the "easy" of checkpoints, but it creates so many problems.
unfortuately that method doesnt work when you fine tune the model, because youre changing 22 billion parameters not a small lora subset of weights. LTX 2.3 doesnt handle nsfw well so hence why eros and sulphur decided to go with fine tuning. Ive tried LORAs as well on ltx 2.3 base model but have been less than pleased with the results. If you find loras do the job though thats an option. I prefer a full fine tune method to change the entire model weights
@sexgod1979 Nonetheless, you're both layering custom data on top of that, and that just creates silos that don't play well. It's just a bad route for things to go. Having a bunch of diffuse standards is plainly bad. That's why Wan rocks so hard. You can essentially use every Wan 2.2 and 2.1 LoRA. Acknowledging that LTX-2.3 sucks as a base, it seems like coordinating on a base finetune and then working LoRA's on that is the better path.
To be honest, right now I'm still doubting why sulphur (not counting 10ero because its base comes from sul) would be fine-tuned purely on data that brains it towards NSFW. Because it will cause the model quality to drop heavily compared to making a LoRA and merging it. (If want to fine-tune, the dataset should be balanced around 70% SFW and 30% NSFW meaning SFW data must be included to prevent NSFW bias.)
@boobkake22 all valid points, I do this simply because its a checkpointn that I want, and I figure its something "someone" else may or may not want as well so I share it. I cannot speak to the larger ecosystem, just something fun for me to create and have with at the end of the day. SFW and other loras may work very well with it, I honestly dont know I havent had the time to test that or worry about it. I can say this with confidence, LTX 2.3 base 'sucks' for NSFW, it has no priors. To get those priors you have to fine tune that model, i've made enough LTX 2.3 nsfw loras at this point to know its not a good base for it and I dont have interest in fighting the LTX 2.3 for nsfw content anymore, I would rather mold a fine tune a model like this instead.
@sexgod1979 does this checkpoint maintain compatibility with loras well? I see ppl making some loras with a sulphur/eros version
@sexgod1979 it would be good to know what your checkpoint is made off. what is the base? what loras did you merge? anything?
@snake88 I combine LorAs fine with all of the models mentioned. Not sure what people are on about. Adjust the strengths. You can't just plug them in and go, that hasn't been true for any model. Some LORAS and Checkpoints require a bit of tuning on the user's part to use.
@snake88 I would expect so, I havent tested any though, but its a normal LTX 2.3 base. I didnt merge any loras into so it wouldnt have a negative effect on those loras but its possible the gradients wouldnt be perfect
@DaddyWolfgang Yeah its all LTX 2.3 base ultimately so I dont see why loras wouldnt work with any base checkpoint
Will test this out now, demos look really promising! UPD: Yeah, really works well. Distilled LoRA at 0.48 for video and at 1.0 for audio - fits fine. Good motion, good sounds.
thanks, I tried to get the motion and audio to be good but its still v1.0 at this point
can this do trans girls/ futanaris? or does everyone have a vagina that's receiving a penis?
any pluses/minuses vs Sulphur/Eros? How well do sulphur/eros Loras with with this? I know it's a big task to finetune a video model.
I havent used sulphur/eros for more than a handfull of videos so I wouldnt be the person who can say if its better or worse. My experience is with the LTX 2.3 base (building NSFW loras) in terms of that I believe this is better.
Will try it out and I really appreciate your efforts. As for requests,... With everything else out there, you'd break new ground if you train your model to understand the concept of what an average penis and vagina look like. Who knows, maybe you've already accomplished it here. I mean, is it really too difficult to use some WWII combat movie clips with closeups of soldiers to understand which way a glans faces??
you might have to explain this a little, as I got lost on the WWII stuff. Like Saving Ryan Privates?
As in... Learning the correct way to wear a fucking helmet! LOL
WW2 imagery truly is the best source for this.
Thanks for the fp8 model !
no problem
You should clarify some things for the community because it's currently confusing and vague.
Is this model a separate fine-tuned version of the base LTX 2.3, or is it a fine-tuned version of the Sulfur 2 - 10eros models? If it's a modified version of the aforementioned models, what is the essence of the improvement?
"You should". rofl.
@sexgod1979 I don't want to offend anyone, but it seems like you're an idiot or just pretending.
@sexgod1979 Okay, Zoomer.
@sexgod1979 Seems you had a fundamental misunderstanding of the initial question: Did you take LTX-2.3, the base model, and fine tune from that? Or, did you take Sulfur 2, and fine tune from that? Or, did you take 10Eros, and fine tune from that? I didn't necessarily see the ask to "compare", other than, if you tuned from Sulfur 2/10Eros, what did you "tune" on top of them with, since they already have some understanding of sex, did you do a general refinement of broad concepts(like anatomy), did you focus on improving different positions, stuff like that.
That is how I took the initial question asked.
@sexgod1979 I'm working with Tenstrip (10Eros) workflow/checkpoint for the last few days. I had a really hard time to get a decent output. Now, with PinkCherry the result is absolutely amazing. This is BY FAR the best model I have used. (BF16 on RTX5090, 96GB ram).
A few changes that are helping me get even better results:
- I've added one more step on the 2nd ManualSigma: 0.65, 0.6332, 0.3425, 0.18, 0.0
- Disable/bypass the VAE Decode (Tiled) node
- LongerEdge slider on max (great addition in your workflow!)
- Added IC-Lora-Detailer with strength of 0.5
- Not sure if this was default in your workflow, but I am using the Rank_111 distilled Lora on 0.6 strength
Still tweaking though, sometimes I get mulitple belly buttons - I think this is a aspect ratio issue on my side.
I hope you keep updating this, it's absolutely golden.
A big thank you from me!
@09franc_aroma872 thats helpful thanks, I didnt really get time to play around with workflows, appreciate the feedback
If possible, could you provide support for various fetishes? (pee, farting, etc.)
ill see what v2 might have in store/
Now there's an underrepresented niche!
@aurelius haha, you're not wrong
Seems to have better facial retention/consistency than Eros by far, though maybe it's because it does less in terms of distance from base LTX. I like it so far, though.
I used some i2v conditioning params in my training pipeline to try to keep faces or image inputs consistent through generations. but I didnt want to over write t2v either so kept it reasonable
@sexgod1979 maybe you can make a i2v only model and go all out? that be great
您好，这个对男同内容有效果么，还是专注男女之间。
thanks for making another alternative base model. we need more of these in ltx 2.3
2.3 1.1 or basic ?
hehe ! this checkpoint is amazing. Lora:distilled-1.1_lora-dynamic_fro9_avg_rank_111_bf16 --strenght 0.6
Experiment add the lora "hentai_voice_ltx23_v2"
sounds spicy :)
What will the lora help with? Wonderful things I hope?
@Goonasaurus_Rex Ltx tends to generate voices lacking emotions. I realized that the voice lora performs quite well (and funny as hell).
This is a nearly perfect checkpoint!
very good face consistency, much better than the others!
very very very good job mate!!!!
Wow! Better than Eros and Sulphur. Tried it yesterday for I2V. Amazing work!
As for ideas: It would be neat if we could prompt for scene changes more directly in a clip. Like if we want to do x for 2 seconds then transition to doing Y in a different location for 3 seconds and so on. But I think it might be an LTX issue more than your amazing work.
Super looking forward to any future releases.
What LLM is completely unmitigated that I can use to add detail and randomization to prompts since LTX likes long ones? Someone had mentioned ?? heretic model or something like that, but I can't remember the name.
you are looking for the ComfyUI-LTXV-TimeGated-LoRA (check out github or custom manager); and I am using the qwenVL nodes or - even better in my eyes - the krea2 prompt generator: you can exchange the qwen3-vl-8b model on your drive with an abliterated one from mradermacher on huggingface. just dirty hack the folder. :o)
Heretic. I always use heretic. You can ignore the errors it throws.
Unfortunately, even though motion seems good, it seems like this checkpoint doesn't know how a penis looks, and often changes the shape of the head of the penis (or removes it) during blowjob. Hopefully OP can add this in training.
Incredible work. This is outstanding. I can't believe my results. Audio is more robotic than Eros, but I can custom that. Big kudos bro.
Would be helpful to cite the models you used for the merge. Something, anything. Given the collaborative nature of the whole enterprise. Just a little bone for the doggies, to make them feel appreciated.
Doesn't seem to be a merge (those tend to suck and be clunky). Looks like a full finetune. And a damn brilliant one at that.
@Goonasaurus_Rex If it is I stand corrected. It's just that there's zero info posted about it, which usually indicates that it's just a repost. Doesn't have to be the full 80 paragraph rainbow panoply... but nothing? That makes the spermaceti sense tingle.
@Ponder_Stibbons Usually I'd agree, but there aren't many good tunes of LTX yet and so far this one appears to be significantly better than the alternatives. Its ability to maintain likeness alone is vastly superior to the sulphur/eros twins at the moment. The darn thing just knows. Even with a basic-ass prompt for a pic like "the woman is posing," it will somehow get a feeling for the woman, style, scene etc and often enough it feels pro-level and appropriate in her behavior. I didn't even prompt it specifically for aheago and it did it on its own for a pic of a cosplay model. Smart little checkpoint.