š„ LTX2 NSFW ā by phr00t
Tested and FineTuned (in progress) By EKKIVOK
Please read ā¤ļø (thank you š):
LTX2 NSFW is an experimental NSFW-focused model based on LTX2, designed to improve adult content generation while maintaining solid prompt understanding, cleaner visuals, and better anatomical consistency.
This model is the result of a merge of multiple up-to-date NSFW LoRAs, carefully combined to enhance overall quality, diversity, and flexibility across different styles and scenarios.
Model is still at work, and newer version will be released ...
Branded to model 6.5 (cause the final model will be replaced with the final one).
āØ Key Features
ā Merge of several recent and updated NSFW LoRAs
ā Improved NSFW prompt understanding
ā More stable anatomy and proportions
ā Better pose and expression coherence
ā Works well with Forge Neo / ComfyUI workflows
ā Balanced between realism and artistic rendering
š§Ŗ Experimental Model
ā ļø This model is experimental.
Results may vary depending on settings, samplers, and workflows.
Further tuning and improvements may be done in future updates.
Use it primarily for testing, exploration, and creative experimentation.
Sampler -> Res_2s
Steps -> 8 Steps (distilled version)
š§Ŗ Testing & Finetune
Testing, evaluation, and finetuning (in progress) :
Thanks for the valuable feedback and testing support that helped improve the modelās behavior and output quality.
This version is based on V6 and is named ā6.5ā, as Iām currently working on a new version derived from it (phR00t release, credits below).
š Credits
Model creator: phr00t
HuggingFace: https://huggingface.co/phr00t
Testing & finetune (in progress, Workflow on the way) : EKKIVOK
š Adult content only (18+)
Please follow Civitai rules and local laws when using this model.
Description
š§Ŗ THIS IS AN EXPERIMENTAL VERSION ā ļø.
š„FINE TUNING STILL IN PROGRESS (6->6.5)š„
FAQ
Comments (60)
Thx for posting The source.but i'm a bit confused. Can you please eplain, why this is Version 6.5 even thought The source's latest model is Version 6? (Edit: Overread it in the discription, my Bad) Also do you have recommend es settings? Is IT a distilled Version? Cfg? Sampler? Steps. Thx in advance.
It is just a reupload of v6, not fine-tuned by ekkivok! ekkivok lied
@shoes22Ā Hmmm you really need to learn to read descriptions .....
@Adaptalab0r it's a distilled version (8 steps) so as a distilled version you need to use cfg 1. for samplers, scheduler, i highly suggesting you to use res_2s :) (i will add those info into the description) :)
@EKKIVOKĀ thx. res_2s oof. Hope IT doesn't kill The mood with Generation time ;-) will Post reaults
@Adaptalab0rĀ In fact..it will. it's a memory killer sampler.
But you can still use Euler / Euler A -> simple. It will do the job and much more quickly :D
But personally i use Res_2s bongo tangeant (don't remember how to write it) on all my WFs...
Not fine-tuned by ekkivok! Stolen, same hash as https://huggingface.co/Phr00t/LTX2-Rapid-Merges/blob/main/nsfw/ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
bruh
Dude....
yep, it's same. lol.
cd7b6dec437c4d7f9974aa9ebcdcff30
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
cd7b6dec437c4d7f9974aa9ebcdcff30
ltx2NSFWDISTILLEDFP8_ltx2NSFW8STEPSNSFW.safetensors
nobody click "show moar" in description
Actually, Proot left this project. If the uploader is going to continue the work, then we should support him. Proot release the merge code so that anyone could continue the work.
Those nipples look weird
yeah the model has serious limits (for the moment)
But i think it will be impossible to do like Wan or SDXL....
LTX is FLUX but for the videos. this is why it's pain in the a** to do something decent with it.
@EKKIVOKĀ The nipple issue is specific to the merge or loras. Default LTX handles nipples fine.
@kudon44Ā hmmmm yeah you are right but it's rough models it need to get more training i guess :)
which workflow were these videos made through?
my special workflows i made for it and release soon :)
Ok, it's time to dive in LTX and test this shit out. As i see, we need time to create as much loras as we have for Wan 2.2 to have similar quality with advantages of LTX, but i hope it will take less time because creators already have their datasets from Wan 2.2 Loras, right?
@Honeyphoria
Yeah, for sure, Iāve seen quite a few LoRAs on the site and some of them are honestly insane in terms of results.
The main issue though is that theyāre trained from a base model that isnāt very censored (like Wan for example, isnāt really restricted). LTX is a bit different, because itās built on a FLUX-style architecture, which you probably know as well as I do is the most problematic model when it comes to this kind of content.
Personally, Iāve never seen a model as restrictive as FLUX. Weāre really far from the SDXL era.
Just look at it this way: FLUX has been out for a while now, and no matter how many attempts were made to make it more āopenā, weāre still very, very far from what SDXL can do.
So yeah, in terms of flexibility and formats itās basically:
WAN ā SDXL Total freedom | FLUX ā Creator nightmare š
@EKKIVOKĀ Thanks for the detailed info and for sharing. It sounds kinda sad :( I had really high hopes for LTX 2. I'm tempted by the possibility of making 20-second shots with lip-sync and, as I understand, faster render times. But since you're working on fine-tuning LTX 2, that means there's still hope and it's not all bad? I'm creating NSFW animations with 3D characters, and I'm thinking maybe I can do something even more interesting by using built-in ControlNet with 20-second videos from real corn references.
@HoneyphoriaĀ Take a look at Wan2.5 it does the same thing as VEO (nice videos with dounds) and it will be released soon (i hope) for the moment it's only via API. but we know wan and it will be released as a public release.
@Honeyphoria
Actually, I just released the model that phr00t made, like I mentioned in my post.
Iām using that model on another PC to do some fine-tuning, but Iām not really expecting huge changes (though I might be wrong).
The main idea is to use it later for merges with other LoRAs that Iāll train using the distilled dev model.
Like Iāve been saying from the beginning, this is all experimental.
The version phr00t made is already a really good base, honestly, a very solid one.
Uploading it to Civitai (on top of HuggingFace) is also a way to let more people (maybe more skilled in this area) know that an LTX2 NSFW model is in progress and already released.
Hopefully, some of them can turn it into something everyoneās been waiting for.
Thatās basically the main reason why I uploaded it to Civitai.
@EKKIVOKĀ Yeah, I was really hyped for the 2.5 release too, but so much time has passed since it actually dropped that Iām not even sure anymore when (or if) weāll get it as open-weight š Probably the Chinese team started figuring out exactly how weāre using it lolā¦ just kidding, Iām sure theyāve known for ages hahaha.
Anyway, Iāll definitely test this LTX version and all your future releases. Honestly, some LoRAs are already putting out really nice results judging by a few videos Iāve seen on Civitai. If the community goes all-in, trains a critically huge amount of high-quality LoRAs, and holy grail fine-tuners like you merge them all into the base modelā¦ we might actually end up with something insanely sweet š¬
If I get any really good renders out of it, Iāll upload them right here under this model thread.
@EKKIVOKĀ hi, im reading a lot of your comments, saying that the nsfw lora's is giving very good results. i tried the phr00t checkpoint with almost any workflow i found, and i get very bad results.... any tips?
@vAnN47Ā i was talking about Wan 2.2. I haven't tried LTX 2 yet
@vAnN47Ā use my workflow it will be released soon, i tweaking it for the best result, honestly Phroot workflow is kinda bad cause it made that very quickly (it's a wan expert not ltx lol)
Just like Phr00t's original. I cannot use this in Wan2GP. I get:
Error
Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.
Can you give me more informations :) thanks i will try to help you on this one :)
I will release a special workflow for this model from scratch.
Maybe you will not get errors anymore :) @wfux
@EKKIVOKĀ Thanks for the reply. I don't use ComfyUI. I use Wan2GP. Whenever I start generating I get this error. I cannot say what the error means. So I cannot say if it's Wan2GP's or the model's fault. I can only say I can use most other safetensors.
https://github.com/deepbeepmeep/Wan2GP/issues/1439
@wfuxĀ I don't know that but you say Wan ? It's a LTX2 Model :D
@EKKIVOKĀ wan2gp is an interface made by deepbeepmeep that runs various video, image and tts models. ltx, ltx2, hunyuan, wan2.1, wan2.2, zit, flux, qwen, etc etc. It is highly customizable but also much easier to use out of the box than comfui. We make "finetunes" of non standard checkpoints to be able to run them in the wan2gp environment. What wfux is desccribing is that they have made a "finetune.json" to load up the ltx nsfw checkpoint you have shared, but it isnt working.
@TheFunkĀ thanks for the infos i'm running it right now with pinokio and this is really good ! but i prefer comfyui ^^ (more control) but that things is incredible !
For your problem @wfux i doubt that custom checkpoint can run with sommething like wan2gp cause it has it own protocol, encoders, text encoders, audio vae that was made to work with dev and istilled official LTX models. That one is a merge and a custom models that can run only with comfyui...
@TheFunkĀ Is it possible now using GGUF models in Wan2GP or it is still to complex for Wan2GP?
@EKKIVOK at least it is possible to run custom Wan checkpoints with Wan2GP, not sure about LTX, but you should create your own profile. There is information about that on their Github repo how to do that. But what is impossible or was impossible is to use GGUF models of Wan in Wan2GP, thats why i dropped it. It's better to figure out Comfy then rely on Wan2GP developers and it's limitations. This is a message more to WFUX than you EKKIVOK i guessš
@HoneyphoriaĀ @TheFunkĀ
Yes, I knew that some external apps had issues with custom models like this one. However, I wasnāt familiar with Wan2GP, as I usually use only ComfyUI and Forge Neo. I have to admit I got a bit confused at first because I didnāt fully understand it, and I had to read it twice before it clicked (i'm french)š
.
I installed and tested it, and itās great, but itās not as good as ComfyUI in terms of performance and speed, even with my 5090.
So yes, itās fast and works well, but itās far from matching the open-source flexibility that ComfyUI offers. There are serious limitations in Wan2GP that ComfyUI will never have :)
@EKKIVOKĀ Yes i can get around 20-25% fater generations in comfyui, but i find with wan2gp i can really max out my RAM and manage longer gens or higher res gens than i can with comfy.
I am running a custom finetune of LTX2 using the heretic text encoder and the old VAE, but the actual checkpoint is the distilled model supplied via wan2gp.
Yes wan2gp does handle custom wan2.1 and wan2.2 checkpoints.
It seems happy to run whatever text encoder and VAE it is directed to in its finetune.json file.
it is a shame that the NSFW here won't run in it. Nevermind, wan2.1 and 2.2 are well tested and developed now, hopefully LTX2 will get there too. The longer gens and audio combo is a lot of fun...
@HoneyphoriaĀ gguf was added last week in wan2gp. Works now
@kronos1959777Ā Nice i will check it out and i will upload it on civitai :)
@EKKIVOKĀ Were you able to do anything in the end?
Hello, does this work with wan2gp v10.98 ?
try to change the finetune file, wan2gp pulls this as ltx 13B
Here is what work for me if the file don't exist in Wan2GP\finetunes create one and name it LTX2_-_NSFW_DISTILLED_FP8_EXPERIMENTAL.json
Copy this into the file
{
"model": {
"name": "LTX2 - NSFW DISTILLED FP8 (EXPERIMENTAL)",
"architecture": "ltx2_19B",
"description": "ltx2-nsfw-by-phr00t-tested",
"URLs": [
"ltx2NSFWDISTILLEDFP8_ltx2NSFW8STEPSNSFW.safetensors"
],
"auto_quantize": true,
"preload_URLs": "ltx2_19B",
"ltx2_pipeline": "distilled"
},
"prompt": " ",
"num_inference_steps": 8,
"video_length": 241
}
@clayton4115Ā yes
Honestly everyone, I'm fine with this. I encourage people to continue my work, because I'm winding down making new stuff myself. My model was based on the hard work from other LORA creators and the entire LTX team whom deserve more credit than I ever should get. Stay safe out there and please, goon responsibly.
I was gonna say that, i read that you were winding down updating this, so, I'm glad someone is continuing your work... thanks for all your work, by the way, I'm really enjoying your AIO version of qie2511
@Phr00t_
Thanks my friend! And thank you for your work and for your attempt as difficult as it was to try to make this model something more open! Personally, Iāll keep pushing your model as much as I can, without huge expectations other than continuing to merge the LoRAs that will be released over time. You did an excellent job making this model more flexible for future merges. Your work is just as remarkable as the LTX teamās, and I sincerely mean that.
@artificialotakuĀ <3 <3 <3
@EKKIVOKĀ I recommend trying out this merging script which handles bringing lots of stuff together nicely if you haven't checked it out already:
https://huggingface.co/Phr00t/LTX2-Rapid-Merges/blob/main/MergingScript/fancy-apply.py
@Phr00t_Ā Omg thanks that's what i searched for ! Indeed i didn't look in that HF folder !
You're a god ! if you get some time in few weeks or month i will be glad to work with you :)
@Phr00t_ i made a script too but i wasn't able to get some good results your scipt indeed is pretty better than mine i think i will take some part of your script and adapt to mine, i really think LTX2 is a really good model (Wan2.5 coming, i hope you're ready for this lol, it will be much easier to work on that one)Ā
gonna miss you king, the qwen aio is a blessing
the included lora is "naked woman sitting on couch" lora.
not only that one.
Here, too, do all the faces look like the annoying face type of a flux girl with excessive makeup? Or is there variety?
there is variety but it's a LTX2 model. so it's basicaly base on FLUX. (see that as FLUX for videos as the main sfw model is)
Anyone ever got this working on SwarmUI? I get error "ComfyUI execution error: Audio VAE config is required for audio VAE" - tried leaving auto for VAE download but nothing happens (and no forced video/audio VAE's for LTX-V2 work either)
nope same issue. Wish there was more swarmUI support. Really no reason to use base comfy anymore
Will this be fixed?
checkpointed models should contain all the VAE's needed but this one doesn't and required Comfy workflows to make it work - they should really fix this and upload a new version here
Is there anyone creating a fine-tune or is everyone just waiting for 2.5 ?
Hello, does this work with wan2gp v10.98?
Details
Files
ltx2NSFWDISTILLEDFP8_ltx2NSFW8STEPSNSFW.safetensors
Mirrors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors
ltx-2-19b-phr00tmerge-nsfw-v6.safetensors