Its not only lora and prompt quality. I have played with LTX2 a lot and the sampler makes a huge difference. Play around with euler, lcm and res2s. Res2s its the highest quality and takes 2x more time than lcm but it doesn't always means it will give the best quality. It all depends on the prompt, lora and if its I2V the starting image makes a huge difference on what you will get.

@ENSIN Yup, I've found RES4LYF stuff to basically be necessary if you want images/videos that adhere more closely with prompts. bong tangent, beta57, and that new FluxEulerPromptAdherence thingy works for all models not just Flux stuff. Res6s with Bong_Tangent has almost always produced the best quality for me in Illustrious, Noob, Flux, but I don't dare do 6s with video stuff or I'd never see a single video! KKKKKKKKKKKKKKKKKKK

ClownSamplersSharkSamplers are all fantastic. I sorta wish these would just become core components in the ComfyUI default suite.

c0ur4ge thank you for your work!
Looks like that all this new way co create a some sort of things, like those type of videos become more about some sort of second job since it required a lot of time spending on research and fixing things to work. It's not about to make generation it for fun or pleasure. but for getting new skills in this days. =D
I spent a lot of time to simple generate something in ok way... Not even close to good result. And i tried only 2.0, looks like is 2.3 want some more skills and time.
Can only image how it hard for creators to create lora or models especially a video.

@c0ur4ge can you just share the workflow you used in the preview images you made ? that would be a starting point for testing and experimenting

take it like it is, this is its creation, you don't even imagine times spent on those training, ok it's not free but meh compare to time spent (and money!!)

Hopefully someone out there with deep pockets is training the actual weights and not loras. Its the only way to get true nsfw outputs. 2.3 has so much potential. You can tell the dataset was carefully put together to avoid misuse. Its even lacking in sfw stuff, sora did my prompts exactly but ltx defaulted back to its training. we have no competition with closed sourced models.

@JaimyB dude, I'm not saying it's not necessary, I'm just saying how it could be better.

@c0ur4ge ok so will the lora eventually be trained for 2.3?

@c0ur4ge Keep up the good work. LTX training is a bit of a PITA and gets expensive fast. Training a titfuck lora for LTX 2.3 I think its done but none the less ive already spent like $30 on this one lora and its still not top tier lol. 95 vids works but if only automated captioning was actually reliable then maybe a big ass 600 vid dataset could be made without me loosing my sanity.

Hey, those lashes are nothing to blink an eye at. 😜

I will say the voice model and being able to prompt audio is just much better than WAN. I can't think of a way to accomplish the same thing during generation or post generation for WAN 2.2. The NSFW audio add thingy is neat but it's very limited and doesn't always generate proper audio no matter what you do with it. And that audio model hasn't been updated in a long time.

WAN 2.5 is dead on arrival since they made it closed source, and I've not much hjope for the WAN team to do anything else at this point.

However, like you said, for video generation there isn't anything better than WAN for local gen NSFW content. LTX2.3 could very well become the king but it'll take time of course.

@c0ur4ge Your models have made pretty much every single local model better in many ways and not just for NSFW but in general. I can't imagine using any WAN model without DR34M4LYAIOV2. It's just too good to not use. Thank you for the work you've put into all of this!!!

It's almost there. Wan still owns in a number of ways, but it seems like it would be a challenge for them to get their models to be small enough for another open release. We'll see if LTX-2 makes them want to compete in that space.

@c0ur4ge They want you to add support for a hard edit, so like "Girl's face, cut to some sex thing". Don't do this. It gets your LoRA killed on Civit if it involves sex. Someone just needs to make an edit LoRA like Cinematic Hard Cut. Go beg them for a LTX version, @dqwddqwdq.

@c0ur4ge It loves it SO much. :P

Those get b4nn3d from CV sometimes, just saying.

In many ways I find LTX-2.3 already better than Wan 2.2

It takes literally 5 minutes to generate a 20 sec clip in 60fps. And the sound is so good now with 2.3.

Wan 2.2 still has slightly better physics and less color desaturation with I2V but I believe both will be fixed with LTX in the next weeks.

The Sound + faster gen time makes Wan simply not worth it for me

@c0ur4ge if you need help I have an rtx 6000 pro that can be used as well

unlike wan, you can easily go 30s long with sound.

Much better, but very few loras.

the foundations are better but in nsfw wan feels more mature and trustable

Better is relative. It does certain things better. Sound and length are the obvious things. Wan is capable of much higher quality motion with fewer visual artifacts. It's just a better generation engine from a baseline perspective, and doesn't inherit the weird qualities of LTX. LTX-2.3 seems like you can make LoRA's with it finally, but there will be a huge gape between them for a while. LTX-2.3 is MUCH harder to prompt (DEMANDS verbosity) and way more likely to fail to follow a prompt. Wan is often very forgiving and most always does what you ask.

The main advantage of LTX-2.3 right now is that it's developing with open weights. Wan went closed, so unless they release a new open weights update, long term Wan will lose out, as LTX will close the gaps and are positioning themselves to be able to take that crown.

That is not right now though.

Wan 2.2 is better than LTX 2.0 but with 3.2, oh boy. The new 3.2 is miles ahead of wan and with time itll beat wan in every category. Of course until a new ai video gen drops that matches sora 2.

The physics though are not that good atm

LTX 2.3 is like 90% of the physics and complex motion adherence of Wan 2.2
Also Wan 2.2 still has the upper edge in prompt adherence.

LTX 2.3 is still my fav go to model because of 30s / 4K / 60fps + sound. Once you have great sound with your video and faster generation, you can't switch back to Wan 2.2

Also I think most of these issues in LTX 2.3 will be fixed considering the rapid jump it made from LTX 2.0

@Johnsmith1969 i can make 30 sec videos in Wan too. the consistency takes a bit of hit though, but Loras do their jobs for character consistency, so only Character loras keep consistency

@damiang surprised how big of a jump it was. 2.5 will be insane

@jd666 For sure, it only took us 2-3 months to go from LTX 2.0 > LTX 2.3. Which means they're keeping the release timeline windows short.

@jd666 Yesss but will it be free and allow local generating?….

It's probably the workflow. Make sure you have everything installed correctly and that you're using the right versions of things (only the 2.3 for the official files, don't mix and match). Try the official workflows if you are not already. They're decent.

@boobkake22 thanks. I figured I wasn't ready to try building a complete workflow for this on my own, so I tried the official template for wan2..2 14B I2V and replaced the high and low noise lora's with these. Is that not the way to go? the originals actually got me better results. very stiff but at least not that blurry.

@tinynimity231 there is a official template inside comfyUI. Also with links to all the correct files you need for that exact workflow are linked in the workflow ;)

@WhatTheGuy you mean including these DR34ML4Y Lora's? that would be great! I can't find it though, I browsed and used the search function. All I found was the wan2.2 template I mentioned in an earlier comment. if I insert the Lora's in that as hirh annd low noise lora's, I get the strange blurry effects and no realistic movements...

@tinynimity231 No. there is a templete button inside comfyUI. looks like this https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fnew-z-image-base-template-in-comfyui-an-hour-ago-v0-i3jxwocfftfg1.png%3Fwidth%3D3456%26format%3Dpng%26auto%3Dwebp%26s%3Db784978c4c3153ba0c2f278624132219cd7ac707 . Then you search for the LTX2.3 distilled. Then you have a correct workflow for LTX2.3 in general. And then you can add your loras there. Most important thing is to get you to the correct base worklow right now =) there are special settings and integrated upscaler.

https://civitai.com/models/2445871

thats the best i2v wf I've used yet.

@brendacaro thanks !!

im actually making a lora but waiting for AI-toolkit to support LTX-2.3 first

20 second VR video, better be a quick draw!

@FantasticFever looping exissts

yes we can, there are some workflows i think.

try sbs. its 3D but still good

@Burgert whats this you say? if we were able to generate vr usable videos, and stitch them together to make something a little more longform. ugh, that would be top!

@Burgert Is there a working native LTX 2 SBS Workflow now ?

VR would be amazing indeed. Side by side Stereo is great too yep. There are good 3D picture / video and 3D vr 180 picture / video players for Quest 3 like Pigasus and also some alternative 3D mediaplayer apps nowadays which also support Network access like samba / and common dlna mediaservers ( same features like pigasus) so there is no local upload needed to the headsets.

If you tag / name the files correct such apps detect the 3D content from scratch so you can just watch your content without a hazzle in alltime selecting "3D SBS / Half SBS or VR " which makes it very comfort and enjoyable.

Some elder 3D / VR apps also mentioned to support 3D / VR but you alltime had to
select the dimensions and picture per picture and video per video 3D / VR switching the settings when playing which is sure a mess.

Pigasus for example created their app for every Meta Headset and the network / file support is very good. They still have a strict privacy policy in their terms so nothing should be collected here.
Still my fav app for 3D picture and 3D VR video playback on Meta Headsets and i personal use it with the Samba support on windows 10 and 11.

Windows 10 and 11 still include the apps shrpubw.exe and compmgmt.msc
which makes it very easy to just select a network share folder and all is set up in a glimps.

@Burgert @Burgert just and FYI ive made loras already for 360 and 180 and created a git with nodes to edit them + a workflow.

@brendacaro https://github.com/yushan777/ComfyUI-Y7-SBS-2Dto3D i use this guys workflow to convert any video to sbs and i use deovr lol but now that i see @Ragamuffin20 has a 180 360 lora, i may be heading towards that! up my porn game a bit lol

No shit distilled is meant to be run at 1.0 cfg

@jd666 no shit i was running non distilled

@FantasticFever why didnt you just say DEV

"Running non distilled"

@FantasticFever well all I can say is, touche....and be more descriptive with yoir prompts. Also if youre running 2.0 prompt adherence is garbage so try 2.3 even though its a little more performance demanding , much better prompt adherence.

@jd666 thanks, noted, I am noticing cfg 4 does over saturate alot no matter what, but of course I was running Euler A cfg sampler so doy!

I didn't realize that; that's why LTX works so badly?

@ForeverNecessary737716 Its an LTX 2.2 model. It barely works on 2.3. I think the Wan reference was a typo. It "works" in I2V, but barely. I've been scarred by some of the T2V outputs lol

@ForeverNecessary737716 there's no open weights for 2.3 yet. So as soon as lightrix release it, we all gonna start training loras

@Sukatroll they've been out for days

I agree, there is still is zero good male models for LTX 2.0 or LTX 2.3 out there... Also the other general NSFW furry one has a stronger male model then this as well, it can do male blowjobs pretty good just with the base model set to lora strength 1.0

Why do I use this Lora myself in LTX2.3, but it hardly works when generating videos using reference images, or the effect is very poor? I added this Lora using the official workflow, and I only tried the missionary position. Can you share the workflow?

@1543674486668 try an unlocked text encoder.

@c0ur4ge I'm not exactly sure how bad it is at prompting without gemma_3_12B_it_heretic_fp8_e4m3fn.

I started with the standard one in fp4, at the time, i found it underwhelming. So I asked perplexity how to make things better. It pointed me there. Put it in. Used your lora. Suddenly im doing stuff ive never seen before. I'd say it's a combination of all these things. My results are not consistent. But the kind of things im creating here, it's way beyond what i'm seeing on preview videos people are creating on this page. That's for damn sure.

It's a LoRA, my dude

look at the info box at the top right of the page. :)

same for me.

i think it works, i just use wan2gp and make the integrated LLM to use the triggerword and rewrite the long example prompts to the base image, maybe you can do something similar with grok since it's uncensord

Use the LTX video plugin, there's a special lora loader thatll make it work

i second this...plenty of other loras that do have a effect...this one seems to have no effect in 2.3

0.5 strength works best but it's still bad.

Someday china will released a wan equivalent open source thats easy to train loras for.

@c0ur4ge well im pretty good with getting what I want prompt wise its just, you can tell each increment ex (0.2 vs 0.6 strength) its a massive difference where the model itslef gets stuck or confused even on 2.0, which this lora is built for.

I think LTX needs to release the training data, or we'll be stuck with guessing.

But of course its risky due to copyright

@c0ur4ge can you possibly share the exact workflows you used for the 2 preview videos ? i think this would tremendously help by setting a baseline on the setup + weights + the prompt

@c0ur4ge @beautiful673 yea the huggingface workflows are very clear and worked relatively painlessly. I just checked out the basic one and was able to generate tests. The voice sampling is mind blowing , especially when comapared to mmaudio, which appears primitive in comparison

@gackt2  correct. now LTX 2.3 just needs some checkpoints trained with improved physics. in particular cumsh()ts once this is done we are all screwed lol....

@c0ur4ge i did try those workflows but for some reason, not getting the same quality as your preview images. so i was wondering if you have changed some things in the workflow or prompted a certain way

Anything that's relatively simplistic is also going to create far less advanced generations. That means anything waxy/flux like that has any kind of simplicity is going to not only create cartoonish badly animated visuals, but its also going to create bad nuances, bad emotions, etc. In other words the quality of the original image in terms of actual realistic, is going to determine the entire sophistication of the animation and voice. Use real photos = exceptional. Use the very most realistic ai images = exceptional.

With the exception of certain lighting and theme situations, which also destroy outputs. Make them basically 900% worse than wan 2.2. I imagine some people on here are running into situations like that.

LTX 2.3 behaves best when it can see midtones and shadow detail, so giving it a brighter, more neutral exposure keeps structure, faces, and motion coherence.

In other words, you want to use a certain picture, up the gamma, or generate bright images, then bring down the gamma afterwards.

This lora was made specifically for ltx 2.0 latent space, creator said he's updating it for 2.3.

Think of it like pa2 vs ps3, sure ps2 games work on the ps3 but the performance will suffer as its not made for that system and is not native to its architecture. Once thid lora is remade for 2.3 increase latent it ahould follow prompts more closely without burning the image.

@jd666 yeah, very little is natively ground up made for 2.3 atm. But a lot of this is also 2.3 and its own quirks, which spill over into what people think this lora is doing (not always, but sometimes).

@c0ur4ge all i can say is im very excited!!

@c0ur4ge It's working pretty damn good so far. Could it ever be made to make oral/deepthroat sounds? Currently the women can talk and moan clearly with cocks in their mouths

@c0ur4ge how it coming along im eaherly waiting!! let me know if i can support in any way!

@c0ur4ge hey there! hows it coming along?

This does not help at all, and I'm sorry. But your post was a funny read.

This is the funniest comment I have read today

There seems to be a very deep life lesson hidden within these Wise Words.

I think what they were trying to say is that tears, like real tears, get interpreted as cum. Like during video generation they are not tears anymore and penises will morph into vaginas to if dildos are visible to enable vaginal dildo interactions. Sounds like LTX issue.

That comment was very poetic though i love it lol

Yea I find the output isn't as smooth\ believable as WAN . But its a great step forward

@gackt2 I got it working a bit better after some experimenting. Nothing like the sample photos from the creator here but it's okay. Wan is still better like you said but ltx seems to be moving faster. One thing I found helped was the load order of the loras. Having the distilled one that's required in front of this one. Messing around with the cfg seemed to do some things too.

https://huggingface.co/DreamFast/gemma-3-12b-it-heretic/tree/main/comfyui

and how do you solve the mat1 and mat2 error, if i try to use this text encoders, I always end up with this error.
I can't find a solution on Google.

@d3ratiz3r930 I have the setup if you want to cross check, my biggest issue was I was using mismatched LTX checkpoints etc. and I also didn't add the Distill Lora! so I can help if you need help!

@c0ur4ge you're work is exemplary! 

Sound comes from the model. Put it in the prompt.

I've tried putting everything wording wise that I can think of to add the male sounds and seem to only get female sounds. (I've even tried, for experiment purposes putting in the prompt no female voices, just to see what happens and all I seem to get is female voices 95% of the time. So, Im still hoping that in a future release it gives a way to prompt the LoRA to add male sounds into the video.

You’ll need to provide detailed descriptions of the background, clothing, characters and so on. As this is quite time-consuming, it’s a good idea to use tools like Gemma 3 Heretic to generate the prompts for you.

Try setting CFG to 1.5 (from 1.0)

@CapAndABull For both High and Low in the LTX 2.3 workflow get set to that?

@jorakoczy387 high and low??? That's wan not 2.3

@c0ur4ge oneiraglyka433's original question was in regards to LTX 2.3, CapAndABull says "set the CFG to 1.5" and in the LTX 2.3 workflow in COmfyUI, there is indeed both a "generate low resolution" and "generate high resolution" section including nodes called "CFG Guider" for both, which is already set to one.

@CapAndABull oneiraglyka433's original question was in regards to LTX 2.3, you said "set the CFG to 1.5" and in the LTX 2.3 workflow in COmfyUI, there is indeed both a "generate low resolution" and "generate high resolution" section including nodes called "CFG Guider" for both, which is already set to one.

So CFG 1.5 should solve it @c0ur4ge ?

@oneiraglyka433 I did one test, both will all settings the same between them, same seeds, etc...except the CFG. From the one test, I noticed a difference in hand gestures that were made during speech. Ill have to try again more and see what diff it makes 

did you figure anything out on this. I think that dr34aml4y is far superior keeping identity from drifting and speach/audio. But i cant for the life of me get it to move, any ideas?

distilled is compatible to me

@iVeo3 Not true. I can take any normal pics and make it work.

Camera fixed on a 45 years old single woman with long straight light brown hair and blue eyes removing her top clothes to reveal her very perky naked breasts and large aureole and erect nipples and start performing a bl0wj0b on a 6 inches erect penis in front of her with her right hand stroking the penis and quickly keep on bl0wj0b frantically. She pushes her head down on the penis and then up repeatedly, She watch the camera all along with sexy eyes. She moans her pleasure all along. keeping her eyes on the camera at all time.

In prompt: (her clothes disappear:2) and if that doesnt work use Posing Nude Lora and .8 high noise.. presto...

@eyensteinproductions Thank you, I will try that :)

@eyensteinproductions Sorry, but i searched for the lora now. Do you mean the Wan one? thats the only one i found. Can I use Wan loras for LTX 2.3?

Yea thats what I was saying earlier, this LTX 2.0 model was heavily trained on women only. Honestly LTX in general does not do men at all well like wan. I am noticing I have to use like 2 different loras just to get a mans penis to properly work in LTX its annoying af compared to wan video which has zero restrictions lmao

@menfq77653 Well I appreciate the effort - it's an unusually effective model otherwise. Keep pushing, we need to properly poison the well of innocence that LTX is trying to be. Have you noticed "vanilla" LTX shames you if you try to do NSFW work it? It's pretty funny.

Yeah, I'm noticing the same with LTX 2.3. No matter what the lora, it just seems to mix up the head and the peener and BJs turn into self devouring abominations. >_<

LTX2.3

ltx requires only one file. there is no high or low

0.7 for this and 0.7 for the ltx2.3_nsfw_furry lora. I have NEVER seen results this good in all my time in generative Ai. Just using the standard ComfyUI i2v with all it's recommended modell and files.

"the furry one"? Link or name?

@slpzlmgvhxdxzuodlc427 judging by another comment they meant: https://civitai.com/models/2310920/ltx-223-i2v-nsfw-furry-multi-purpose-sex-lora

That's the one.

You're using them both at the same time?

@FartingMcCracker Could you explain the setup more? You use this model for High and the other for Low or how is your setup exactly? Would be much appreciated!

I think you're referring to the LTX2.3 version?

@italianbeef @Eshinio @ToastyOat:

I am, both at 0.7
No "high" or "low" because...
I'm using the LTX 2.3 version.

@Eshinio I'm using ComfyUI, and the workflow i the standard "LTX-2.3: Image Audio to Video

Upload an image and an audio file to generate a high quality video with synchronized lip movements using LTX 2.3...." you can find in the templates. I just copied and pasted 2 more LORA nodes from the one that was behind the subgraph.

Start with following the link at the description:
"EXAMPLE WORKFLOWS HERE! Shout out to RuneXX!"
Take the JSON file -> drag it onto comfyui -> it will complain about not having a ton of nodes probably -> google how to get the nodes.

Start with LTX-2.3_-_I2V_T2V_Basic_GGUF.json

or the LTX-2.3_-_T2V_Basic.json

You can do it manually with 'git clone <urL>' to the github repo into the custom_nodes folder, or use the comfyui manager to install as many missing nodes as you can and then manually get the rest if needed.. this should help. There are better guides for custom nodes out there, look for that terminology.

Train an orgasm Lora?

@WeBeJavn  that's the thing, there are good trained loras for orgasms, i'm asking creators their experience with those. coz i hardly see it anywhere. or, exceptions prove the rule.

@c0ur4ge did you get offended? this question was probably not for you, if you have never tried to incorporate an orgasms in your generations!! but that raises another question, doesn't it?
we can do this all month but i don't enjoy making direct insults. but that also raises another question, doesn't it?

It has more to do with prompting maybe, than a lora

@c0ur4ge when the update arriving? Im so keen!!

side note anyway can help with with differnt ways or styles to make NSFW Videos?

This is a LoRa model, it's a supporting file for a main model and it can't work by itself. If you have a high vram gpu, you can download Pinokio, install Wan2gp in it, and download this LoRa and add it to Wan2gp.

ready? u sure?

It doesnt do it for me, i think usually there is something wrong when this happens. Either prompt or settings. BTW as weird as it sounds a lower strenght high noise lora works better for me.

what went wrong in your life to the point where you actually had to type these words? jfc.

@jstchksomething ahh.. come on.. where is the love? oh. yes no butthole no love..

yo the questions on this website are so wildly specific

@jstchksomething  not everyone is as pure as kim jong-un!

@c0ur4ge Thanks for your reply but i use wan 2.2 14b

@c0ur4ge Alright thanks it works now! With the "Wan2.2_I2V_PainterI2V_Workflow_Kenpechi_v2.4" workflow.

I got good results by raising the lora to 1.41

@arcendolf67 When I raise the high noise lora too much, it starts to change the face expression. I mean, not the facial features, the person remain the same, but the expression, and I don't like that. Have you managed that by any chance? Also, I assumed you were talking about the high lora on 1.41

I gen at around 720p but these are some of the things I came across that affected skin texture while genning at that resolution.

1) If you're using WAN mix models like "Smooth mix", don't, as it can smooth out skin textures. Stick to the original WAN models.

2) Low model lora stacking; try not to stack too many loras for the low model as it can also smooth out skin texture. I just stick to lightx2v lora (at 1.00 strength), Dr34ml4y i2v (at 1.00 strength), and lastly the general NSFW lora (at 0.40 strength because if I go higher the women can end up with extra "fleshy" bits).

3) Kijai released updated lightx2v low step loras about two weeks ago, in my testing the low model lora retains slightly better detail than the version 1022 that I was using before. Give that a try; https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LoRAs/Wan22_Lightx2v

4) Don't use LCM as a sampler; while it is fast, it can smooth things out too much. Use something like euler, euler_a or something like ddim (my preference).

One thing I can think of that I haven't tried (for video) is to add a film grain node after the vae decode. Film grain can visually help mask the lack of skin texture a bit.

Hope this helps you.

@marcociuffi05678 oh, then you're using wan2.2. That actually takes a lot of tweaking. especially depending on the starting image. I would set the image resize to some very small for testing and only adjust one thing at a time until you get the results you want.

@c0ur4ge How should I set this up

@ziwxt123420 with courage and wizardry

@c0ur4ge AWESOME!!! your a legend

I sent you a message btw please let me know if you can!!

@c0ur4ge Can't wait! Your LTX2.3 preview lora is the must have for NSFW motion. :)

@c0ur4ge let it cook as long as it needs to, no need to rush

@c0ur4ge eagerly waiting lol