The breasts in LTX-2 look very poor, so I trained this LoRA to address it. It was trained using a simple set of 30 images. While it's not perfect, it is a significant improvement over the native model and is definitely usable. The training did not include faces, so there are no issues with 'face pollution' (facial distortion).
Trigger Word: BestBreasts
Strength: 1.0 (You can use a lower weight during the first sampling stage, but I recommend at least 1.0 for the second stage/high-res fix.)
Tested using the ltx-2-19b-dev model. I haven't tested it on distilled or quantized versions, so I'm not sure how it performs there.
The next phase of my project is to attempt training for the pussy.
It seems many people are unaware that LTX-2 provides an official trainer. The installation is straightforward, and the documentation is quite comprehensive. I highly recommend checking out the link below—let's work together to grow the LTX ecosystem!
https://github.com/Lightricks/LTX-2/tree/main/packages/ltx-trainer
Description
FAQ
Comments (35)
very cool to see the first LTX-2 Loras! which software did you use to train this? edit: works with the distilled model
LTX official provides a trainer and a comprehensive user manual, which makes training a LoRA quite straightforward. You will be surprised by how fast LTX-2 is—not just in inference, but in training as well. On an H100 machine, training 6,000 steps takes only about an hour to complete.
hm I just have RTX 5090 ... hm maybe I need to wait until Musubi tuner implemented LTX-2 lora training.
@WhatTheGuy The official trainer offers gradient checkpointing and it's even enabled by default in the sample configs. In my experience this option turns "only on H100" to "anything that can run the model". I also have a 5090 and I definitely want to try this trainer myself. The docs are surprisingly detailed and describe various training scenarios. The effort put to this model release is HUGE!
Great job man! I'm looking forward to pussy lora.
So are the default breasts basically just... ugh... male bodybuilders?
They are like poorly done plastic surgery breasts. These weird plastic bulbous things. And maybe like they took some testosterone supplements and hit the gym.
You are the hero we don't deserve.
The default breasts are hilarious, truly.
They're male.
They look like character's eyes from South Park
is this for model stage 1 or model stage 2? are the lora nodes similar to wans high and low noise?
LTX-2 is indeed very similar to the high and low noise mechanism in the Wan2.2 model, but it isn't split into two separate models. The first stage works like Wan2.2’s high-noise phase, establishing the movement and general appearance, while the second stage of denoising focuses on finalizing the details.
As a result, you need to apply the LoRA across both stages. You can use a lower strength for the first stage, but it's best not to go below 0.5; otherwise, the model might fail to generate exposed breasts and may keep the character in bras or undergarments instead.
@adf992000 thanks. I did both anyways as it made sense. Thanks for the tip on strength hopefully itll lower generation time. If it is indeed the same as wan then we should start stating in the description.
So just to clarify, its a single model that acts as both high and low noise but still need to have its own lora node for model 1/2?
And so the depravity begins!
Playing the saint on Civitai? Really? 🤣
@adf992000 there's nothing saintly on my hard drive 😆
Did did those 30 image include a diverse set of ethnicities and breast sizes? 😅
Unfortunately, I just used a set of lower-quality images I found online and picked a few based on my personal taste for training. That’s why the breast size is pretty much the same across the board—it’s hard to make them larger or smaller. I'm currently using zimage to supplement the dataset. Haha, I wonder how 'AI training on AI' will turn out!
@adf992000 what framing/res do the images need? I could probably find some real images that match
Great work! Please make NSFW loras for LTX2 (penis, pussy, blowjob, cowgirl). I'm very curious and excited to see this kind of lora. The realism of LTX2 is incredible, I think we'll reach a new level of realism for NSFW content with NSFW loras for LTX2.
nice~
Dis you use the Abliterated text encoder to train? I want to try my penis dataset for it on runpod. anyone set up the trainer there already? I want tutorials before spending money training
As far as I understand, nobody should be using abliterated models with existing DiTs. It will more than likely only make things worse. The text encoder is used for precisely that, encoding the meaning (embedding) of the prompt.
Abliteration is lobotomization (only forgetting, not learning anything new), in an attempt to reduce the likelihood of token sequences which lead to refusals. That's not relevant here. We're essentially showing the LLM the prompt, taking an MRI of its brain, and seeing what lights up in order to create a "semantic map". The DiT has learned (or will further learn) how these maps relate to pixel distributions. The coherence and alignment of the existing mappings could be affected if you use a lobotimized text encoder.
It's only during inference you should worry about it, because some LTX-2 workflows use the same text encoder as a generative model for prompt alignment. That should be disabled or some other uncensored/abliterated model should be used specifically for that part (and the system prompt should be modified to remove any censorship instructions).
No need to use the Abliterated text encoder; the standard one doesn't censor the text anyway.
@xto @g1263495582 I actually used the Abliterated Text Encoder in my previous runs. I wasn't sure if it was making a big difference, but after seeing your feedback, I switched back to the standard Text Encoder for this version.
I just checked the 3k step epoch, and the character is still a bit 'stubborn' about open-leg poses, which feels different from the Abliterated version. It might just be undertrained at this stage—I’ll have a better idea once it hits 6k steps.
@adf992000 How is it going now? Is that encoder actually effective?And does it still consistently flag the 'legs-spread-apart' posture?I'd like to understand its NSFW detection capabilities in detail for my practice. I would greatly appreciate a response.
When you get around to training the sex LoRAs, make sure not to train it on "doggy style." It should give some pretty entertaining results when people use that as a prompt.
k....
im getting a lot of weight errors in comfy with the fp8-dev model, and the ablirated gemma
I knew this was already baking in the oven. Thank you!
yeah it definately fix most of the plastic balloon effect that base model had
it does not discrimnate both women and men of LTX 2 have the same boobs now. :D
Guide from the LTX team on how to train your own lora: https://www.youtube.com/watch?v=sL-T6dsO0v4
Could you do one or on your next version have the aiblity to show huge fake looking breasts. It seems every time I try to use I2V the breast size changes if they are covered for even a moment.