Phr00t

/

LTX2-Rapid-Merges

like 287

NSFW LORAs for LTX2 are still pretty early and genitals are particularly weak. Using I2V can help.

Make sure you are using my workflow. Others very commonly have problems with shiny and oily problems.

Hi, just wanted to report several problems I noticed in the NSFW v6 version:

1. Female skin tends to become excessively shiny/oily very easily.
2. Both female and male private parts (genitals) appear quite exaggeratedly swollen.

I'm using the LTXV-DoEverything.json workflow. Could it possibly be something I'm doing wrong on my end?
Thank you for looking into it!

Female private parts (genitals), (nipples), with LTX-2 are indeed very swollen horrors lol. Most likely will be less horror as better loras comes out and are integrated.

NSFW version when you have characters wearing clothing it in most cases will result in nipples being very noticeable or even poking through the clothing. Also it has the strong habit of undressing your characters in a very weird way. I assume it is best to use the SFW version if your characters are meant to keep their clothing on and for nipples not to be visible.

Yeah, got it. Tried adding negative words to block stuff — did literally nothing lol.

Yeah, got it. Tried adding negative words to block stuff — did literally nothing lol.

Well, CFG of 1 ignores negative prompts, so yeah, that's expected.

You really need to use I2V or even "first to last frame" to get the most out of the NSFW merge. I'm actually trying to train a LORA myself as the current LORA landscape is rather lacking, making us have to put up with these shortcomings in the meantime.

Yeah, got it. Tried adding negative words to block stuff — did literally nothing lol.

Well, CFG of 1 ignores negative prompts, so yeah, that's expected.

You really need to use I2V or even "first to last frame" to get the most out of the NSFW merge. I'm actually trying to train a LORA myself as the current LORA landscape is rather lacking, making us have to put up with these shortcomings in the meantime.

Thank you for actually taking the time to train a better one yourself ！

been having a miserable time @phroot, trying to train nsfw loras for ltx2 using the ai-toolkit and have tried a variety of steps , epochs, batches, gradient accumulations up to 3000 to 5000 on an RTX 6000 pro blackwell card. MY config shows it's using the text encoder "gemma_3_12B_it_heretic_fp8_e4m3fn.safetensors" but there seems to be too many god dam flavors of abliterated gemma 3 that claims to do it all nsfw / uncensored wise but I think it's very far from the truth. have also been using a mixture of claude ai and ltx's own guidance , ostris also posted a 40 min video of him making a ltx2 lora of carl sagan but so far I've been making crap outputs from judjging the samples it produces , getting monster blobs.