This is a minor finetune-level adjustment of LTX 2.3 using very undertrained and low influence nsfw training data from the Sulphur/LTXXX finetune project (on-going).

I didn't train it but have permission to share it.

Original Lora.

All preview start images are generated fictional characters. This model cannot undress or 'nudify.'

It's not strong enough for the fully nsfw T2V use without loras. For T2V try the much higher strength LTX2.0 version which is a full strength merge (can cause issues).

Workflow:

Try this workflow, it is being maintained and I can't update mine.

My original experimental sampling workflows:

My experimental i2v 10 Eros Workflow.

Raw WF link.

That WF was used for all previews, all red nodes are change points to modify the output. I never dialed in the audio normalization and CFG settings, so be careful about audio levels. If they are consistently problematic swap the SDE sampler to er_sde or redo the audio normalization. It's very experimental sampling that does good adherence and consistency, and is close to being as good as a 13 step SDE/Euler split sampling can be. I've worked on it for weeks using the finetune and model, then figured out how to bring the most out of it for LTX2.3. Idk if using the default model with it would be good, and be careful with the CFG level in the first high noise steps, that's where the audio issues come from but that adds a high level of control to the output.

I also never even tested this merge with the triple sampling methods or the normal method. It may only work in my WF as good as previews indicate. I only spent half a day on them pretty much made all of the previews in one sitting, so if you have good prompt, lora stack, and images lined up they will work well. Everything about this is very early and experimental.

It's still LTX2, and is not strong enough to be used on it's own. It's made to push out non-nsfw influence from the base model so that nsfw 2.0/2.3 loras can work better and 2.3's improvements can maybe shine through.

It's still extremely prompt dependent--use LLM enhancement. It understands pretty much all contemporary adult themes. If it doesn't work right away; first pass resolution is very important and can make a large difference. If the video isn't animating much it will probably be that. Adjust the video size node megapixels node and aim for 480 vertical, 512 square, or ~700-ish for horizontal. Adjust the megapixel node in the WF, most normal size ranges are 1.0-1.3.

The other motion constraint is the usual pre_process and i2v_conditioning node. That song and dance should be familiar to anyone who's used LTX2, you need to mess with those a lot.

The Sulphur finetune that powers this is ongoing, covers pretty much all adult-nsfw concepts. That project is swapping to a more expanded dataset and targeting a real LTX 2.3 train that will be much better than this. This is barely representative of it's final state, it's maybe 15-20% of it's training so far since the dataset was expanded. I'm only involved with it as a tester/implementer and have permission to share this.

I'm also leaving the country for a month and won't be doing any gens, trains, or merges, and can't fix this if it sucks, so I hope it doesn't and works for everyone like it does for me. I'll be back in April to revisit this and adjust the mix and WF then.

This model has a few do's and don'ts:

Do:

Use loras.

Use prompt enhancement.

Try video-to-video and extend or add audio to Wan 2.2 outputs, especially nsfw ones.

Try weird concepts, you'll be surprised what it can animate.

Test nsfw loras that you trained/are training with it.

Adjust all the settings on nodes that are red to finetune the workflow.

Try it out in different workflows.

Use frame conditioning. Using klein/qwen edits can be powerful for complex concepts like penetration, transitions, undressing, cumshots, position swaps, etc.

Don't:

Use it on real images or illegal content. It will not undress. That example uses an off-site expose lora.

Train on it. The final trained Sulphur version may be trainable but not this half accuracy mix.

Expect it to be perfect.