<h2 id="trigger-word:">Trigger word:</h2><p><strong>Doggystyle. They are doing it doggystyle.</strong></p><p></p><h2 id="you-need-to-setup-the-scene:">You need to setup the scene:</h2><p>(brackets text is optional)</p><p><strong>The man is positioned behind the woman</strong> [with his hands gripping her hips tightly.] <strong>The woman is bent over</strong> [with her arms resting on the hood of the police car.]</p><p></p><p>I recommend to use this lora with my penis ltx lora. And caption:</p><p>"The man's penis slides in and out of the woman's ass. "</p><p>Use 1 str on this lora, and 0.5-0.6 strength on the penis lora.</p><p>Note: Very unstable, but works half the time. Other times result in extra body parts.</p><p>Do not need use penis lora for i2v, t2v only.</p><p></p><h2 id="setup">Setup</h2><p>I made a tittydrop lora using a similar method and it came out perfect at only 1k steps. You can teach motion quite easily with LTX it seems. For this lora my goal was to have a simple phrase that will setup the motion for doggy style sex (from the side view). I quickly scraped around 30 clips, trimmed them to 2 seconds, converted them to 24fps, then trimmed/padded them to all be 49 frames exactly. I made sure to crop out faces and any watermarks. Then I used my captioning tool with grok to describe whats in the image but to ignore the hands and do not describe the action and just stick "They do it doggystyle." at the end of the caption. I went through and manually tagged the tattoos and piercings since grok is bad at that (still missed a few...) I spent around total 90-120 mins max setting up the dataset/sourcing clips.</p><h2 id="training">Training</h2><p>Trained on runpod with rtx 6000 ada. Float8, with 4bit text encoder. I used the same dataset two times, once with "train i2v" enabled, and the other with that not enabled. To train the lora on both t2v and i2v. Cache latents on for all datasets. Resolution 256x. Caption dropout rate 0.01. The best result was 3.3-4.2k range (very common I see for LTX loras). And after 4.2k its clearly over trained with every generation having body horror.</p><h2 id="results">Results</h2><p>I was not going to release this originally. it was more to learn about training and see what we can do with very limited video only dataset. As I rushed through the data preparing stage. But it worked well enough that I'll share. I think this needs to be redone completely to be actually stable in v2. But its better than no doggy style lora and does work.</p><p></p><p>I2V works better with 3300 steps version, but I like how 4200 version works with the penis lora on t2v. So I put both, I highly recommend just using 4.2K version. And if you want to see penis shaft use the penis lora 0.5 str with it. If you are having trouble, use my example prompt and change things around. That one seems to work well. I have not tested i2v much though.</p><p></p><p><strong>Known issues:</strong> because I didn't caption the hands (bad idea in the end), sometimes hands shift to 5 fingers or arms deform. Just reroll with a new seed or modify prompt. I say 50% of prompts works fine. Sometimes its just body horror. It tends to crop out the men/women, so prompt in the camera shot better or state "woman's face is in frame at all times" etc. It does not do initially insertion, I got it working with penislora but its not worth it. This dataset is purely mid-fucking only.</p><p></p>

Files with SHA256: 597eecf5a0a372f81d64374653026d48698d2800583cfcfef4c2ee5b0ee8b530