A LoRA which reliably depicts the insertion of fingers into a vagina for a woman in the classic position of lying on her back with the camera showing the action and her face in one shot<a target="_blank" rel="ugc" href="https://riotmodels.com/generation3dx"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/51748d1a-26e8-401a-98bb-7ac6659a7ccb/original=true/51748d1a-26e8-401a-98bb-7ac6659a7ccb.jpeg" /></a><h3 id="minimax-h3-version">Minimax H3 Version</h3>Trained on my RTX5090 using ai-toolkit. Steps 0-2000 trained on 32 close-ups to capture the micro detail of the fingers and the distension of the labia. Steps 2000-4000 trained on both close-ups and 30 medium shots to allow the model to understand the body positioning better and prevent zooming or cutting into close-ups during inference. Final 4000-6000 steps trained with the close up and medium shots plus 28 full body shots to enable full body inferences without H3 cutting to close-up.Usage Notes:<ul><li>For best results I highly recommend you use <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/10Eros-Max/tree/main?not-for-all-audiences=true">TenStrip's 10ErosMax for H3</a>. Using the LoRA on the H3 base model is fine but H3 doesn't understand labial and vaginal distension really at all and this LoRA can't do all that heavy lifting. 10Eros helps massively in this area, but without this LoRA, 10Eros struggles with this activity. The combo of 10Eros and this LoRA is the golden touch. My video showcase demonstrates the before/after effect of this LoRA on 10Eros.</li><li>Strength 0.8 has been used in all testing and showcase videos. Different strengths will likely produce different effects. Experiment.</li><li>Works with both FL2VA and Ref2VA models (10Eros is a hybrid checkpoint). As always with H3, tight and detailed prompt structure is key when dealing with two-person shots. See my showcase videos for examples.</li><li>The training set featured various different women all lying on their back with the camera either close up or at some distance away but within some range that was a "frontal view", so this LoRA still struggles if your angle is too much to the side, or the woman is in a different pose.</li></ul>A big THANK YOU to all of you who donated their buzz to the Early Access unlock in less than 6 hours! If you like my work and wish to support me, then please consider signing up for a paid sub to <a target="_blank" rel="ugc" href="https://riotmodels.com/generation3dx">my riotmodels page</a> where you get to see the full scenes of the Seven Sisters doing their thing.<h3 id="ltx-2.3-version">LTX 2.3 Version</h3>Many months ago I tried to use the same training set that I successfully used with Wan on LTXV2. It failed miserably. Now, with LTX2.3 and specifically the 10Eros merge, I thought I'd give it another go. After three different training runs, layering 3 different datasets of medium close-up (29 vids), extreme close-up (28 vids) and full body (28 vids), all manually captioned, I am now satisfied that I have a production quality LoRA that will faithfully produced the desired motion, insertion visuals and sounds correctly! This has been my toughest LoRA ever!To use this LoRA start with the 10Eros_v1.2 checkpoint and no other LoRAs, and use the sample prompts that come with my Showcase. The addition of other LoRAs and/or different checkpoints will not guarantee results. LTX, like many diffusion models, is incredibly sensitive about finger motion, so anything that might disrupt the latent space can break motion. Enjoy! And please share your generations!<h3 id="wan-2.2-i2v-hi-and-lo-noise-versions">Wan 2.2 i2v Hi and Lo Noise Versions</h3>My detailed training process for Wan motion LoRAs <a target="_blank" rel="ugc" href="https://youtu.be/HyICHBL26KU">is covered here in my new YouTube tutorial</a>.<a target="_blank" rel="ugc" href="https://civarchive.com/articles/27580">I have also written a series of articles here on Civitai on the subject of Video Motion Training.</a>Notes for use:<ul><li>No particular activation keywords, just describe the motion that you want (Wan 2.2 seems very tolerant of natural language vs specific phrases)</li><li>Example core prompt: girl pushing fingers into pussy, two fingers, solo, targeting her vaginal entrance</li><li>These phrases were included in the training: - fast rhythmic thrusts - medium rhythmic thrusts - slow deep thrusts</li><li>Suggested extras: - wet fingers - slick fingers</li><li>Suggested facial expressions: - her expression is of intense pleasure with mouth open and brow frowning - she smiles while looking at you</li><li>Trained on solo girls but worked on a series of lesbian couples consistently!</li></ul>Has been trained using a tight set of 28 videos of 10 different girls lying on their back with the motion of their fingers going into their pussies. Trained locally on my RTX5090 using Wan2.2i2v-A14B Hi and Lo noise models and Musubi Tuner with 1e-04 LR for Hi and 5e-05 LR for Lo. for 2nd pass refinement, 256x256 resolution, network dim 32, 90 frames.The LoRA seems powerful enough to get the girl to finger her pussy even if her hand is away and off to the side of her body. And also powerful enough to get her to do this in various different body positions. See the showcase for examples of this generalised behaviour. I put this down to the power of Wan 2.2, as the same training set with 2.1 was much poorer.T2V UPDATE:After burning 18 hours of time on my GPU (expensive with shitty UK electricity prices), I have been unable to get T2V to work properly, despite the same training set. Since I never use T2V personally, I'm not motivated to put in further time and cost into trying to solve this. However, if someone is really desperate for T2V then please reach out if you would like to contribute funding to it. Sorry guys, but I tried, and my time and resources are limited.

Fingering Pussy

First Minimax H3 version. Used best with 10ErosMax merge checkpoint at 0.8 Strength.

Files with SHA256: 8aeb24f44ec0648cc1cec654da0d0a48fe02eec5ac45e0d112796ec460a829f7