v0.3 was dual trained so the same file should be used for both high and low noise.<h3 id="this-is-best-used-as-a-nsfw-enhancer-for-slime-girls-rather-than-a-standalone-motion-lora.">This is best used as a NSFW enhancer for slime girls rather than a standalone motion LoRA.</h3>I got tired of waiting 10-15 minutes for a slime girl animation only to be disappointed by the transparency not getting animated right. In that regard this was originally meant to be similar to <a target="_blank" rel="ugc" href="https://civarchive.com/models/2322841">Translucent Penetration</a>, but there is also some builtin motion enhancement.Since I have a 4080 (16 GB VRAM) and I couldn't train a Wan 2.2 LoRA with a huge dataset, I decided to use 2 separate 16 video datasets (16 FPS, ~480p, and 21 frames from each).The trigger word is <code>slime_girls</code> but I find that a prompt like this works pretty well:<pre><code>slime_girls, aqua slime girl with a see-through gelatinous body, standing sex, rough vaginal. The man is excitedly thrusting his penis into the slime girl's pussy repeatedly, rhythmically fucking her hard. High detail transparency shows internal penetration through her body.</code></pre>For best results I would recommend using with <a target="_blank" rel="ugc" href="https://civarchive.com/models/2040641/smoothmix-animations-wan-22?modelVersionId=2376136">SmoothMix Animations XXX</a> (or something equivalent) and a checkpoint with good NSFW understanding. I usually use <a target="_blank" rel="ugc" href="https://civarchive.com/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors?modelVersionId=2761725">DaSiWa WAN 2.2 I2V 14B</a>. The starting image also needs to have good transparency or it might not get the effect right.Bonus tip - if you are using the lightspeed version of DaSiWa or another checkpoint that has the Lightx2v LoRAs merged, you can stack a <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/WanVideo_comfy/blob/main/LoRAs/Wan22_Lightx2v/Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_260412_rank_256_fp16.safetensors">high noise Lightx2v LoRA</a> on top of your high noise LoRAs to boost the motion. Start with a weight of ~0.5. Higher weights give stronger boosts with higher risk of quality degradation and/or nonsensical movements.

Slime Girls

Apparently I had the time steps reversed for high and low noise during training for v0.1. This should do a bit better with transparency. I haven't been able to get a better high noise version yet but v0.1 high noise is pretty decent when used right.The recommended strength is ~0.7 instead of ~0.3 now since it actually kinda works.The file size being 4x smaller is correct since I wasn't properly combining the weights before.

Files with SHA256: 8a5f8e35931b185c101f3facbb538646a09085325c49dcbd53b66e5789135c98