I had good success with already positioned initial frames and combined with "Self Forcing"; 3-5 steps usually enough when using with it.As usual I just picked some generated images randomly to animate from related models, kudos to creators.No trigger word used directly but tongue job is common in all. I recommend focusing on the action describing the subject with the concept related words. Some nsfw words included in the train set such as: cum, blowjob, tongue job etc.Here’s an example prompt I used in some of these generations—feel free to alter it heavily:<pre><code>A woman is performing a tongue job on a circumcised, erect penis. She is licking the head of the penis.</code></pre><h2 id="wan-2.2-experimental-update:">Wan 2.2 Experimental Update:</h2>I start exploring the new Wan version, the two model approach seems to be adding better movements and prompt adherence but causing some other problems such as rapid movements or little artifacts here and there. I have trained a lora for each version, a high noise lora for movements and low noise lora for finer details. You can also use low noise lora directly like wan 2.1 if you want.Note: If you have any concept requests or want to support my experiments, feel free to message me. Testing different settings and learning new training recipes for wan 2.2 seems to be expensive for my taste, I am open for runpod gift vouchers for that purpose... (Begging part) I use total steps of 6, 2 for high 4 for low part using self reinforcing lora (for 2.1) on 2.0 strength, rest is same as official comfy workflow. This is just early access and experiment version, use it at your own risk xDAs usual:Please use the model responsibly. No real person faces were used in the training.

Files with SHA256: 9286cde277ad82f1424e41f924d089003f305e81339afb25ab3513f0d02590b7