This is an approach to get more realistic cum out of our beloved diffusion AI as most models were a let down in that regard. Since the training included various "body shots" the results are not limited to the face alone. It's up to your prompting how much white stuff is being applied... Results may vary - some models generate far superior results than others. Interestingly enough, HassanBlend (1.5+) was giving me a real headache until I tried others, which worked very well in a convincing way. Also check out my other <a target="_blank" rel="ugc" href="https://civarchive.com/user/redhome">embeddings here</a>.Version 3.5 is not based on new images, but was slightly refined (training settings, image descriptors). This version will be the last for the forseeable future.I tested this mostly with Dreamshaper 3.3 (title image), URPM 1.2 &amp; ProtogenX53. See the images for detailed prompts. VAE used was "vae-ft-mse-840000-ema-pruned".Again: it is still most effective with inpainting. (I suggest using URPM 1.2 for the best results.)Inpainting: I used a CFG scale 6 - 9, plus denoising strength at about 0.25 - 0.5. Changing "Only masked padding, pixels" to values &gt;24 and &lt;60 also helps to get a more realistic "flow of liquids" in some cases. The usage of "Only masked" in "inpaint area" is strongly suggested!Pro advice: do not use "Restore faces", if you want cum on faces, as well.Just use the embedding name as a trigger word after putting this into your webui "embeddings" webui folder. You probably know the drill.

realcumAI (Likeness embedding)

I completely rebased the source (amateur) images. By throwing out most of the more blurry ones, the results get a lot more crisp and work with most 1.5 based models more or less without flaws. img2img/txt2img was a bit of gambling with version 2.Dreamshaper (3.3) gives awesome, sometimes really artsy cumshots, too.If the resulting cum is too white, try moving the trigger word more to the back of your prompt. This will lower the attention and results in fewer "splats", but more convincing ones.Dataset images: 84 Batch size: 14 Gradient steps: 6 Learning rate: 0.05:5, 0.01:35, 0.005:260, 0.0005 Steps: 300 Latent sampling method: deterministic

Files with SHA256: 3215c06efbcfb8dab33e08763e0562804cc75fac414853b46cde67cf783e8298