Big White Cock Frontal (WAN 2.2 T2V)

If you like my "work" please consider tipping my electricity bill: https://ko-fi.com/loralai82564

Upcoming loras: general purpose futanari/cowgirl and a really good realistic cumshot lora (both for Wan 2.2 I2V) which are coming along very nicely. In final testing stages.

This is a lora for frontal shots of a big white cock in flaccid and erect form, there is no trigger word. Please check the sample videos I posted for some ideas of prompting. It's quite flexible, the most important words for the penis are:

"erect standing penis"
"penis hanging down"
"flaccid penis"
"penis sticking straight out from her/his shorts/skirt/pants/jeans"

"she pulls down her shorts and her penis springs free" (if this causes too much penis bouncing, try "penis springs free and hangs down" etc, or just follow my example prompts and hunt for a good seed)

This lora works very well with character loras as there are no faces in the training data, so it should play nicely with them. It's not just for women/futanari, it should work absolutely fine for male characters too.

Recommended settings:

Without speedup lora (I use 20-30 steps usually):

20 steps

Beta scheduler

CFG 3.5-5.0 for high/low models

9 steps on high noise, 11 steps on low noise with shift 5

11 steps on high noise, 9 steps on low noise with shift 8

Lightx2v works pretty well with the low noise model, I do NOT recommend using the lightning lora for the high noise model. The bathhouse sample and the library futanari sample were both using Light2xv on low noise model.

With speedup lora:

20 steps

Beta scheduler

CFG 3.5-5.0 for high noise, 1.0 for low noise model

9 steps on high noise, 11 steps on low noise with shift 5

11 steps on high noise, 9 steps on low noise with shift 8

When using the speedup lora you can probably use less steps than 20 but that worked quite well for me, the problem with the speedup lora is it can mess up with the shape of the penis, so beware.

Training information:

Rank 16, alpha 8, discrete flow shift 3.0 for high noise training, 1.0 for low noise training

Cosine with restarts, 2 cycles, 1e-4 learning rate,

20 images, 768x768, 5 clips at 200-240p each 3-5 seconds,

3700 steps for high noise, 4000 for low noise (overkill, I could have used less steps for sure)

Upscale information:

I use Davinci Resolve to interpolate to 30fps and use its AI Super Scale for a bit of added clarity, I prefer this to using upscale models and RIFE in Comfy.