WAN 2.2
This one is pretty damn good if I do say so myself. Swapped out some of the other penis LoRAs for WAN 2.2 in my tests and this v2 blows them all away. Your results may vary, so look at the showcase prompts for help. This can create videos or images of male anatomy from a prompt, in realistic detail. It can also preserve the genitals if they are in your start frame for image to video workflows. Detailed anatomy prompting helps, see showcase for v1 and v2 for example prompts, or use local LLM Qwen VL or Grok for help with dick anatomy terms.
And it knows cum this time. No other LoRAs were used for showcase videos except for Lightx2v. For the cumshot, describe how much and exact details of where it goes. Beware saying 'massive cumshot' or you'll get a firehose like velocity and volume shooting out (see showcase).
You need both the high and low models in your workflows for video creation. The high noise model has its own version tab instead of the Required Components panel I orginally used. That would prohibit it from being used to generate videos with high+low loras on this site. The high noise part gets the motion, and low noise gets the details.
You can use the low noise with WAN 2.1 or its variants such as the SCAIL-2 or one of my favorites, Bindweave. You can create very high resolution images also the t2v model or with Magic WAN that are quite good (see showcase).
See showcase for prompts and workflows should be embedded in most mp4 files. I try to put the technique in them, t2v, or i2v. Some use RIFE or FILM interpolation for the frame rate and TensorRT or Nvidia RTX image scaling as well to get beyond 480p/720p. If there are some you like and the workflow is not available, let me know.
Check out https://coachbate.ai where you can see my latest work and follow me for free.
v2
Built on v1, adding 60 videos and the latest image dataset with the goal to fix the backwards or upside-down anatomy. It still can produce backwards glans, but before it would do it 90% of the time when viewing the underside of the penis, and now completely opposite, failing about 10% of my tests. Prompts can help if you get that. use 'detailed defined glans'. I guess it's hard to get rid of all the backwards dick heads in this world. Trained to a total of 15k steps, 20 hours of additional GPU time for version 2, stopping and adjusting along the way. Loads of work
v1
Trained on the same dataset as versions 1 of my Z-Image Turbo and LTX-2 LoRA, but for WAN 2.2 trained with the T2V model, which can be used to generate penises from your detailed prompt, or keep them in your scene if using i2v workflow.
Description
Trained with my latest dataset of ~60 videos and 200 images. The High noise LoRA is a separate version but also available on this page under Required Components section.
FAQ
Comments (10)
thanks for version 2 😍🔥
my please. please use the Add Post below and share one of your creations
Woah! Using your both of your low loras it does solve the problem of reversed peepee when visible-no visible state!! 👏🏻👏🏻👏🏻
The second version should be strength lower and the 000009750_low 15 points above and below the v2 and that makes the trick!! (i2v)
you're using the high noise lora also, right? Pretty sure that's where it gets the shape and low noise is for details mainly. for i2v with a penis in it, it shouldn't change the orientation of the penis even 1. it was the t2v or txt2img that had issues with v1 and backwards. If you include that you want the "Frenular delta facing camera" or away from camera, that is a nice new trick I learned
@coachbate No, just the low loras, and perhaps this is helping: "((Anatomically instantly correct depiction of the man's penis in ventral view facing the camera in a very detailed high quality frenulum and urethral slit that goes through and marks a division of the left and the right of the shaft of the peni's head of the man's penis and pale-pink color across all the frames.))"
And I am testing with https://civarchive.com/models/2468330?modelVersionId=2775271
And even the size is finally normal.
@TieFighterPilot it was made with the intention that there will the 2 stages: one high noise and one low, each with their corresponding lora (like all the other wan 2.2 workflows use). It voids the warranty so to speak otherwise if doing videos. You need them both. That model you linked to even says they want 3 high steps and 5 low noise steps after that. The high noise lora is under the recommended files section
@coachbate Ah, yes, that lora at the url already shows a 🌶 but not outside, I wanted the end of the clip too, so yours was the perfect combination because it gives the part that it left and with tricks with noise you can achieve it, I tried other ones but created monsters 🌶 or reversed but yours interprets good that noise that I was trying to manipulate.
@TieFighterPilot please share by clicking add post below. i'm curious
@coachbate I posted a short test lcm for this demo I had to go even higher in strenght, the v2-low at 0.46 and the 000009750 at 1.65 and the ModelSamplingSD3 at very low shift for micro detailing 3.00, I wish I could use a higher shift to focus more on woman's face consistency, if I add at any value the high noise then the 🌶️ head gets reversed, I wish I know how to train a lora I hope the math of this experiment gives you some clue or something 🤷🏻♂️ and thanks a lot for your work!
@TieFighterPilot if you aren't using the high and low noise, I can't really offer any advice. The v2 was based on v1 which was training only on images, which could be why you need a low shift for it. However, v2 was also finetuned for the details. Maybe I'll finetune some more. You shouldn't have to use two penis loras, esp since mine have different triggers. All my current models use p3n15 and before used bwc