This is an improvement on my original Super Cockskin Lora - https://civitai.com/models/1106422

This model is intended for creating images of ???? intact men.

This model offers many improvements over the original model. I'm still committed to making improvements.

Settings:
Sampler: I typically use Euler A + normal
Steps: 30 steps has been best for me. Sometimes going lower will lose the concept
Image size: 1024x1024 or a similar equivalent size with other ratios
Strength: 1+ (it can go a little higher without overdoing it). Sometimes going up a little like 1.1 shows the concept better.

Tips: Use both your clip-l and t5xxl. You can pull up one of my posted image and click on nodes to paste it in comfyUI for an example of what nodes I'm using and how I'm prompting my clip-l. From what I know about flux you can prompt with tags in clip-l and with natural language in t5.

Trained concepts. Please note that not all concepts are 100% understood. Prompting will have a great affect on the success:

1. Nude men - you don't need to prompt ?????. ???? man by default should have a penis.

2. Erect men - ????????, ?????, ????? ????? were used. ????? ????? is the most understood, but doesn't prompt well. The ????????? show up more with other actions.

3. Specific ??????? - ????????, ?????, shaft, ???????, urethra - try "color coded ??????? photograph", "blue ????????" or something like this to kind of test what it knows. It only has a weak understanding of most of these.

4. Oral ??? - the trigger words ????????, ???????, ???????? were used, but ????? in mouth is the most understood. Tongue in ????????, biting ???????? were used but it doesn't usually have enough detail to tell. You have to be SPECIFIC ABOUT POSITIONING. Just using the trigger words isn't going to get consistent results. A standing and a kneeling man gave me the most consistent results.

5. Anal ??? - this was trained but I haven't had any success yet

6. Masturbating - ???????????? ????? ?????, ????? in hand were used. "A ???? male ???????????? his ????? ?????. ????? in hand." - Rubbing ????? was not used but seems also work

7. Masturbating with fleshlight

8. Anal ????????? - doesn't really understand ???? or ????????? well. You might get lucky though

9. Self Suck - Autofellatio, selfsuck, ????? in mouth were used but the action doesn't look natural

10. Cum - ???????, ???????, ?????, ??????????? ???, - these were used but doesn't look natural and often gets skipped.

11. Pee - ????, ???, urinating were used - doesn't look natural and often gets skipped or has the wrong color

12. docking - this is a tricky concept, try docking penises

13. Inserting in ???????? - finger in ????????, touching ????????, ????? in hand - these are phrases it was trained on, but doesn't work consistently. "Photo of a ???? man with his index finger touching the tip of his ????????", is more consistent and will usually push it in a little.

14. Holding ???????? - intended to be like tugging or pulling but misses a lot. The other terms got confused too often so this one seemed like it would work best. "Stretchy ????????" wasn't used in the training, but seems to help flux understand.

15. Hairy/very hairy. Also try hairless.

16. Anime & comic art - During training the anime images got worse and worse as the photo style bled into other concepts. I added anime and comic art regularization images to try to bring the concept back in. I personally think its a bit better than the original flux's anime style.

This model is done training. It got a little overbaked towards the end. I'm still working on making improvements.

For anyone looking to try to recreate a similar model, see my tips:

1. This model was done with 4 network rank, 4 network alpha. Going too high seems to make it learn faster but less flexible, IMO, and overbakes quickly to look like crappy photos

2. Proper captioning is key to learn the concepts. The main rule of thumb is if you want something to be default and baked into a concept, don't caption it. For example if you want all the ???? men to have flaccid ???????, then on your basic images of just a ????? guy standing flaccid, don't caption the ????? at all. Only caption the ????? in non default images like ????? penis.

3. This photo was trained off of mostly old or very low quality photos. If your image sucks, make sure you to caption that its blurry or old or low quality photo. Then when you actually prompt your image you don't use "blurry" and it will be more likely to have everything in focus. That said, blurry images can still bleed into concepts. Also important to caption if it is close up or not. If you don't, you will be more likely to get close up images by default.

4. You can kind of color code ??????? to help add definition. In gimp you can add a layer to the image with transparency, color something blue on that layer, then merge it down. Export it as like png to your img folder (i had a separate concept folder for color coded ???????). It kind of bakes out a little, but it helps string things together. Keep the original with its regular captions still.
   
   For example - "color coded ??????? photograph of a ???? man. The ????? is blue." OR "color coded ??????? photograph. close up image of an ????? ?????. The ???????? is red. The urethra is white. The ??????? is green." Or "the ???? also known as ????????? is blue" (not quite sure the best way to do AKAs in flux).
   
   It does seem to bake out overtime but retains the general location of everything. Otherwise when I prompted ???? (after over a dozen epochs) it was still pulling up images of puppies. But one or two epochs with color coded anuses and it was mostly learned.
   
   I read that there was some kind of heat maps or something in these AI models and that at some point you may be able to have more control, but for the time being this is kind of a hack that may help. It is similar to masking, but doesn't block out sections of the image.
   
   When you go to prompt, it will not show the colors on the body parts unless you specifically prompt "color coded ??????? photograph." and isn't going to bleed into the default generated images (unless maybe you have too large of a collection of these color coded ??????? photographs.)
   

5. This model was trained in batches. I started by using a huge 650 dataset of images for the first few epochs. These contained all concepts. During training I would save the state. Every so often I would stop and cull images I no longer needed and continue the training.

6. Some images kind of held this back. For example, the aesthetic I was going for was to have the opening of the ???????? visible, but many of the images I had were taken at an angle that hid the opening. So now the model will often still hide the opening at the wrong angles like straight on.

7. I did not use regularization images in the regularization folder, but I did add regularization images in folders. I had an anime folder, comic art folder, etc with a handful of images to retain those styles. Otherwise they start to look more and more like photo style.

   All of that said, I am working on another model with a slightly smaller dataset, but this time instead of doing all of the concepts at once I'm going to work from basic to complex in stages (solo male ???? -> ????? -> bodily fluids and touching -> duos/ sex)
   
   DM me on here or on Discord if you need help with a similar model on Flux. I'm not a pro but I'll be happy to chat about it.