>>35256 Thanks. It gets the tapes right more often than not. Maybe 3 or 4 out of 5. The CFG and the checkpoint really drive it though. I dunno if these grids are too small after upload, but here's the full set for CopaxTimeless, Pony, Pixel, and Juggernaut. I have no idea how to prompt Pony, but it seemed to already understand diapers. The lora forced the white diaper blue tapes style when it worked.

>>35281 After looking through all that data I can see where you got some of your training data. Would you mind uploading the json file you use for kohoya so I can see how close your settings are to mine? Here is mine for anyone who wants to try it out. I typically try for 10 trainings of each image and a batch size of 10.

>>35288 Sure. We are pretty different in terms of settings. I gave up on Adafactor and Constant. Dadaption Cosine was what I ended up on. Prodigy was alright too. The 15 epochs wasn't really any different than the 10 from the model prior, but I added 2 more ass shots to correct an issue there and wanted to see if I could overtrain, but Dadaption doesn't really do that. I used 32 full body images after the ass shot addition. Took 12 of them and cropped them to close up diaper shots, and did a second copy of the close up with the background removed. 7, 1, & 2 repeats respectively. WD tagger. I tried natural language captions from CogVLM, but they didn't give as good results. As an added bonus, here are some more furries since I figured out Pony a little better today.

>>35291 I like the baby lion. :3

>>35288 How are the other LoRA types? I been just using the normal original vanilla LoRA. I haven't really bothered to look in to the other variants - as I think I still got a plenty to master just on the regular. There are some actually really interesting kinds available there for sure with really cool properties. Oh. And I recommend trying other base models (Which been trained instead of merged). I have solved many problems I come across by using a different model. Those LoRAs also work in all base models - only struggle with the merge extravaganzas which barely perform on general level to begin with. However if it doesn't train on the SDXL base model, it ain't training like that at all and you need to rethink. >>35291 Adafactor constant is not that good really. Really hard to adjust. Adafactor seems to only work with Adafactor scheduler. I personally use cosine for all and it works well. I use DadaptAdam with Cosine and it always works. Might not be the best, but the one I know will work every time. Most difficult part is setting the repeats to just right amounts. It seems to be around 7-10; 6 if the thing is obvious but below that it is just nothing. Above 10 it becomes inflexible. But if you struggle with bias to style, like it making only photographic, only anime, only painted like... etc. Throw in few random images of different styles. Doesn't need to relate to the subject itself. Like if I want to train something with only photographs, I put in a random cartoon character and it adds a lot of flexibility. And so forth. Also you can increase image quality (if your model is learning "low quality" attributes) but putting in few really good images of unrelated matter. Same thing with fixations. You can put in like partially exposed diaper, and you get extra flexibilty. But treat this like spice, just a small dash is enough.

>>35306 I can't say I notice a whole lot of difference training a locon. I had read some guides online that said they were better for lower end cards. I only have 8gbs of ram and have ran into memory constraints before. I may try retraining a model using a different model trained on sdxl just to see how it performs vs a standard sdxl model. Just trying a lora with dadaptation caused my training time to spike to over 48 hours. I'm guessing my 8gb 3070 just doesn't have the ram for the job. I really need to try adding some closeup crotch and butt shots to my training but in the past that skewed the results too much that prompting a full body shot was hard, but I only recently realized you can specify how many times it trains specific images and these last few posts reminded me I should try that.

>>35309 Gradient Checkingpointing was the biggest vram saver option I found.

>>35309 If you been doing the base 40 repeat for every training then yup... That's bit excessive. My 4060TI with 16gb does 6-9 batch size at around 15gb of Vram and FP8 training. I have read that if you do all the VRAM optimisations use Cudnn and FP8 you can do SDXL on 8gb no issues. And I believe it because lowest I can run with basic settings in like 9 gb of VRAM. Also keep in mind that you don't need HUGE pictures, the training seems to be better the more sizes you got.

Not a lot of images for an imageboard

Pony is basically cheating at diaper content

The lora definitely doesn't know what I guy in a diaper looks like XD

>>35316 I'm sure we can get that guy who spammed every single thing they made on that (bing?) image generator to spam again.

This diaper has pockets! I think I'm going to retire the dude prompts...

>>35344 This is a common issue with prompts including "standing" and "wearing" in men. I just put "pockets" "hand in pocket" in to negatives and that usually does it. Also "dashion, posing" generally fixes it also.

Anyone have any experience adding AI generated images with bad tapes and trying to teach the concept so that it can be used in a negative prompt?

>>35353 It's easier to teach the correct presence of tapes. SDXL doesn't actually react well to broad negatives, you should only negative prompt specific things. "Girl", "Sunset", "Potato"... etc. The reason the tapes are wonky is because the AI doesn't know what they are supposed to look like correctly. This is why we train the models - we want to add concepts and finetunes we want to add specific informations about a specific concept.

>>35084 Like... 12 versions later I think I'm getting the LoRA where I want it. No longer biased with style. Still not getting the camo pattern with green and blue present reliably nor that "leather tag" which was the design of 2016(?) goodnites for boys. However it did learn this twunk/athletic body also, so I have to add negatives for that.

>>35357 What have you been adjusting over those 12 iterations?

>>35365 Primarily the dataset and it's repeats (and amount of epochs)*. Little bit timestep (currently 50 - 1 000) and then trying to find the best Rank and Dim. This is important because you need to have as high Rank as possible, but not more than is needed (or it learns irrelevant things). *This is difficult because 5 was too few, 6 was weak, 7 was bit too much. However more epochs changes this a bit. Another thing I been doing is that I been trying to spot unnecessary features it picked up. Since I used adult dudes wearing goodnites as the dataset it started to learn the masculine features. Then I cropped, deleted and added compensating pictures trying to balance it out. Oh... And biggest thing I figures out. Variety of resolution is better than high quality high resolutions. I get way better results by having pictures ranging in resolution. Smallest dimension I have is 632 and biggest 1240. It seems like smaller sizes prevent the oversized diaper that doesn't scale down to the subject's size in proportion, and it locking in to a specific area. Like past few iterations I been able to generate without issue so that the diaper can be mostly out of frame or covered by a shirt, and it still holds the correct details and form. Earlier on it was dead set trying to push it in to the frame causing all sorts of proportional issues. However detail bleeding is still an minor issue. Like bedsheets or shirt pulling that camo pattern into itself. But generally this seems to go away with adjusting prompts. (Which I'm worse at than in training). Also I can't figure why it wants to keep putting breast pockets into shirts. The picture posted is from last iteration to the earlier pictures.

I'm close to having something worth sharing. There are a couple issues that I think I am addressing in the current run. If they are sorted I'll post the lora for everyone to use.

>>35377 It's out on civitai. It's never going to be perfect, but given the complete lack of any decent diaper xl lora or checkpoint I figure it's good to get it out there. I've got a laundry list of things I have noticed it is overtrained on, but usually those can be fixed with negatives. Please check it out. If you find any weaknesses, let me know and I'll add them to the list if I don't already have them noted down.

I made a quick pacifier lora. SDXL isn't the best at inpainting but it still works half decent.

>>35461 Nice, are you gonna upload it to civitai? If you can inpaint a bunch of 'fullbody view, medium distance' generations and post them here, it would be appreciated. Im hoping SD3 lora's will have better quality and detail, but we will see.

>>35521 Nah, I need to generate a ton of training images so I can get a more randomized set of faces before I'd think of uploading it. It looks way too much like my training images at this point..

>>35536 I am by no means an expert, but I find captioning consistent names for redundant training people effectively hides the uncaptioned uniqueness of the particular person.

Try this lora out, the trigger is "ddiap3r" or "fullbody view of girl wearing a diapar,ddiap3r" So i been using a simple vroid poser to make a dataset, gonna work on making custom diapers lora's for pony by re-texturing it, It's not bad, and the 3D model effect doesn't bleed into it. I'll also try making a Paci Lora this way for pony. Enjoy i cleaned out my mega of old models and threw it in there. https://mega.nz/file/QetT0AxA#wKJfhyvdCGxw4NU-TypycGBa7I0lmA4cRtRHOPICcEQ

>>35584 Yep, I mostly have solved my issue with faces. You really can't tell where my training data came from now. Next will be getting colors to be a bit more random, or making it not so hard to prompt a color.

This thread is falling down the catalog pretty quickly. I don't have any content to contribute but someone should start a new one.

>>35644 Try setting adding training images where the face is partially covered/cropped out, but pacifer visible. I been wortking on a tape gag lora and that helped a lot with preventing it training facial features. Also I think blurring faces (but not the tape) helped. I also threw in few cartoon faces with taped mouth for good measure. After I realised that I started to utilise it in trying other things and it helps a fair bit. I been also trying to make a welding mask LoRA and I realised cropping tactically and having the welding mask in different parts of the picture helped. Give it a try. If it doesn't work for you then just revert back to your old means.

>>35647 How do you handle getting images that are cropped, but still large enough to be useful in a training set? Or do you just grab an image, crop out part of the face, and just leave it at 1024x1024 and just notate in the caption that it's a cropped image?

>>35648 My smallest training image on SDXL has been 380x380 and I had no problems. I make variety of different aspect rations. In photoshop I just crop them (do some adjustments to and quickly remove things like text which I know to mess up things) then scale so that biggest dimension will be 1024. By my experience small traning images do not cause issues AS LONG AS THEY ARE SIMPLE. Because keep in mind you can do random cropping while training when using Kohya. I don't use any captions at all. So I can't tell you anything about that.

>>35644 Kohya has a color augmentation option under advanced. It supposedly shifts the hue a bit during training to weaken the specific color link of the objects. Might be an option. I have used it as I have been focusing on a white diaper. >>35646 Be the change you want to see in the world.

>>35388 Adding less perfect shots at lower repeats really expanded the flexibility (check out the old vs new for the ballerina set). I'm up to 104 images now. Still a few issues that I have noticed, but I should be able to straighten up. Usually it's either a typo in a caption or one image that really confuses the training. The sock/thighhigh fetish was due to Base SDXL having no idea what a thighhigh is. Thigh high socks, or knee high socks cleared it right up. I spent a lot of time trying the words I was using the captions in prompts to see if Base SDXL understood them already. Ass is out, Buttocks is in. Really helped, but took forever. I haven't figured out the nipples through the shirt issue. There's only a few boobs in the training set and I captioned them all with nipples, breasts, topless. It appears I lost a bit of 4 tape consistency which I think I can get back by adjusting the close up repeats. The 60 new images watered down the close up to other ratio.

>>35046 >>35013 >>34897 Only remotely fappable pics itt. Aspies need to stop spamming 30 variations of the same shit.

I been working on Abu's King diaper. Using Prodigy now which is good, but I struggle to get the fine details reliably. But I been able to keep it versatile.

>>35658 Very good! Can you post somewhere your LORA or your checkpoint file?