>>42413 >jarring cuts >Twitch screen when on Instagram 6/10, still funny though

>>42417 >source: dude trust me

>>42390 more runway

>>42413 >last political post dude, why don't you reach across the aisles and do diapered tradwives too?

>>42413 Better than I expected. Wish it was voiced and had sound effects. >>42434 Diapered tradwives are one of my favorite concepts to come out of the modern bullshit political discourse.

>>42436 >ywn change the mucky huggies of your prim and proper Christian housewife as she bakes you and your children a delicious chicken dinner in your suburban kitchen in the South

>>42413 what app are you using? this is pretty good

>>42443 The pain is real bro.

>>42436 Here's one of a reporter that is in over her head while reporting in Lebanon. It uses an AI voice generator and sound effects are from various sources. I'm sure someone with better skills could produce something way better. However, I'm glad AI has progressed to a point where I can bring some ideas to life even if they aren't 100% perfect. >also, political post, oops >>42391 I used Hailuoai. They let you upload an image and turn it into a video with a prompt. It's what I used for >>42118, >>42119, >>42138 and >>42259 >>42451 Hailuoai again just like with the video in this post, but with some editing, inpainting, etc.

>>42426 >t. babby's first political experience

>>42428 Can you do one with a visibly messy diaper? Or will it not let you?

>>42464 >6 tapes Apart from that kek. Feels like something that was a shitpost from the Middle East thread.

>>42464 My favorite part of this is the implication that the cameraman is there only partially to focus on the actually reporting but primarily there to make sure we get close ups of the reporter's diaper so that the average viewer at home can see.

Illustrious models and the new loras for them are pretty nice for a poofy and detailed diaper style.

It seems like big has died for content like this, don't you think, no alternative pays well?

You heard about space diapers, but have you heard about stasis tank diapers that don't absorb the breathing gel?

Well, I'm going to share what I was doing on Bing these days .

Guess which ones are SDXL Base + the Diaper XL lora vs SD3.5 Large with my initial Lora results. I'd say I've got a pretty solid lora ready to go at this point. Just waiting for a checkpoint...

>>42470 the fact that he's even CONSIDERING it should give you pause.

Currently training a pacifier lora, results look promising!

>>42667 i have to admit, whilst i think most ai art looks like shit, it is morbidly fascinating seeing the weird stuff people make with it

>>42668 let people cum

With the current state of ComfyUI, diaper loras and available anime models, it seems like it would be possible to create a visual novel without any drawing ability. Lack of image consistency is still a problem, but IPAdapter and other nodes to transfer styles, faces or general appearance are finally in a good state for SDXL (Pony, Illustrious) and mostly available for Flux. The image was created with Gimp. I used an Illustrious + thiccdiaper lora gen that I removed the background of with Inspyrenet Rembg and added the characters to a separately generated background. The thiccdiaper lora is pretty neat, very consistent, and only rarely puts the tapes on wrong. Much better than the SD1.5 loras I have tried, mostly because Illustrious is a great model. It would be neat if the same author made a normal diaper lora for Illustrious, I can't seem to find good other combinations for SDXL for anime style images.

>>42793 >seems like it would be possible to create a visual novel without any drawing ability. You clearly know more about the state-of-the-art than I do, but my impression (merely as a passive consumer) of it is that AI still has a very limited expressive range. Your own pic related is a good example: very conservative posing and perspective, generic character designs, can't handle too many characters in the same scene all at once. You could for sure still make a VN out of that, it'd just be kinda plain is all. The writing would have to carry it. What would you have to do to get it to produce more ambitious art than that? Not to sound like a snob, it's just that I still haven't seen anyone do anything more interesting with AI than the kind of slop that Lilith had already been churning out for years... by hand.

>>42800 >You clearly know more about the state-of-the-art than I do, but my impression (merely as a passive consumer) of it is that AI still has a very limited expressive range. I don't know very much, everything is evolving way too fast, and some modules cherry pick their examples so it is hard to tell what a module is actually good for, i.e. what deficiencies it provides a fix for. >Your own pic related is a good example: very conservative posing and perspective, generic character designs, can't handle too many characters in the same scene all at once. That's true, I'm not sure how much of the fault lies with the technology, though. Besides the fluff to make decent pictures and the lora trigger words, I merely added 3-4 tags, and "take_your_pick" is essentially ordering SD to do this standard pose. It is hard enough to come up with something original, but then the danbooru tags often behave in an unexpected way and you have to figure out how to actually get the model to spit out your idea. Do you have a fleshed out idea in mind that wouldn't be conservative posing with a boring perspective and generic characters? >You could for sure still make a VN out of that, it'd just be kinda plain is all. The writing would have to carry it. What would you have to do to get it to produce more ambitious art than that? I'm not really sure. It kinda feels like it can do anything (especially if background and characters are generated separately), but I'm drawing up blanks if I actually engage with the program. >Not to sound like a snob, it's just that I still haven't seen anyone do anything more interesting with AI than the kind of slop that Lilith had already been churning out for years... by hand. I share the sentiment that AI has yet to show that it can be used as a proper tool, especially if you want to tell a story. Churning out single diaper pictures isn't the problem, but getting it to output exactly what you want, consistently, is. What kind of content do you consider to be good? Any examples plus reasoning might help me, I've personally not been able to pinpoint what I dislike about most of my outputs made with AI.

Trying to make my own LoRA, but getting the picture set right is not very easy for SDXL. This picture happened by accident (prompt randomizer), this messy "anime artbook" style has a weird feel to it. I wonder how much more advanced our LoRAs would be if people openly shared their picture sets. It's not easy to find good pictures with enough detail, tagging them properly (or not at all) hugely influences how the LoRA performs (in either direction), and sometimes removing a single picture improves an already good LoRA substantially, so the more people fiddle around with it, the better it should become. Hosting a community like that in a way that allows for anonymous participation is probably not easy. Besides that, I also wondered if there's interest for an AI request thread. People describe the setting, characters, diaper style, etc, and people who want to play around with SD can try to make good pictures with the description as a guideline. I'd be up for it, maybe others like the idea, too?

>>42807 I really didn't have anything too specific in mind, and not everything has to be super dynamic and complex. In fact something like >>42824 is very much in the range of being "creatively viable" like I was thinking. Personally I think it's the difference between the character being in a "pose" and more so the character being in the middle of an action. Pic semi-related.

>>42825 "being in the middle of an action" adds something to my own mental model of what kind of pictures I might find more interesting, thanks. Some time ago when SDXL was still new and 1.5 got all those fancy modules I tried to create images where two people interacted in some way, e.g. by adding OpenPose and trying to make two people fight with wands, but that wasn't easy to do. Maybe you could say it still looked too much like they were posing. I'll try to play around with some minor environment interactions, like drinking some water from a glass, or playing video games. The bullying scenario from your picture looks interesting, too, but might be really hard to recreate.

Made my own LoRA for Illustrious/SDXL. Worked out well enough with my old unoptimized 1.5 dataset. Third image was edited, SDXL goes wild with patterns on diapers for some reason, but they are easy enough to remove. I've tried to upload the pictures like 15 times, but seems like 8chan now blocks Tor again. I won't risk uploading ecchi stuff over my normal ISP as I reside in a country that might jail me for doing this. Guess you won't hear much from me for a while.

>>42240 new ABU Little Kings LoRA https://civitai.com/models/963628/illustrious-abuniverse-little-kings-diaper-abdl >>42824 I've created a disk where I've put the datasets I used to train my LoRAs. I'll try to keep it updated. There are also some LoRAs that weren't published or were removed by civitai. Hopefully this will help us create more quality stuff. If you have any questions, go ahead and ask https://mega.nz/folder/HxoinZ5R#JrIhGXdqwZRDmGlnfPepGg

>>42841 Very cool of you to show your datasets. it makes sense but for some reason I thought pony and illustrious needed art images to train on. At first blush, the datasets seem huge with a ton of variety. Is that required or just something that you found that worked?

>>42533 >source: dude trust me

>>42841 Thank you so much. I still haven't looked at it in detail, but you collected a huge amount of pictures, which is something I haven't tried before. Have you tried tagging with quality tags? I've personally not gone over 100 pictures for the omuai diaper lora (was deleted on civit.ai recently, throwaway email didn't work anymore), so I threw out lots of pictures that I didn't like the quality of or that I suspected to lower the gen quality (did many gens, some had weird colors/image details that could be traced back to a single picture). With your approach, though, I wonder if tagging the lower quality pictures with "medium quality" or something similar might improve the LoRA. Not that I think any of your LoRAs need improvement, by the way. With SD 1.5, going for a pictures set size of 1024px worked much better despite the base model being 512x512. Maybe I should increase the pixels with SDXL beyond 1024. I usually went with a LR of 0.0001 (text encoder 0.00005) for 1.5, and now with SDXL (Illustrious 0.1), I'm experimenting with a LR of 0.00002 and disabled text encoders, and the results are maybe a bit nicer than with 0.0001 (three retries, ~80 pictures, 40 epochs, seems to work fine from epoch 5+ onward, but doesn't get fried even at epoch 40). There are a lot of buckets for the various resolutions, I'm considering reducing them by editing pictures manually or changing the bucket_reso_steps so there's a more reasonable amount of image resolution buckets. I have no idea if that'll help training SDXL, but it had a positive effect with SD1.5. I did over 40 experiments for the omuai lora, and I'd say there's like five distinct feels/bleeding effects to them, with 2-3 being mostly usable. Reducing the dim and alpha with a special script resulted in the one I released, and it actually behaved better (more generalized, less bleeding) than the "uncompressed" one. If you want me to try anything, feel free to ask.

>>42853 I've been building up a collection over the years, so I thought I'd try to add as many images as I can to the dataset. But personally I am not so happy with the results, because there are often inconsistent shapes and often my LoRA has trouble rendering tapes. In the last three branded diaper LoRAs, the amount of illustrations to total photos is about a couple of percent. So, I will try to create at the next opportunity a more niche diaper LoRA that will have a smaller dataset (maybe 50-200 images) and try to select only images with a clearly visible diaper (high resolution, good lighting and no weird camera angle) and a higher ratio of illustrations to photos than there was before. >>42855 I haven't tried using quality tags yet, but would like to try it the next time cooking LoRA. As for resolution - I tried to do datasets based on 1536px or 2048px resolution. Images that were above that resolution I left untouched, and those that were below I upscaled using waifu2x

>>42854 read project 2025, which your people pretended was a hoax until he won and you said that it was not.