AIDL: OmoAIShi Changelog

V17: Pee-ka-poo Edition

New update for AIDL: OmoAIShi (the diaper LoRA). Moar data means more better, right? Here's V17, with 725 more images in the dataset with perfectly balanced concepts.

Also included are versions for two very new models: Chenkin RF trained at 1024px, and Anima trained at 512px (the model was trained on 512px anyway). Anima is based on Cosmos-Predict2 2B and supports NL and much more compositions than anything SDXL based. There's also the NoobAI RF test LoRAs because I had those ready too. Didn't train a NoobAI VPred or Lumina 2 version and there were too many images to do a run on 1536px (sorry!).

A new concept added to the list are covered diapers, where you can have diapered characters under pants and they will have bulges, complete with waistband peeks. There's new data for each of the preexisting concepts too, which makes generation of them better and more accurate to the training data..

• Added more data throughout all concepts. Image counts have been curated so that each concept is exactly balanced against the highest count concept.
• Added new concept: covered diaper. This allows you to have those cute diaper bulges and waistband peeks
• Versions added:
• NEW 2026.04.03 Mugen @ 1024
• Anima @ 512
• Chenkin RF @ 1024
• NoobAI RF EQ-VAE @ 1024
• NoobAI RF FLUX.2 VAE @ 1024
• Illustrious 2 @ 1024
• Versions Removed:
• Lumina 2 @ 1024
• NoobAI VPred @ 1024
• Illustrious 2 @ 1536
• Replaced NoobAI VPred version with NoobAI RF (Rectified Flow) version
• Removed messing self tags from pooping in diaper images since a secondary trigger is redundant
• Removed plain diaper and clean diaper tags (and their pull-up variants). They were detrimental to generation and variations, since ideally a clean plain diaper should be what it defaults to anyway. Should help make x tape diaper and wetness indicators more consistent.
• Replaced some upscaled Pixiv preview pics with their higher-res Fanbox variants.
• Included tag and caption droupout for NoobAI, ILXL 2 versions
• Increased Network Alpha to 32 from 1 on SDXL-based models.
• Added new diaper design: attento
• Remove useless anime, drawing tags which ended up killing artist styles.
• Fixed artist names to match the Danbooru/Sankaku/ATFBooru names.
• winopz -> win opz
• oolong763 -> kyoju
• shinobu (appacandy) -> john doe shinobu
• azuma (pixiv 38220217) -> azuma (azumapic)
• stng_nicchi0609 -> ikayaki-san
• kamabok0 -> akakama
• kohaku (pixiv 29695938) -> kohakuwestale
• aiueo (air377) -> air377
• aonori_sss -> aonori sss (this is mainly a formality of changing underscores to spaces)
• BBCode converted readme for XenForo, Invision-based forums instead of a lazilly done screenshot.
• This update has made OmoAIshi's image count surpass 1500.

V16: Tag Restructure Edition

Welcome to V16! Refined tags, another new customizable diaper element, and the new Neta Lumina architecture brings AIDL to the world of Diffusion Transformers. While the Lumina variant is the standout variant of AIDL, the SDXL variants can still hold its weight well. With the Gemma 2 2B TE that Neta Lumina has, such concepts like covered diaper are possible without having any images of that concept, as well as the ability for natural language to be used.

Wetness indicators are now tagged to help stop the wetness indicators from appearing during pantypooping. You can also go and set up the wetness indicators to any color the model recognizes. The most trained colors are yellow and blue, which are common colors used for diaper wetness indicators, but if you use Neta Lumina, you can also set it to any color it knows.

Since TA went SFW only, and neither SeaArt and PixAI want to host the Lumina version, so this LoRA will only be on here, HF, and GitLab. I will try getting AIDL on alternative model sources such as LoRA Torrent and Diffusion Arc as well.

• Added handmade natural language captions to diaper and peed diaper concepts images in the dataset.
• The diaper and peed diaper concepts are hand captioned.
• Other images and newer images will use a combination of three VLMs summarized into one caption with tag grounding: JoyCaption Beta 1, MiMo Vision 7B, and Gemma 3 27B. This will be done on V17.
• This is to ensure that this LoRA is easily usable on models such as ~~SD3.5~~ Neta Lumina and any other future natural language models.
• Added artist tags to all identifiable artworks due to an accidental discovery where AIDL also learned artist styles whn there was enough data.
• Added new diaper design: merries
• Tagged wetness indicators as an attempt to separate wetness indicators from soiling (apparent when doing panty pees/poops).
• Redone prints and made sure to catch as many prints as possible, including the use of polka dot diaper, orientation indicator, size indicator
• Added LoRA Manager metadata

V15.1

• Added natural language captions to the peed diaper concept.
• Added new Illustrious 2 variant trained on 1536x1536. This also replaces the Illustrious 1 variant, as it also works on Illustrious 1
• Switched to training with 67372a's Kohya Easy Scripts instead of bmaltais' kohya_ss
• Moved internal AIDL stuff to a GitLab group.

V15

• First public alpha of OmoAIShi
• Added more clean, wet, and messy diaper art
• Added more clean and wet pull-up art
• Added new concept: open diaper
• This update has made OmoAIshi's image count surpass 1000.

V14.4

• Added damp diaper tag to help prevent very lightly soiled diapers from being used on the peed diaper tag.
• Lowered LR to 0.0002 (pre batch scaling)

V14

• Added more clean and wet diaper art. Mainly of the two-taped diaper kind.
• Added more pantypooping images.
• Added new tag: medical print. Use for diapers that look like Attends/Abenas/your average adult diaper. Keep in mind that you will still need to use the print diaper tag.
• front-print diaper now has new images for landing zone prints.

V13

• Added more pantypooping and pants pooping images.
• This update has made OmoAIshi's image count surpass 750.