mrpopsalot/aidl-mini

AIDL: Mini

These models are designed for **NSFW** purposes, trained and tested on DRAWN CHARACTERS ONLY.

EXPERIMENTAL MODELS. TRY AT YOUR OWN RISK.

Overview

AIDL: Mini is a small version of AIDL: OmoAIShi meant to be used to easilly iterrate on training and experiments on architecture and setting (with my poor little RTX 3090). Instead of 1700 images throughout 9 concepts, there are only 80 images throughout 4 concepts. The images are taken from V16's dataset.

• Trained Models:
• NoobAI-XL Rectified Flow: Bluvoll/Experimental_EQ-VAE_NoobAI_tests/NoobAI-RectifiedFlow-test-step486k @ 1024
• FLUX.2 Klein 4B: black-forest-labs/FLUX.2-klein-4B @ 512
• Anima: circlestone-labs/Anima @ 512
• Developed by: Mr Pops Alot
• Model type:
  • SDXL-based: LyCORIS LoCon
  • All DiTs: LoRA LierLA
• License: Apache 2.0

This is the second LoRA of the AIDL series. AIDL: The AI Diaper LyCORIS is intended to a series of LoRAs that teaches models about diapers and other ABDL related concepts.

Features

• Tape-on diapers: Two and four taped, printed and plain diapers are supported.
• Peed diapers: Wet diapers of various stages, from slightly wet to completely leaking. Speaking of leaks, they now correctly go at the side of the thighs.
• Diaper and panty pooping: Soiling of various stages, including blowouts have also been trained, as with double pee and poo soiling as well. Pants, shorts, and dress pooping is also covered. Compatibility is not guaranteed (there is less training data). You may see some exposed scat.
• Consistent results across different character positions: This model has been trained on a variety of styles and poses, ensuring correct generation on many poses and whatever artstyle you want to use, including standing, squatting, lying down, from below upskirts, and more

How to Use

Generally, most things that you can do with OmoAIShi, you can do with AIDL: Mimi.

Diapers

Use the general term: diaper, or in a diaper.

• You can also specify the tape count using two taped, or four taped.
• Prompting plain diaper will get you a diaper with no designs, while print diaper will get you a printed diaper.
• front-print diaper and back-print diaper can also be used to get you a print on the front or back. Use Danbooru's list of prints.
• Add in wetness indicators by prompting yellow wetness indicator or blue wetness indicator.

Diaper Soiling

Peed-in diapers: Use the general peed diaper tag to get wet diapers. From there, you can also go and set how soiled the diaper should be. Without it, you might get unexpected diaper blowouts.

1. wet diaper
2. soaked diaper
3. drenched diaper
4. leaking diaper, pee on legs

Poopy diapers: Use the general poopy diaper tag to get messy diapers. From there, you can also go and set how soiled the diaper should be. Without it, you might get unexpected diaper blowouts.

1. messy diaper
2. packed diaper
3. loaded diaper
4. diaper blowout, poop on legs

smell can also be added to your prompt to give off that soiled smell effect.

If you want more soiled diapers, you can increase the weight of the tag to suit your needs. Tags works well between 0.9-1.5 weight.

Anima Note: For poopy diapers, use pooped in diaper instead of poopy diaper. The latter term does not work on Anima.

Pantypooping

Use the term pooping self. You can also use the terms poopy underwear or poopy clothes to have either their underwear be soiled or their lower clothes (like their dress or pants) to be soiled.

Multiple underwear types supported!

For diapers please use the poopy diaper tag. Pooping self can also work but poopy diaper is the recommended tag for diaper pooping.

Bias, Risks, and Limitations

Due to the deliberately limited number of images in the LoRA, the LoRA may not be fully trained and may have issues such as improper tapes.

As well, there may be some limitations with this LoRA:

• Panty poops may happen unprompted if smell is prompted
• Peed diapers may have poopy diaper bulges unless poopy diaper is put in the negatives
• Poopy diapers may have unprompted blowouts.
• Although you can use Rouwei Gemma in the SDXL variants, you'll only get pull-ups

Increase the strength of the LoRA if needed to achieve your desired results.

I am in no way responsible for any content that the user creates with malicious intent while using any of these models. The user assumes complete responsibility for the misuse of these tools and any unethical content created with them. Follow all local, state, and federal laws as well as online rules regarding this content and generations with this LoRA. I am not responsible if you are arrested or banned from an online service for using this.

This LoRA violates CivitAI's terms of service (their TOS does not allow diapers nor scat). Do not upload generations from AIDL onto CivitAI.

Training Details

80 images have been used in the dataset. Images used are exclusively drawn 2D art.

No synthetic data has been used for the model. Tagging (Danbooru tag style) is all manually done, as automatic taggers do not recognize the concept in question well.

Training Procedure

Training configs are provided with the weights, but as a TL;DR:

SDXL-based (Illustrious, NoobAI)

• Program: Kohya sd-scripts Bluvoll fork
• Optimizer: CAME
• Scheduler: Rex
• UNet LR: 5e-5
• TE1 LR: 3e-6
• TE2 LR: 1.5e-6
• LoRA type: LoCon
• Epochs: 10
• Network Rank (Dim/Alpha): 64/32
• Precision: Full BF16
• Hardware: 1x RTX 3090 24GB

FLUX.2 4B

• Program: Musubi Tuner
• Optimizer: CAME
• Scheduler: Constant
• DiT LR: 3e-4
• LoRA type: LierLA
• Epochs: 80
• Network Rank (Dim/Alpha): 64/1
• Resolution: 512px
• Precision: Full BF16
• Hardware: 1x RTX 3090 24GB

Anima

• Program: Diffusion Pipe Bluvoll fork
• Optimizer: CAME
• DiT LR: 1e-5
• LoRA type: LierLA
• No LoCon needed as they'res no Conv layers.
• Diffusion Pipe only supports LierLA models.
• Network Rank (Dim): 128
• Precision: Full BF16
• Hardware: 1x RTX 3090 24GB

FAQ

• Why is the dataset and training configs public? As AIDL: Mini is a tiny version of OmoAIShi, designed for easy iteration, it makes sense to release the dataset so you guys can go and further improve the mini model and use it to train on other models.
• No pull-ups? No pull-ups. I have only trained diapers, diaper spoiling, and pantypooping.

License

This LoRA model is provided under the Apache License 2.0 license.

Special Thanks

• DangoKnight for training some of the test versions of this LoRA
• Yatogami for contributing to the dataset.
• Hugs288, Lopi99, and others in the NoobAI discord to for helping me get my LoRA training settings right.

Model Card Contact

I can be contacted on:

• Matrix: @mr_pops_alot:cutefunny.art
• GitLab: gitgud.io/mrpopsalot
• All The Fallen: @MrDoesNothingAlot
• Discord: 474149843724664854

Like my work? Support me by giving me a tip. This model will always be free, and open source. There will never be any paywalls, hidden fees, or any content locked behind a subscription.

Support	Regular options Crypto
Chat

I do not have a SubscribeStar, as they do not seem to want to approve my account. I have never and will never open up a Patreon, as they are infamous for their ban waves and attacks against the NSFW community. Please support me on these alternative services.