Custom huge dohonkerbonkeros diffusion model v(-1) (Patreon)

Linked is a *very* early preview ckpt of the very WIP model I'm working on. I've been aggressively exploring strategies and rewriting portions of training code to finesse the least-destructive training possible. Attached also are some example raw outputs.

Mega.nz Download : (1.99GB) *(Trained at clip skip 2, 512x512 base resolution, default VAE)

TLDR Limitations and Best Use
This is just a preliminary "hey look I'm not dead, here's what I'm working on" sneak peek, and by no means is to be considered something worth even a "v1" tag. This model is still heavily undertrained, and will probably be more trouble than its worth for anyone not intimately familiar with danbooru tags and negative prompting

- For anything passable, the tags "worst quality, cartoon, retro style" in the negative prompt field are basically mandatory. This is a symptom of it being undertrained.

- Sizes larger than 'gigantic' are particularly undertrained, and will typically output low quality results

- I accidentally nuked the "Flat Chest" tag out of existence (you may think "this is a hyper model who cares" but the advanced diameter appreciators among you will know; small thing next to big thing makes big thing look even bigger)

- Currently only trained on hyper breasts, and only 1/4 of current hyper breast dataset at that (planned is hyper breast, belly, butt, futa .. the whole 9 yards)

- I ask that any lora/finetune posting to citvai or alike wait until the base model's made "public" in a week or so, just to be fair to all patrons who have been incredibly patient with being involuntarily dragged along with my diffusion shenanigans

Boring Technical Details
The effort in performing a minimally destructive training is mostly paying off, if you see the attached comparison photos attached (upper row is base model WD, bottom row is training), so far the sizes I've added are coming with very little 'extra baggage' - meaning, while all other tokens + the seed remain the same: the only thing that should change is the prompted breast size, which is nearly the case. I'm still trying to further improve on this. In theory, the "perfect" training results in a model where asking for a different bust size *only* changes the bust size.. if I could figure that out, then it'd mean the gates are basically open to add new content indefinitely without prior model knowledge being 'forgotten'. "Catastrophic Forgetting" is a well researched topic in GAN and diffusion models, where the more training you perform, it "forgets" prior training: this is most visible in some of the lower quality ""finetunes"" floating around, where the model output reliably reflects a certain style, but contains very little flexibility to produce anything but standing/leaning cowboy shots of medium bust girls facing the camera. I'm not sure that this could ever be "beaten" (and if so, it would be by some research team at harvard, not some rando trying to proompt the biggest horkiest tiddies on their 3090) .. but I think that with care it may be able to be mitigated to an almost trivial amount