Update 1 - Fixing BeatriXL's errant NSFW

I've begun training the controlnet CN7 and CN10 dataset into Beatrix to counteract the NSFW. It should be a hell of a pose lora when it's ready; and it'll eliminate most of the errant NSFW that shows it's head when it's unwanted.

Roughly 150k controlnet pose images tagged with correct depiction and offset rotations.

I once had a simple question;

"How do I condense Flux camera angles into simple prompts?"

So my solution was the original Simulacrum dataset, an upgrade from the old "Consistency" dataset I had made at one point.

from side, from behind, from above, from below; the primary "from" tags that PDXL learned just fine and Flux failed to learn. Turns out FROM is a very specific associative linker with many angles other than a couple plus odd sex tags.

Okay well, 4 angles isn't enough.

• facing viewer, facing the viewer, facing viewer, etc

  • These cannot be used to train the torso, only by indirect association. Not only that, but the model itself seems to snap necks and leave torsos the opposite direction without enough training.

  • Danbooru tagging, Gelbooru tagging, and many other systems simply did not cover the majority of angles; so I ended up needing something akin to a dataset to control those angles.

• e621 tags are closer; front view, etc

  • They are good tags, but the dataset associated with the sourced images... aren't... human primarily, so they aren't very useful.

The Simulacrum V4 approach;

• camera angle fixation

  • from front, front view, pose, <other>

  • from side, side view, pose, <other>

  • from behind, rear view, pose, <other>

  • from above, above view, <other>

  • from below, below view, <other>

  • It's a simple spell but quite unbreakable. The weakest being from front and front view, required the strongest and most images to solidify PDXL Autism originally, so I fixated on a simple policy; plain English for Flux, and that carried heavily over to SDXL-SIMV4.

  • Designed with variations of every angle for multiple poses.

    • from front, from side, side view

    • from behind, from above, above view

    • etc

  • These were carefully sourced and fixated to be designers of an immediate scene. Entirely based on CONTROLLING chaos, and providing solidity to the systems.

When it comes down to it, the camera angle defines the scene; which is why I continued down the rabbit hole until I had an entire grid structure plotted and built around that concept.

The problem with the grid structure I constructed comes down to just a few pieces in the end; the primary one being the lack of data. I simply don't have the time to tag and caption the data correctly, nor do I have the computation to train the required data, nor did I have the expertise to train the AI related to identifying those and integrating them automatically bypassing the multitude of tagging policies that make no sense in systems like Danbooru or Rule34xxx.

Sourcing much data from many sources solved many of these issues, and yet introduced many unexpected new ones.