MY T5XXL-Unchained finetune is not SD35m or SD35L even though they were paired.

So lets do a tiny recap shall we.

SD35 launched and was met with some backlash due the licensing shifts.

The original founder of Stability AI resigned his position, believing that no single person nor entity should be able to control systems like this.

So fast forward. One of the primary distributors of SD35 variants (dismal at best) has now removed any traces of those - my obviously SFW and liminal models included with any sort of NSFW variants. I really liked those liminal models you know. It's annoying that they were pulled, as uninterested as people were in them.

Given a bit of time I utilized a training process that finetuned a SD3.5M model ALONGSIDE the T5xxl variant that I was also finetuning. This caused some very interesting effects, and allowed the T5xxl to stay... somewhat compliant with all unchained-t5xxl compliant models.

Essentially this means they were an experimental parallel training, and this is two entirely different models existing side-by-side, while an entirely different variation of loss shifted not the T5xxl with SD35 - but the opposite, where the SD3.5 would be altered due to the changes to the T5xxl.

Now... my upload that coupled the unchained-T5xxl finetune directly with the SD35M weights lora was removed.

This is legally a T5XXL finetune coupled with a SD35M lora - which makes the standalone unchained-t5xxl VALID to be used based under the new rules of CIVIT and not the SD35M lora. The training issued to the T5xxl variant was trained not using dual-contrast or L2 or any of those KohyaSS losses - but trained utilizing an INDEPENDENT AND UNIQUE RESEARCH PROCESS TO ABSTRACT POWERED RESEARCH that allows the t5xxl to exist simultaneous WITH ANY CORE MODEL TRAINED ALONGSIDE IT, utilizing the captions from the images themselves and a form of T5 training that I have ran on multiple occasions for multiple variant T5 finetunes in the past - best showcased on my T5-small variants. Which is why said T5xxl functions with multiple models, and why the T5xxl can still have a conversation with the user.

So the unchained-T5XXL finetune is only going away if google changes the license, and even so I'm an independent non-profit researcher who licenses all of my new work with MIT licensing - meaning you can use it however you please. Make money with it, retrain it into a duck, have it explain to your grandmother quantum physics.

https://huggingface.co/AbstractPhil/T5-Small-Human-Attentive-Try2-Pass3

This can be seen in a multitude of systems such as the alucard walker system that proves finetunes completely independent of diffusion models to be valid and usable, even when the model itself has never once seen the T5xxl. This is shown to be obvious fact with the Alucard walking system, and more advanced trainings like the actual trainings I performed on the T5 small variants showcase even better control of the alucard system on models trained completely unrelated to an encoding structure when the correct similarity modifiers are applied to the features themselves.

https://github.com/AbstractEyes/comfy-clip-shunts/tree/dev2_electric_boogaloo

This unison-training is part of ongoing research and this system is in fact INDEPENDENT of the SD35 and thus this is a LEGALLY EXEMPT system trained simultaneously and not directly intersecting with that particular core model.

This T5xxl targeted weight changes finetune and the subsequent finetunes FROM THIS T5 FINETUNE, is invalid to the original SD35M terms - as the finetunes themselves are targeting not the SD35 but ALL t5xxl variants. Thus when directly decoupled this is a perfectly safe and viable T5xxl finetune for any sort of reuse.

This unchained-t5 is a viable and legally safe distributable variation of the T5xxl, as it was heavily trained on both synthetic data and captions - but not because of SD35M.