The long awaited T5 guidance for clip models V1 iteration into the ComfyUI suite.

The sliding window for 2048 larger context size IS NOT READY YET. It's got settings on the t5 loader though, so just ignore them like they aren't there for now. I'll push a change and make an announcement when that piece is ready.

I'm tired boss..

You can get the comfyui workflow FROM the images here; and you don't need to download the shunts from here at all. I uploaded them for posterity.

This isn't a base model in the terms of a SDXL model. It's an adapter meant to be used with specific nodes - built to introduce T5 guidance to CLIP models.

I tested it to be working with SDXL for now, last I checked it was having some sort of error with FLUX that I'll need to work out. I'll push some changes and make a flux workflow soon enough.

Check out the article I haphazardly threw together. There are many threads to this tapestry so unraveling them will take some time.

https://civarchive.com/articles/15340/t5-guidance-for-clips-comfyui-release

You can find the huggingface repos here;

https://huggingface.co/AbstractPhil/t5-flan-base-vit-l-14-dual-stream-adapter

https://huggingface.co/AbstractPhil/t5-flan-base-vit-bigG-14-dual-stream-adapter

And the dataset here; roughly 20m samples per shunt checkpoint, not the whole billion.

https://huggingface.co/datasets/AbstractPhil/human-templated-captions-1b