This is a full restoration of Zero0Int and AbstractPila work on the CLIP-L model.
In 1000's of test images the FP32 model corrected bad anatomy in cases where the FP16 model failed (Seed to Seed)
In most cases the images have little difference but when differences exist the FP32 model was reliably more accurate (As it should be mathematically)
Note the vision models use the VIT-L vision blocks Zer0Int has a new finetune of the VIT/Vision model here
MIT License
Copyright (c) 2021 OpenAI
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Description
FAQ
Comments (13)
i grab all your tools thank for doing all this for us
awesome work, does this work in reforge too?
It should in forge, though it has/had some issues with the GPU slider overriding the FP32 CLIP command
I suppose my next question is if this works with sd3.5 as well or not. Think I'll try it out at least. Thx for working on this.
I think Abstract said the CLIP-G was trained with SD 3.5 in mind
I sense... a 3 on the end of something... I dunno Gabe Newell can't figure out a 3, we'll see if I can pull off a 3.5 instead.
Is that a Tree(3) reference, where you one up some one by adding a Tree(3)+ 0.5 - LOL
@Felldude I would NEVER make a multi-layered programming and game development industry joke. I'm a serious professional.
sips scrabble pieces from a martini glass and they spill everywhere
Anyone have a workflow with this working? I use a loadclip node and put it in place of the built in clip only to get "mat1 and mat2 shapes cannot be multiplied (154x1280 and 2048x640)"? Using the first pdxl model.
You cant use these with PONY models. Only PONY trained CLIP L and CLIP G are possible to be used and those are in every PONY checkpoint.
@Mescalamba Correct, the CLIP-G that is posted here can be used with PONY but not the CLIP-L models - I probably should have categorized it with the PONY CLIPs rather then by the original finetune poster
@Felldude While on it, do you think it would be possible to actually make PONY L clip from these? I tried merging them a bit, but its not particularly great (mostly hands get fked up).
@Mescalamba I think Astra changed the tokens so the need to use PONY CLIP-L remains