CivArchive

    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)

    Cosmic_CrafterFeb 4, 2025· 2 reactions
    CivitAI

    i grab all your tools thank for doing all this for us

    LyloGummyFeb 4, 2025
    CivitAI

    awesome work, does this work in reforge too?

    Felldude
    Author
    Feb 4, 2025

    It should in forge, though it has/had some issues with the GPU slider overriding the FP32 CLIP command

    MysticMindAiFeb 4, 2025· 2 reactions
    CivitAI

    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.

    Felldude
    Author
    Feb 4, 2025

    I think Abstract said the CLIP-G was trained with SD 3.5 in mind

    AbstractPhilaFeb 4, 2025· 4 reactions
    CivitAI

    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.

    Felldude
    Author
    Feb 4, 2025· 1 reaction

    Is that a Tree(3) reference, where you one up some one by adding a Tree(3)+ 0.5 - LOL

    AbstractPhilaFeb 4, 2025· 3 reactions

    @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

    emails_save_livesFeb 5, 2025
    CivitAI

    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.

    MescalambaFeb 12, 2025

    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.

    Felldude
    Author
    Feb 13, 2025

    @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

    MescalambaFeb 13, 2025

    @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).

    Felldude
    Author
    Feb 14, 2025

    @Mescalamba I think Astra changed the tokens so the need to use PONY CLIP-L remains