CivArchive
    KitchenSink2fp16 - KitchenSink2fp16
    NSFW
    Preview 297345
    Preview 297363
    Preview 297361
    Preview 297360
    Preview 297359
    Preview 297358
    Preview 297357
    Preview 297356
    Preview 297354
    Preview 297353
    Preview 297352
    Preview 297351
    Preview 297350
    Preview 297347
    Preview 297346
    Preview 297371
    Preview 359012
    Preview 359011
    Preview 359010

    KitchenSink2.0 Final

    fp16 LOW VRAM (sfw)

    A 768-1200px General Enhanced Model
    1TB total merge

    • SD 2.1 768+ based. YAML included! (fp16)

    • Resolutions of 768x768+ to 1200x1200+

    • Built for artistic, can do realistic.

    • Listens very well to natural word prompting.

    • Simple to complex - type a few words to a dictionary!

    • Call to it! Cartoon for cartoon, realistic for real!

    • Strongly suggest using face restore on humans.

    • 3 Sink flavors have same outputs! Pick any, all the same!

    • GRID samples included to download and reference

    • SETTINGS included below!

    • Website gallery attachment limits sorry! (site flagged suggestive here as nsfw. they are safe with clothing on all sfw models!) 100+ samples spread over 3 flavors of the model! Full gallery in one place found here:

      SFW gallery > https://www.artstation.com/artwork/NyyyGN

      NSFW gallery > https://www.artstation.com/artwork/aoooy9



    EDIT: Openart made a fantastic PDF booklet on natural prompting which this model likes very much! This will help talking to it and getting the best results. =)

    Note: All 3 variants (fp32/fp16/nsfw) are the same output. Model description on all 3 is near copy paste !! Rename both file/yaml as needed - civitai changes the file names often!

    Hello everyone! This is my goodbye present to the community. =)

    "Everything but the kitchen sink"

    I give you the sink then, a big boy incremental by month merge of every working SD 2.1 768 based model from august until this post. I put the long boring complicated description on how it is blended in the 1.8 version text box, which is archived under my profile.

    For those up to date with my 1.8 -> this one rolls in all of feb/march until the 20th, into each their own 1.8 & 1.8L, and those sloppy drunks haha were additive blended into each other to make the 2.0! =)

    You can skip this mixing info if you want:
    IF the setup images are too small, link for full size here!
    https://www.artstation.com/artwork/ZeN8G0
    I cannot control how badly this website ruins uploads sorry!!

    SETUP NOTES

    The PURPLE changes below are to modify the default size from 512 into 768 for 2.1 models so you don't have to change it every time you open SD.

    You do not have to do this if you don't mind =)
    Don't forget to save your changes!

    The WEBUI-USER launcher edits are recommended !!!

    > Use the LEFT side for regular computers with recent modern video cards to run the 5Gb model flavors in fp32 at max speeds.

    > Use the RIGHT side for older low vram 8gb or less video cards. This will tweak the startup to use all fp16 top to bottom when it can. (see below)

    ** the fp32 vs fp16 argument **

    There is none. Just empty air talk. I put a grid in the samples to prove it. you can test yourselves on both versions. It's math in the end. =)

    So why is this even an issue? Honestly? Because early video card manufacturers lied. Mhmm. Older cards might actually do real or fake fp16 for math (it's really a crippled fp32 chip doing 32/2 math...heh)
    This was only discovered when you go plug in math for something recent like ai, using the chip for alt purposes than video games, and it bugs out... but the 32 math works... wait what? that came later..how.. people did checks. yelling. yada yada. that's the real story why we have a f16 mess. Not knowing if you have a crippled older card as there is NO way to verify as far as I researched...the --precision autocast in the above is the safest detour. (this forces fp16)

    For you? At home? No difference on modern cards. Older 8GB cards? I suggest the fp16 to give plenty of overhead when doing larger images. I used a dinosaur GTX1070 8gb for everything here. you can see at 1200 rez it maxes the card in the details under the prompt =)


    legality / copyright

    Do whatever you want. It's all nonsense.

    All models were based on the laion-400 dataset which is a German grey area scrape to begin with. SD, MJ, the lot. So at a root if you are not litigating vs them - you are already failing as everything is based on it top to bottom.

    In my non lawyer, logical opinion - it's not my 5gb or her 5gb model... it's the laion400 vs anything in court as it's the daddy. Other people can have their opinions this is just my way of looking at it logically. I can't claim copyright on anything that was in a scrape of the internet for 10+ years in public domain. You just can't naturally by polling the data, so the ai combining it through a straw is no different - You still slurped from a giant grey area pool of data. I just can't see how anyone on either side can waste time on this without pulling in the original company to the fight.

    So do whatever you want as I'm almost positive that will never happen as it's zee germans that made the L400 scrape, which happens to be physically hosted on massive multi-petabyte server in east germany, and billion dollar tech companies used it as their base too. cough cough. yeah.

    The Germans made the scrape - other humans split it into a million personal unique models spread across the internet - I put it all back together again. Nothing more, nothing less.

    Enjoy the model responsibly !

    / End



    Personal notes

    Everyone is capable of doing as they please, including walking away. That is the choice I'm making. I do not like the direction people are dragging and flooding all the ai resources with their nonsense, disgusting and quite frankly illegal material. I just don't want Hanson kicking in my door by association because I visited a site for another model. If you remove the nsfw filter and browse this site raw, you will see exactly what I mean. Nobody is saying anything, reporting or taking action to clean anything up. The adult in me is disgusted.

    So I'm out! Was fun while it lasted. =)

    Links under my profile to reach me for special requests, media, or for private offline Ai r&d work.
    Coffee link is there too if you want to help me destress with a brew.

    BarQ!

    Description

    Low vram 8gb or less flavor of model. THERE IS NO DIFFERENCE TO THIS ONE OR THE 5GB VERSION FOR OUTPUTS. This is only tuned for low vram ~ I got you fellow gtx10** dino carders. =)

    I Have this working on a 1030 3gb vram model as well by adding --lowvram to the gui command line after the rest. =)

    (note: dont use this extra command on 8gb+ cards!! just use the fp16 and my config. It's the fastest and best I've found so far to date.)

    FAQ

    Comments (2)

    fablegeniusDec 16, 2023· 1 reaction
    CivitAI

    any plans to make checkpoints for 1.5 or XL?

    BARQ69
    Author
    Jan 3, 2024

    My apologies for missing this message.

    1.5 is far too old sadly for resolution sake (check out the new SD Turbo as a modern replacement to it that is faster and produces better quality!) to be making a model at this point in time. (I honestly was never a fan of 1.5)

    As for XL, well, I won't lie! A week after release I already had one trained but decided to keep it unreleased here and on hugging due to the issues I brought up months ago along with civit here recently getting cut off until they clean up their acts moderating illicit content. It still is disgusting and terrible honestly.

    Due to this, my model is now called VarieXL (in case you ever see this on an image tagged in the wild!) and will be for commercial release soon into my own public ai service company as it was trained to do 1920x1080p of sorts making it unique in this model world. Yourself, as well as others will be able to directly use the company ai for private projects retaining all copyright, As I own the commercial model and know how I "ethically" trained it due to the new emerging laws I can grant it as well knowing it's cleaned up. ;)

    Even more exciting is the new SDV-XT video model which uses the special XL model as a base to make the images - its looking sweeeeeet! as they say. I've already posted some video samples in a discord channel and everything will be ready soon for public release.

    PM me directly if you wish to know more. =)

    Have a great day!

    Checkpoint
    SD 2.1 768

    Details

    Downloads
    654
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/21/2023
    Updated
    4/30/2026
    Deleted
    -

    Files

    kitchensink2fp16_.yaml

    Mirrors

    CivitAI (83 mirrors)
    Other Platforms (TensorArt, SeaArt, etc.) (1 mirrors)

    kitchensink2fp16_.safetensors

    Mirrors

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.