[Return] [Bottom]

Posting mode: Reply

BBCode
(for deletion)
  • Allowed file types are: gif, jpg, jpeg, png, webp, webm, mp4
  • Maximum file size allowed is 25000 KB.
  • Images greater than 255 * 255 pixels will be thumbnailed.
  • 99 unique users in the last 30 minutes (including lurkers)




    So, I made a seconde version of Sagittal Plane Cutting LoRA with Pony and Illustrious

    https://mega.nz/file/sOtkSYhA#7DLvrMwc1bQME-TDiG3_L-dF6eglzuZUm4JSp3IMq5M
    Pony version, recommended strength 1 ~ 1.5
    Kind of weak by default, kinda works with strength 1 in base checkpoint with nude, but with more clothes and additional Loras, Higher strength is required.

    https://mega.nz/file/QXdlBZiZ#Klm5doZRheqYsGZWx7bYpSGjivbSRWFnOOHmI2wxgVw
    Recommended strength 1~1.2
    Better than Pony in pretty much every way.

    Trained with 271 images
    THIS WAS A FRIGGING TORMENT
    It was hard to get some good training images (There was like 1 sagittal plane cutting for every 3~4 transverse plane cutting) and even harder to get good images, so I struggled with the severe quality degradation.

    I had to consult with ChatGPT and had to to pull out all the tricks such as locking out block dimensions, using reg images, two-phase training and shit.

    I'm considering making a more stable version of Transverse Plane Cutting LoRA using the new skills I learned making this.

    Here are some sample images starting Gamia Q who always seems to get cut in half. Pony 6 base version.
    Viewing last 50 posts
    >>
    >>11009
    do you still have the dataset? i'd love to train this on chroma or maybe even qwen
    >>
    >>11565
    Here it is
    ttps://files.catbox.moe/hag7dv.zip
    >>
    >>11600
    nice. browsing through it, there are quite a few censored and just extremely low quality images. for chroma/qwen i might remove these and add all the natural language that the models require. but honestly from my testing even a simple phrase is probably good enough.
    >>
    >>11565
    Let me know if you are successful with chroma, its soo good at generating nsfw, would love to do an ayaswan with chroma.
    >>
    This works surprisingly well. Nice work.
    >>
    >>11619
    still training it. it "works" but it's not at the same level of quality i would hope it's at. i think i have to retrain and further clean the dataset.
    i left the booru tags and removed some of the censored images.
    for my next try i will add a short sentence to describe the scene as chroma just loves that natural language bullshit. this should help with generating way more specific stuff like "the woman on the right has been sagittally sliced while the woman on the left laughs", as an example. the current lora just applies the effect to everyone 80% of the time.
    >>
    261086
    does any experts give her a Sagittal Plane Cutting? thx.
    >>
    >>11691
    here
    chroma lora, try this please. should work with any art style you throw at it. you can "activate" it by using "she has been sagittal cut" for instance. it should be vague enough that you can prompt it pretty openly.

    if you have multiple characters i recommend you either inpaint or gitgud at prompting with chroma. can be a major hassle but can pay off.

    use it at weight 1.

    https://files.catbox.moe/fvcrdr.7z


    also, if someone can post the transverse plane dataset i will train that for chroma too.
    might revisit the sagittal lora but for now it works well enough.
    >>
    >>11620
    I was able to get WAN2.2 to animate this image really nicely.

    This new checkpoint seems to be able to do a lot of interesting things: https://civitai.com/models/1981116/dasiwa-wan-22-i2v-14b-lightspeed

    With just that and this lora (https://civitai.com/models/1811313/wan-dr34ml4y-all-in-one-nsfw) I have been able to get it to clean up imperfections in lora-tortured generated images.
    >>
    can you reupload?
    Catbox seems to have gone the way of the dodo as well. :(
    >>
    >>11717
    >she has been sagittal cut
    Thank you! for putting in the effort
    i'm not sure i'm prompting it right, or maybe its the gguf version, let me try the Chroma-HD
    >>
    Decided to try it
    >>
    >>12386
    Second thing
    >>
    >>12387
    Third thing (Technically my first thing but I shared it last)
    >>
    File: guro_00032.webm
    (5.68 MB, 992x1280)
    5950751
    wan2.2
    prompt: a woman is standing in a japanese room. a man off-screen slashes her from top of frame to bottom of frame with a greatsword. she is cutting in half vertically. a huge amount of blood splash out from inside her. she falls backward one on each side.
    >>
    posting some of my creation
    >>
    >>
    >>
    >>12512
    (O_O) Oh!!! Wow!!! That's really hot. Such a sexy body, and lo at those tits bounce! and the detail in the organs is amazing!
    >>
    >>12512
    Looks hot! Can you share your workflow?
    >>
    This LoRA is amazing, it's a shame civitai doesn't allow anything remotely fun. They allow the wildest grossest niche sex stuff, but even hinting at violence is a non-starter. Watching this thread for more!

    What kind of keywords are useful? I'm using "sagittal, bisected, (brain \(organ\))". Trying to hint at "heart" tends to introduce heart shapes instead, and "organs, spine, intestines" tend to make the cross section really sinewy sadly.
    >>
    looks cool if you don't ask too much of it! Thank you!

    It'd be cool to have a version that just did the silhouettes of organs, or just a blank bisection, like manga do.
    I don't like realistic/real hi-fi gore, it's way too much... the subtle painterly manga bisections from samurai stuff is way hornier.

    also holy shit AI checkpoints using booru/pony are tagged so poorly when it comes to gore. gore artists staying employed
    >>
    Re up?

    The file you are trying to download is no longer available
    This link is unavailable as the user’s account has been closed for gross violation of MEGA’s Terms of Service.
    >>
    >>15018
    read the thread, there's a catbox up >>11388

    btw, very good job with this one, op. can't believe how well it works
    >>
    >>15015
    Try adding 'intestines' and 'organs' in the negative prompt.
    >>
    >>
    >>
    I'm really liking this chroma Lora. Do you have anything else for chroma?
    >>
    Tried it out again (I was the guy who did Rouge, Charlie, and Loona)
    >>
    >>15236
    no, but i'm willing to train more on chroma if people post the datasets.
    >>
    :D
    >>
    >>11388
    these are so good, using the illustrious ones. any good phrases or words to help trigger the effect?
    Also, are there any differences between the LoRAs *_05.tensor to _*_20.tensor etc, or are they just different versions?
    >>
    >>15985
    Different epochs. Sometimes some of them perform better than the other. Occasionally the image might break under certain conditions, the style gets affected too much, then try use different epoch.
    >>
    >>11388
    I think catbox is also down. Could anyone re-upload it?
    >>
    >>17401
    No need, just search for the keyword, someone uploaded them on other platforms.
    >>
    all links down
    >>
    >>11388
    Links is dead =(
    >>
    >>12912
    It's rly cool!
    Whant to learn how to make this!!!
    >>
    Does anyone still have the lora? All links are dead and would be grateful if somebody could re-upload it. Thanks a ton!
    >>
    >>20559
    ttps://files.catbox.moe/bzrnve.Zip
    transverse plane cutting - pony (password : transverse)

    ttps://files.catbox.moe/qm8hhq.Zip
    transverse plane cutting - illustrious (password : transverse)

    ttps://files.catbox.moe/fvzstd.Zip
    sagittal plane cutting - pony (password : sagittal)

    ttps://files.catbox.moe/0bwgab.Zip
    sagittal plane cutting - illustrious (password : sagittal)

    there are multiple files in each zip, each file with different epochs. Usually all work well but if it sometimes break, try using a different epoch.
    >>
    >>20658
    Thank you very very very much!!!!

    (It would be even greater if you could reupload the training set too)
    >>
    >>
    8575575
    I tried merging multiple videos.
    >>
    Will OP or someone train a lora for diagonal cuts? like how Nakoruru is sliced in Samurai Shodown? AI is junk at internal organs, so one trained on just silhouettes of the cross section and getting the separation correct, would be good and wise...
    >>
    >>20658
    Have you thought about using 7z archives with the file names encrypted? I can only imagine that the unencrypted file names are why the previous links got deleted since the current ZIP archives, despite requiring a password, openly display the file name contents without a password.

    And maybe sharing some SHA256 checksums too? Or heck just throwing an actual checksums.sha256 hash file into said 7z archive(s)

    (I'd consider doing it myself but my upload is rather slow - like 1Mbps aka ~200KB/s peak if not doing anything else)
    >>
    >>21215
    Making a LoRA, especially something complicated like bisection requires lots of training data, and diagonal are very rare, especially when you want it to work on 2D. (too much 3D images in LoRA tends to makes the output look 3D as well)
    That's why I also gave up on coronal plane cutting.
    >>
    >>21371

    Is there a list somewhere of all guro related loras?
    >>
    >>20658
    >>21370
    I was humoring myself and looking into whether I actually could in fact do an encrypted file names 7z archive, but realized something:

    Catbox says they max out at 200MB even when logged in, and the largest of the existing LoRA Zip archives on Catbox is nearly 600MB!

    What secret am I missing for uploading >200MB files?
    >>
    >>22728
    I don't know about catbox, but in case they simply lowered the max filesize limit while keeping the old files:
    you can span a single file into multiple zip files
    https://duckduckgo.com/?q=multipart+zip+file
    >>
    >>22738
    I'm well aware of the splitting of archives.

    But I actually checked for old snapshots on web.archive.org and it shows that catbox has had a 200MB limit for every year since 2015 (not a typo)

    [Top]

    Delete post: []