Designed to look innocent, but she has attitude.
Every image starts with natural light, realistic skin, and a photographic feel.
She already has a personality, so treat her nicely.
Ignore every suspicious LoRA you found at 3 a.m.
Ready for an easy first date? Grab the wolFkroW_SSB.
Expect only the essentials: a few nodes, a prompt, and you’re on your own.
X marks the spot - if things get weird, tell me what you did. I know her.
Links:
Buy models or support - https://boosty.to/bl4cksnow or https://boosty.to/llikswonskcalb or https://app.lava.top/blacksnowskill
News, Exclusive - https://t.me/blacksnowskill
Description
INT8_ConvRot
FAQ
Comments (17)
Well, that’s it—you brought this on yourself. Now I’m going to flood this place with penises! 😄
Friends, this is simply an incredible result. Thank you for your work!
Not sure if I'm missing something or not but the X version and X (standalone) are both just the model weights without any text encoder or vae. I checked your attached workflow and it loads the text encoder and VAE separately instead of using the checkpoint loader which is what you would use for an AIO.
Also, any chance for BF16 version? I like to make my own quants.
My workflow is only for model X.
X (Standalone) loads via Load checkpoint
bf16 may appear later.
@llikswonskcalb I'm confused.. so the X (Standalone) is supposed to be the AIO? But you have the text encoder and VAE as optional files so you would still use the diffusion model loader and not the checkpoint loader...
@jmblsmit129 Optional files for those who already have TE and VAE; they can use the 7GB model as a diffusion model. The one that weighs 12GB already includes everything necessary and can be loaded via load checkpoint.
@llikswonskcalb I downloaded one of your example images and looked at the workflow and it's clear now! I think the biggest thing that confused me was the file size and calling it "standalone". The size is the same as a fp8 quantized UNET. Calling it AIO would have been less confusing but I could just be weird LOL.
will the mxfp8 work in 40 series?
It’s possible, but difficult
Don't bother. It's far slower than int8-convrot. I gave setting up the AIO a shot and it took 16 seconds longer to generate a single 1.2MP image with no LoRAs than using any of the 13GB int8-convrot models plus the extra 5GB for text encoder and VAE.
Sadly, much as I love the preview pictures this is not a 4000 series release.
Int8 is up and there is much to see what with the things and all.
This sucks. Because it looks fantastic.
With a lowly 4060 it takes me 20-30 seconds to generate a 1.2MP image using int8-convrot models... this AIO saves me around 6GB off of that stack. But it takes about 16 seconds longer per image.
It would be really awesome if you would put up an f16 so we can convert it ourselves if you don't want to do that. It's a magic format. A 4060 8GB can run at 95-100% utilization on 30GB model stacks in int8-convrot. I think it's because witches.
I will do my best for you, my brother.
you're welcome
@llikswonskcalb My hero. Guess I know what I'm doing tonight. :D
Down from 37 seconds per image with no LoRAs to.... 17 seconds. Actually pretty much a new record for a 4060. Thank you. :)