BloodOrangeMix (BOM) HuggingFace Model Card
All credit for this model goes to the original author(s) and the repository maintainer on HuggingFace, WarriorMama777. Please consider checking out their repository on Huggingface and giving it a like! Also, check out their profile on Civitai!
Compatible ControlNets available for this model at Difference ControlNets.
For SFW/Soft NSFW, see SFW/Soft NSFW. For less hardcore NSFW, see NSFW. For semi-real style, see AbyssOrangeMix2 - Hardcore. For an illustrated/painted style, see EerieOrangeMix2 - Hardcore.
Additional Terms of use
Clearly indicate where modifications have been made if used as a basis for any further work (i.e training, merging/mixing, extraction).
If you used it for merging/mixing, please state what steps (i.e. recipe) you took to do so.
From The Repository (WarriorMama777)
"This is a merge model that improves on the AnythingV3, where NSFW representation is not good.
It can produce SFW, NSFW, and any other type of artwork, while retaining the flat, beautifully painted style of AnythingV3. Stable. Popular in the Japanese community."
Description
This is an inpainting version of the model and is only meant for use with img2img and inpainting. This should NOT be used with txt2img and likely any attempt to will fail with an error.
Orangemix VAE is baked in.
FAQ
Comments (17)
Thanks for the pruned model. It's inconsistent, but I occasionally get this error message when generating images:
modules.devices.NansException: A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
I'm using "automatic" for the VAE selection since it's baked into the model, but I've also tried with explicitly using the orange VAE instead.
I don't get this error using the HF version.
Hey @kurokuro, I have a theory this might be related to the VAE being pruned and/or fp16. I am going to investigate this further. In the meantime, adding --no-half-vae in your webui-user (bat for Windows, sh for Linux), on the commandline_args line, should solve it, note, this will use more VRAM, using --xformers (typically always recommended for Nvida cards) and --medvram can offset it.
Hello @kurokuro , having difficulty reproducing the issues, I turned off --no-half-vae and it is still working.
I have uploaded a new version of the VAE that is pruned, but still fp32, try this one and let me know if it resolves the issue for you.
As a precaution, I have updated the VAE download on all of the models to the pruned, fp32 version, just in case. The baked in VAEs are still pruned, fp16.
Sounds good, thanks; just grabbed the updated VAE. Will post back with results (or if I don't see any errors in a reasonable amount of time).
@kurokuro how was it, did it work?
Clip skip 2?
What are the triggerwords? When using this it is not like the images. Suckin off, fuckin, etc...
Hello @Masolad ,
There are no special trigger words, these models are based on Booru tags. In order to get coherent sexual images, you would want to replicate the tags used by Booru in similar images. See https://danbooru.donmai.us/wiki_pages/tag_groups for reference.
@Masolad If you look at the preview images on this site, most of them will have an ( i ) in the bottom-right corner of the image. Click on that and it'll tell you the exact seed and terms that were used to create that image.
what dose it mean by "hardcore" NSFW. im not sure what that means. whats the hardcore in this example? not to get gross or disturbing but depending on one persons definition of "hardcore" might mean nothing
You are correct but usually NSFW means nudity, very revealing clothes, lingerie, etc while "hardcore" tends to mean explicit focus on genitals, explicitly sexual poses, sexual activity, etc
Hello! Is there any known downside by using pruned safetensors over the original? Be it loss of quality, more processing time or anything? This file is 1/3 of the original size. Thanks for your upload.
This checkpoint generating comics. It's interesting. It save the same outfits, tattoos etc. on characters in different "frames". Sadly, that frames often merging in one. You need carefully change cfg parameter + some extra tags in negative prompt to avoid it.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.