Workflow text2img MOSS-Image Local Uncensored Text-to-Image Generator

[ferris09]

https://www.reddit.com/r/StableDiff...simage_5b_apache_20_on_a_12_gb_rtx_3060_int8/

https://github.com/OpenMOSS/MOSS-Image

This is the first image generator that I've tried that can generate genitalia and sex acts without LORAs.

 

[ferris09]

https://www.reddit.com/r/StableDiffusion/comments/1x8awgl/mossimage_showcase/

 

[clockoff4]

Damn, it truly is uncensored out of the box for basically everything including violence. Don't ask how I know.

Spoiler

she bites his dick off

 

[nookkeeper]

I've been on the image generation space since sd 1.4 got leaked and this is THE most uncensored and capable model I have ever seen, literally sd 1.4 levels of uncensored (nothing). This could genuinely replace grok completely.

 

[clockoff4]

Not bad at all for a base model. It's pretty easy to prompt too.

 

[ferris09]

I still get quite a lot of crazy body horror stuff, melted faces, etc.. It's definitely not a perfect model. It can struggle really badly in scenes with more than one person. The workflow I've been using takes 80 seconds per image. I tried a workflow that took 12 seconds per image but the prompt adherence and quality were so bad it was unusable. This is the most uncensored model we've ever gotten so I'll give it some leeway on the other stuff haha.

 

[clockoff4]

Yeah I don't mind a longer cook time for images as long as we're talking minutes and not tens of minutes. For me it's been about 150 seconds with default settings at 2 MP (5070 Ti at 90% power).

 

[pondgasket]

I can't really stomach that kind of stuff, do you think its better for someone like me to wait until Ai gets so good that just doesnt happen?

 

[ferris09]

AI is advancing at an extremely fast pace. If you want to wait, that's totally up to you. I get lots of good results from Moss-Image, it's just kind of hit or miss. One image will be perfect, then the next image might be totally messed up. When you're generating an image every 12 seconds though it's easy to mass produce them and then skim through for the good stuff.

 

[EllisFuse3]

What are your specs and what is your workflow if you don't mind me asking?

 

[pierdrift]

I was impressed how well it does cosplay chicks. Did a Peach/Daisy tongue out together one with minimal prompting and it was spot on, down to the earrings.

 

[DiaryLena]

Someone is training a LORA for editing, this could be really big considering how uncensored this model is https://www.reddit.com/r/StableDiff...aring_both_moss_image_edit_methods/?rdt=51929

 

[clockoff4]

Well, I was running bf16 because "it's just an image model, what could it cost?" but nah there's no point. The difference is very small and int8-convrot is almost three times as fast.

 

[LenaByte64]

So this is exclusively T2I ?

 

[DiaryLena]

Yes, currently there's hacky approaches similar to Krea2 and some people are training it for i2i though it's way too early. MOSS themselves imply reference support is coming through, so we can be hopeful.
"We have found that simple SFT can enable features such as GIF generation and reference-based generation in MOSS-Image-Base. Stay tuned."

 

[ShedRobin]

man this is really good.. omg, how is this even for free?

 

[BexMeter31]

has anyone tried this

does it maintain facial geometry compared to reference image if pose is changed ?