Freedom.Redmond is here!
You can currently test it at this link thanks to the makeai. run API.
I'm grateful for the GPU time from Redmond.AI that allowed me to finish this model!
This is a generalist model fine-tuned on SD 2.1 768X with a focus on generating images in >1024x.
The model has a high capacity to generate realistic, artistic images, cars, people, and a wide variety of themes. It's a versatile model.
This model will serve as the basis for a dozen models and LORAs that will come specialized in specific themes.
I highly recommend DPM+ SDE/2M or 2M SDE settings with 30 Steps. In resolution above 768x.
I also recommend testing some prompts with or without negative prompts as there are cases where the results are also interesting without negatives.
I really hope you like the model and use it.
If you like the model and think it's worth it, you can make a donation to my Patreon or Ko-fi.
I want to give a huge thanks to the people who helped me these past three months:
Mousewrites, PeePa, Kaz, Queria Star Morta, theovercomer8, Nawnie, Freon, Kohya
Follow me in my twitter to know before all about new models:
https://twitter.com/artificialguybr/
Update: I have authorized commercial use of the model in image generation services. You just have to explicitly credit me @artificialguybr (thanks if you link to my twitter account).
Description
FAQ
Comments (19)
Ok, but why is it censored? nsfw models for SD 2.1 practically do not exist
because internet crybabies forced people behind SD to censor it. Im pretty sure its just too much effort at that point to make uncensored SD 2.1 so stick with SD 1.5 for that, its not like 2.1 is that much better anyway.
@bitzupa I agree but I would like to see 2.1 uncensored models both real and anime, just out of curiosity
sd2.1 sucks balls. it always creates these weird elongated human faces and weird textures. its def a downgrade and f*ck the crybabies
Didnt work for me, get The size of tensor a (1024) must match the size of tensor b (768) at non-singleton dimension 1 error, even when change res to 768X1024 or 1024x1024, get same error, only have this problem with your model.
If you look at the console log when you generate it explains how to fix this.
there's a checkbox under Settings > Stable Diffusion
"Upcast cross attention layer to float32"
You need this if your card doesn't support half precision (ie; it is old)
I'll save you some time though, if you don't have enough VRAM to gen a 1024x1024 image this model will be worthless to you.
Check if you have controlnet enabled.
same here, can't generate anything. no controlnet, im on float32 and I have a 4090 + 96gb of ram.
Update - got it to work but cannot use any of my Loras or Lycoris...
@cyrilstyle its 2.1 model so yeah nothing from 1.5 will work unfortunately,
Thanks, got it working :-)
I've been running a few prompts testing the model and have to say it's very nice!! Finally a model that's not filled with anime and naked chicks! It generates great at 1024x1024. I ran some tests with multiple samples and steps ranging from 25 to 45 and 90% of the images are great! I'll post some samples later today. Nice work!!
I'd like to see what you've made!
Great work. Thanks for making it 2.1 and tagging me on reddit. I'm after photorealism only and I found a couple other models to be slightly better at eyes than this on average. They have bigger issues though. This model has very good lighting and coloring, great detail and style, great likeness with training. To help fix the eyes a little bit I ended up merging this as an ADD DIFFERENCE with 2 other models. Works fantastic.
If you're ever able to improve eyes in an update, I'd definitely go all in on it (I make 500+ images daily)
Doesn't work. outputs are black images.
Don't fully get this model. On Reddit it was explained how it would not need any negative embedding. All sample images now have a negative embedding + uses a lowered cfg.
The images were generated by users in a closed beta. You don't need in many images (some even look better without than with), but it's still an SD model. You can and should test the best configuration that works for you.
About the CFG; it's the same reason. These were images generated by other users that I just posted. I wasn't checking parameters of each images.
does this require a config?
great model, btw
On an M1 mac I get "RuntimeError: expected scalar type Float but found Half" // is there a way past this? I know M1 unideal, but this is the first checkpoint (of many) I've had this issue with.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.