THE MODEL IS DOWN FOR NOW!
I will wait and upload it again when everything is cleared with the licence. Thanks for testing and supporting this project.
In the meantime you can might take a look on my other checkpoints. Links are down below (Recources). In the "training data" you will find some images I made for my latest SDXL project as placeholder..
This is a very exprimental merge between my SDXL enhancer and SD3. I merged my own enhancer and a model I am working on at the moment (DEMONCORE 4 "NeoDEMON")
It somehow worked out...
Its based on the "clips included" SD3-model. The main work goes to Stability AI. I am just experimenting with their checkpoint. https://civarchive.com/models/497255?modelVersionId=568278
To get it working correctly hook the clip from the model directly into the conditioning..
First things first.. This is a research model.. I published this to get images from you as experiment. I am not using it for commercial use.
About Version 2:
This time I completely removed the SD3 text encoders and replaced them with my own SDXL text encoders from my SDXL checkpoint Colossus Project 10B https://civarchive.com/models/147720?modelVersionId=557262
This model is a hit or miss one.. in my opinion better than the standart SD3 model.
Some words about the showcase here: I normally do no more than five trys for any showcase images. So I did the same here.. Some images turned out great others funny and some are like burning trashcans. Scientific speaking: This model can't compete with the original Colossus SDXL model despide sharing the same text encoders... Try it yourself and compare it. I see this as funny expriment. We will laugh about this model when the fully trained are released.
HOW I MERGED SDXL CLIPS INTO SD3
https://civarchive.com/articles/5698/how-to-merge-sdxl-clips-into-sd3
Is it a good model?
To be honest at the moment SD3 is bad compared to SDXL. And I mean it. Its may good in text following but try to create a woman laying in grass.. Also you will have problems with hands, feet, faces of woman, oversaturation, coherence problems and so on.. In my tests this version boosts more details than the original version.
Well.. test it.. post some comparison images between the original for example.
This is nearly a base model so it won't be as good as fully trained SDXL checkpoints yet. In the meantime here is an alternative to SD3---- it works much much better and can create woman laying in grass.. https://civarchive.com/models/155977?modelVersionId=576344
Will a fully trained model will be better?
Yes and I will keep on working on this and do a proper training. Again this is just a early test. see it as it is... I didn't even get Forge running with it yet.. So well I have to dig into it more... (like we all)
Have fun creating!
Licence:
This is a non commercial model. Please read the licence here:
In short.. Don't use this to sell images. Only for fun and science!
https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE.md
If you want use SD3 commercial go to OpenAI and use the original model.
Credits:
Thanks to the guys from Stability AI
https://civarchive.com/models/497255/stable-diffusion-3-sd3
UPDATE:
A small update: I will NOT start training SD3 further.. My reason for this is the mess of licencing.. just read the fineprint of their licence! (Or watch a video about it)
- Creators can hold responisble for the images users are creating for example.
- Creators have to destroy their works after the payments end. Including already trained checkpoints and loras
- Creators must pay or also the costomers have to stop using their products.
- Creators can only create 6000 images/ month
This is nothing else as a anual fee of 240 dollar for every creator.
I may also delete this model soon. Use SDXL instead. Don't do even train on SD3 images..
I may leave it up for a couple of days in the hope that the licence will change..
Thank you for testing it.
Oh, out of rage I might release a SDXL SD3 killer.. I held it back the last days because of SD3.
Best regards Afroman4peace..
UPDATE 2:
I've been in contact with some of the devs.. seems like getting some donations over kofi doesn't or getting buzz is not commercial use for them. At least what was told me... That means I will keep working on this as side project. However my main focus will stay SDXL. Anything which I make money like on tensor .art will be stay SDXL. This includes creating images for professional use... That means that I will continue working on this.. will low priority. I made up my mind and will not delete this checkpoint here...
I still hope that they will change their licence.. Just be aware just in case.. Really don't use anything in here commercial. Don't use the image results for training your models. Be aware that this licence is a strict one... and if Stability AI asks to delete my opinion.. I will delete anything here with it. So don't rage if this model will disappear.
And I will this release the SDXL checkpoint mentioned above. (the image in the gallery is from it). Not because if rage this time. I want just provide a open more working alternative to SD3.
Best regards Afroman4peace
Update 3:
Here is the checkpoint I meant. Have fun with it
Description
This is a test with a complete SDXL-Clip. I chose my latest model Colossus Project XL 10B for it. To get it working you have to use the clip from this model. Not the clip loader. (see in the description
FAQ
Comments (18)
how did you manage to merge two models from different architectures?
I will add comparison images next.. but it doesen't work in Forge for me yet so bare with bad images. I also need to test a lot
gives exact same results as the base sd3
First SD3 community fine-tune in the world LET'S GO
First of many, historic moment.
You definitively need to explain how you did that in detail.
真给力
So basically you merged their CLIP models.
Most interesting indeed. Perhaps this will be easier than expected.
Mind that having two fine-tuned text encoders (moreover tuned on another models) inside SD3 could fix most of the issues of this new base model
For anyone asking how to merge SD3 clips with SDXL ones: I made a simple guide here: https://civitai.com/articles/5698
I am testing and working on SD3 since the first second it came out. I was and I am still thrilled about it. But I have to tell you my thoughts about it: Please this is just a base version but here are my observations for now. The first image I get is kinda ok.. (My SDXL checkpoints for example Colossus Project XL 10B are still better). But when I try to upscale it the image gets worse with fragments. Even the prompt following was sometimes worse than SDXL. Especially for women and dynamic positions. Getting SD3 to work now it should will getting a challenge. To be honest I'm very disappointed right now. I know it's a base version but still.. the censorship makes everything worse... Yet alone the licence policy. Even paying 20 dollars per month is too much for right now. I still will try my best to get it working but the motivation is not high here right now. The masochist inside me is screaming to try training it. May it has some potential..
My main questions at the moment are:
How to get rid off the annoying censorship?
Am I even allowed to get rid off the censorship?
Is it worth paying 20 Euro per month to work on a model that is heavily censored?
Is it worth my time to train it?
@stability ai.. If you read this may answer these questions. I know that you worked hard on this and I appreciate it. I don't want to bash your work here..
The only thing I hate the most is the censorship which broke the model in my opinion. SDXL is far better at the moment.
Tell me what you think about it..
But waht about t5xxl clip? It is one of the main features of SD3
It's very cool to see people experimenting with it instead of just complaining. Do you think this merge can be trained? I would like to give it a try
Well, it looks that its time for the NOT NAKED mummy nun rises again for another experiment.
A small update: I will NOT start training SD3 further.. My reason for this is the mess of licencing.. just read the fineprint of their licence! (Or watch a video about it)
For Commercial use..
- Creators can hold responisble for the images users are creating for example.
- Creators have to destroy their works after the payments end. Including already trained checkpoints and loras
- Creators must pay or also the costomers have to stop using their products.
- Creators can only create 6000 images/ month
This is nothing else as a anual fee of 240 dollar for every creator using SD3 remotely commercial
I may also delete this model soon. Use SDXL instead. Don't do even train on SD3 images..
I may leave it up for a couple of days in the hope that the licence will change..
Thank you for testing it.
Oh, out of rage I might release a SDXL SD3 killer.. I held it back the last days because of SD3.
Best regards Afroman4peace..
Are the creators of SD3 seriously trying to force how it is allowed to be used. This is the internet.
Our new Community License is now free for research, non-commercial, and commercial use. You only need a paid Enterprise license if your yearly revenues exceed USD$1M and you use Stability AI models in commercial products or services.