Let's make it short: You can't with the normal version but you can with a special branch of the software.
Do you want to know more ?
Prompt for the picture above:
Soldiers fighting using conventional weapons to fight insectoid aliens in the desert.
Steps: 24, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 3.5, Seed: 2945699879,
Size: 1024x1024, Model hash: 92db4295e9, Model: sd3_medium_incl_clips_t5xxlfp8, Version: 1.9.4
(the goal of this picture is to optimize cfg and steps for DPM++ 2M)
- Go to the github of Stable Diffusion and once there, make sure that the "sd3" branch of stable diffusion is selected.
If you click on this link
You should be directly there.
- Click on code and select Download.
Then unzip to the directory of your choice.
- I will suppose here that you already have all the requirements that are needed to install a normal version
of stable diffusion (like Python). Read the docs if you are not sure.
So open the directory where you unzipped the project and double click on:
webui-user.bat
This create the virtual environnement.
Wait that the installation is finished.
At the end, a browser window should open.
- Download the models that you need to use the software:
Go here:
https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main
It is one of the following files:
sd3_medium.safetensors
sd3_medium_incl_clips.safetensors
sd3_medium_incl_clips_t5xxlfp16.safetensors
sd3_medium_incl_clips_t5xxlfp8.safetensors
(one is enough, it is your choice)
You can put them in this directory:
stable-diffusion-webui-sd3\models\Stable-diffusion
In the latest release of this branch of Automatic1111, there is already a CLIP directory here:
stable-diffusion-webui-sd3\models\CLIP
With these files inside:
clip_g.safetensors
clip_l.safetensors
(The files are downloaded by the software when you load the model).
In case this doesn't happen, create the directory and put these files yourself in this directory.
The files can be downloaded here.
Refresh the model's list and now you can generate pictures with a Stable Diffusion 3 model thanks to Automatic1111.
The Euler model seems to work the best for me.
LMS works decently.
Heun is good.
The results with "DPM++ 2M" tend to be more "artistic" (at least as long as the prompt is not too long, if not the results are garbage too). It is pleasant to see. The picture at the beginning has been generated with this sampler.
I hope this tutorial saved your time.
(Yes, i am aware of the controversies with Stable Diffusion 3).