• Ratings
• Pros/cons
• Prerequisites
• Description
• Installation guide
• Further details on model parameters
  • Conditioning timestep start/end
  • Model sampling shift
  • Steps
  • CFG
  • Denoise
  • Sampler name and scheduler
• Alternative model versions

Accuracy / Quality ★★★★☆ 

Flexibility / Features ★★★☆☆ 

Data security / Privacy ★★★★☆ 

Before this model can be used the following prerequisites have to be met: 

1. The device used should have access to a GPU (either from NVIDIA or AMD) and a VRAM of at least 4GB, though 8GB is recommended.  
   To find out how much VRAM you have available, open the ‘dxdiag’ program on your device, navigate to the ‘Display’ tab and look for the value next to ‘Display Memory (VRAM)’. This should be at least 4000 MB or 8000 MB for this guide. 
2. Have Python 3.7 or a more recent version installed on your device. 
   If you don’t have Python installed yet, you can do so via the WUR Software Center or via the Python website.  

Next to generating text the most well-known application of GenAI is the generation of images. Doing so via online tools often subjects you to severe restrictions in format or frequency of use. It is also more difficult to document the use of such online tools as their availability is not guaranteed over time, and thus the reproducibility of your work may be in question. An alternative method is therefore to use a locally installed image generation model. The most well-known of these models are the Stable Diffusion models developed by Stability.AI.  

For this guide we will use Stable Diffusion 3 as example. However, alternative versions of this model are available and are listed at the end of this guide. These alternatives may require different workflows and supporting programs, which can often be downloaded from the model repositories on the HuggingFace website.  

1. Go to the Stable Diffusion 3 Medium (SD3 Medium) page on HuggingFace. 
1. Log in and sign in with your HuggingFace account (or create a new one if needed). Once logged in, scroll down and click the ‘Accept’ button to agree to the licence terms. The usage of this model is allowed for research and educational purposes and limited commercial use. 
2. Go to the ‘Files and Versions’ tab on the SD3 Medium page and scroll down to the file ‘sd3_medium.safetensors’. Download this file, as well as the following files in the ‘text_encoders’ folder. 
   • clip_l.safetensors 
   • clip_g.safetensors 
   • t5xxl_fp8_e4m3fn.safetensors 

You may also want to download the example workflow to make your initial use of the model easier. These are found in the ‘comfy_example_workflows’ folder. The ‘sd3_medium_example_workflow_basic.json’ file will be used for the remainder of this guide. 

7. Install ComfyUI from the ComfyUI repository on GitHub. 
8. Move the downloaded model from step 2 to the folder ‘ComfyUI_windows_portable\ComfyUI\models\checkpoints’. Also move the ‘text encoders’ downloaded to the folder ‘C:\ComfyUI_windows_portable\ComfyUI\models\clip’. 
9. Launch ComfyUI. Launching can be done by opening ‘run_nvidia_gpu.bat’ for NVIDIA devices, for example. 
10. On the bottom right in ComfyUI click on ‘Load’ and open the example workflow downloaded in step 2. 
11. Once the workflow is open, go to the top-left block called ‘Load Models’ and select the appropriate checkpoint (the downloaded model) and clips using the arrows. 

8. In the block below this called ‘Input’ you can enter the positive prompt (what you do want to see in the image) and a negative prompt (what you don’t want to see in the image). Also available is a setting for the width and length (in pixels, should be in multiples of 64) and the ‘batch_size’, which defines how many different images the model should generate (using the ‘Seed’ number indicated at the top, adding ‘1’ to the seed for each next images required to satisfy the ‘batch_size’ setting.

9. Once all these settings are in place you can click on the ‘Queue Prompt’ button. The image will then appear on the far-right side of the workflow in the ‘Output’ window. You can track the progress throughout the workflow (the active section is highlighted, and the generation process is indicated with a loading bar in the ‘KSampler’ box next to the Output).

In the Stable Diffusion 3 Medium basic workflow there are several other parameters that may be modified to change the model behaviour.

This guide used Stable Diffusion 3 Medium as example, but other models are available. These may require different workflows, hardware specifications, or understanding of the methodology.