How to generate the highest quality AI images with Fooocus

• GPU Requirements: Ensure you have a powerful GPU (NVIDIA recommended) for faster rendering. A minimum of 8GB VRAM is suggested for optimal performance.
• Pre-trained Models: Download the relevant pre-trained models (e.g., Stable Diffusion variants) for your preferred art style or application.
• Prompt Mastery: Learn to write detailed and descriptive prompts to guide the model for precise image generation.
• Optional AI Upscaling: For better final image quality, use AI upscaling tools like ESRGAN after generation.

These basics will help you set up Fooocus for high-quality AI image generation.

To start, the first thing to know about Fooocus is its commitment to abstracting away many of the complicated settings required to make high quality generated images. They outline this in their tech list on the github page, but let’s go through each of these improvements here.

• Style—The Fooocus V2 style system allows for MidJourney style prompt expansion to alter and improve the text representation using GPT-2 at low cost. This prompt expansion allows simplified inputs to generate far more varied and complex outputs, and the motifs can be mixed and matched to the user’s preference.
• Native refiner swapping inside one single k-sampler—This allows for seamless carryover of the base model’s parameters to the refiner by maintaining the sampler. It is also used in the AUTOMATIC1111 Web UI.
• Fooocus compensates for the lack of cross-attention in high-resolution Stable Diffusion XL outputs by adjusting positive and negative signals. This modification addresses the resulting reduction in classifier-free guidance contrast at these higher resolutions.
• Self-Attention Guidance—Following the research shown here, the Fooocus development team has implemented a form of Self-Attention Guidance at a low setting. Along with the negative ADM guidance, this helps mitigate a smoothing effect commonly reported in image outputs with SD XL.
• Automatic implementation of the “sd_xl_offset_example-lora_1.0.safetensors” LoRA model at a strength of .1. Their experiments determined that values below .5 were always better than those without.
• Sampler parameters are finely adjusted and optimized by the developers.
• Implemented a hard resolution setting based on the optimal aspect ratios for image synthesis with SDXL.

These cumulative features make for an extremely low level of coding knowledge to run the image generation after setup. Now that we have discussed what makes the platform so nice to use, let’s walk through setting up the UI before comparing its pros and cons.

Setup

To run the demo, we first need to install conda.

In the terminal, we are going to begin pasting in everything needed to run the notebook. Once we are done, we can just click on the shared Gradio link to get started generating images.

First, paste the following into the terminal:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Follow the instructions in the terminal to complete the installation by answering yes to each question when prompted. When that’s complete, close the terminal using the trash bin icon in the terminal window on the left, and then open a new one. This will complete the Miniconda installation.

Afterwards, we need to paste in another set of installs to complete the setup:

conda env create -f environment.yaml
conda activate fooocus
pip install -r requirements_versions.txt

These will install everything needed to run Fooocus in the Notebook. Additionally, it’s worth mentioning that the application will download a Stable Diffusion XL model from HuggingFace for us to use on launch. When we paste in the final snippet, we will see that occur first before the application itself is launched. Be sure to use the public link so that we can access the link from our local machine’s browser. The process to complete the launch and install may take a couple of minutes.

python entry_with_update.py --listen --share

Now, we can begin synthesizing new images. This will share many similarities with familiar Stable Diffusion and MidJourney pipelines, but have some obvious differences in implementation, some of which we covered above.

The first thing we want to do is show a quick test of the basic image generation using the base settings. To do this, all we need to do is enter a prompt and hit ‘Generate’. Fooocus will have automatically downloaded a merged model, namely “juggernautXL_version6Rundiffusion.safetensors”, that is capable of handling a wide variety of both realistic and artistic medium outputs. Above, we can see an example generation using the Fooocus web ui’s basic input. The default settings will generate two images at a resolution of 1152x896 (a 9:7 ratio), and we can watch the diffusion process occur in real-time.

A comparison of three prompts across the different performance values.

From here, we can begin to look at different advanced settings for the Fooocus by clicking the toggle at the bottom of the screen. The advanced settings will be displayed on the right, in convenient Gradio tabs. The first, and probably most important tab, is the ‘Settings’. Here, we can see one of the first things abstracted away from a typical pipeline: the Performance settings. These are pre-set and optimized to run a different number of diffusion steps to achieve different qualities of images at different speeds. Speed has 30 steps, Quality has 60, and Extreme Speed will run for 8. While the application seems to have the capability of using other K samplers, it seems to be hard defaulting to use “dpmpp_2m_sde_gpu” for all performances.

Next, we have the resolutions. These are all optimized resolutions to run Stable Diffusion XL models in particular. They included this hard restriction on image sizing to further optimize users’ outputs; images generated at suboptimal resolutions are far more likely to appear strange. Above is an example generated with a 704x1408 resolution.

Finally, we have the negative prompt. This functions as a sort of opposite to our regular prompt, and any tokens we include will be discounted as much as possible by the model. For example, we may use a negative prompt to remove unwanted traits from our generated images or try to mitigate some of the inborn problems Stable Diffusion has with certain objects. In the example above, we used the same seed as we had in the previous generation, but added a negative prompt to try and alter our output a bit. You can assign the seed by filling the field left by un-toggling the Random box below the negative prompt field.

In the next tab in the advanced options, there is “Style”. These are the GPT2 enhancements provided by the model to expand our prompts. Try out different styles to see what sort of effect they can have on the final outputs. We particularly recommend using the ones above for all generations.

The next tab is the model tab, and it is probably the most important of them all because it lets us switch out the main checkpoint and LoRAs. It has nice sliders that allow us to easily adjust the weights of any additional models, allowing for a more simplistic way to blend the traits of two LoRAs. The application will also automatically download and allocate a weight of .1 to the ‘sd_xl_offset_example-lora’, which we can choose to remove if we so choose. If we download a model during a running session, the refresh all files button at the bottom will allow us to refresh the available model lists.

The final tab, advanced, has our Guidance Scale and Image Sharpness sliders. The guidance scale controls the impact the prompt has on our final input, and works best in a range of 4-10 in our experience. The image sharpness value will sharpen images, but increase the uncanny valley effect if raised too high. We suggest leaving it alone unless adjustment is particularly needed.

“a scenic Texas sunset”

In addition to the advanced options toggle, there is also the input image toggle right beside it. This is how we can do Image-to-Image generation in Fooocus. Illyasviel, the creator of this project, was also the creator of the popular ControlNet. They have combined Image-to-Image with a robust ControlNet system to automatically use ControlNet to guide inputted generations. It comes with three options: Upscale or Variation, Image Prompt, and Inpaint or Outpaint. Let’s look at what manipulations we can make to the scene above using this prompt:

New prompt: “a scenic sunset in Hawaii”

• Upscale or Variation: there are 5 settings: vary (subtle), vary (strong), upscale (1.5x), upscale (2x), and upscale (fast 2x). We can use the first two to introduce additional variation to our image with our prompts, as shown above with a subtle variation. The strong variation will make even more changes, and the upscale options will make our images higher resolution

• Image Prompt: This is comparable to Image-to-Image modeling with Stable diffusion, but instead of a single image, we can input multiple images. This allows us to vary our outputs using the inputted prompt. Check out the example above where we added a tree to our output using a different tree as input to influence the final result.

• Inpaint or Outpaint: inpainting allows us to add images to a masked region of the inputted image, and outpainting lets us functionally extend the edges of an image to predict what would be outside the “camera” aperture. We can use it to add new things to our images, like adding this lifeguard stand.

We can even use all three together to great effect! Be sure to test out all of these to see how they fit into your workflow, as using existing images as bases or editing them to work can be much more efficient and effective than generating wholly new images.

We’ve explored what makes Fooocus a standout tool for running Stable Diffusion, but how does it compare to other popular options? With alternatives like AUTOMATIC1111’s Web UI, ComfyUI, MidJourney, PixArt Alpha, and DALL·E 3, it’s essential to understand where Fooocus carves out its niche. Designed by Illyasviel, the developer behind ControlNet, Fooocus targets low-code users who want a streamlined image generation experience—without the complexity of systems like A1111 or ComfyUI.

Further, Fooocus offers greater flexibility and local control, while remaining more accessible than traditional UIs. That makes it ideal for newcomers to AI image generation or coding. For advanced users who need more customization, we still recommend exploring fast Stable Diffusion implementations via AUTOMATIC1111’s Web UI or ComfyUI on DigitalOcean Droplets.

Thanks for reading! If you’re ready to get started, don’t forget to explore our DigitalOcean GPU Droplets for running Fooocus and other Stable Diffusion tools with ease—and check out our other guides on AI, ML, and image generation workflows.

References

• Getting Started with Stable Diffusion on DigitalOcean
  A beginner-friendly guide to setting up and running Stable Diffusion using GPU Droplets.

• FLUX.1 Text-to-Image Generation on Cloud GPUs
  Walkthrough for experienced users looking to run FLUX using cloud GPUs.

• Stable Diffusion 3.5 Large with DigitalOcean GPU Droplets
  A setup guide focused on using Stable Diffusion 3.5 Large workflows with GPU Droplets.

• Training a LoRA model for Stable Diffusion XL
  Explains how to train LoRA models for SDXL for custom image generation tasks.

• TPU vs GPU: Choosing the Right Hardware for Your AI Projects
  Covers the different options available and how to select the right one for your AI Projects.