Styling with Fooocus

Styling with Fooocus


Alright, folks! Let’s dive into the exciting world of image generators. Today, we’re testing Fooocus - a software that creates beautiful images using various models and prompts. We want to see how simple it is for users and how creative its output can be. To make things reproducible and comparable, I’ll be using the same seed (123456) and resolution (1024x1024) and extreme fast setting (except in one case) for every generation.

Now, buckle up as we take a black box testing approach - enabling different styles one by one to generate a wide array of distinctive outputs. This way, we’ll have plenty of examples to share and discuss! So grab your coffee, sit back, and let the magic unfold before your eyes!

Lets compare

In this next section, we will be comparing various scenarios to see how noticeable the differences are. By examining these distinct setups, I aim to provide you with an understanding of Fooocus’s capabilities and how it handles different prompts and models. So, let’s dive into the first scenario!

Version difference

We start our journey by comparing the performance of the juggernaut_v7 model against the v6 version (which is the default model in the software right now).
My expectations were high, as I anticipated that v7 should bring a noticeable improvement in image quality and finer details.
It wasn’t deliver too many improvements. In most cases, the upgrade did not result in better visuals, some noticeable skin tone changes to make it photo realistic but mostly those.
Prompt: redhead pirate girl

Mk Encaustic Paint
Mk Bauhaus Style
Photo Neon Noir

However, there were some styles where the difference between the two versions was hardly noticeable. But that doesn’t mean v7 isn’t worth trying out – it’s simply a testament to how well-optimized its predecessor was!

Mk Blacklight PaintMk_Blacklight_PaintMk_Blacklight_Paint
Mk Patachitra PaintingMk_Patachitra_PaintingMk_Patachitra_Painting

Link to v6 Gallery
Link to v7 Gallery

Speed vs Quality

In another easy comparison, I set Fooocus to generate images with the quality setting after trying the extreme fast setting.
No surprises here either – the images with the quality settings are absolutely stunning! However, it’s worth mentioning that Fooocus does have some default settings in place when changing to quality, that can be undone or added for fast setting, but it means that it’s not just a sampling step change.
But hey, who doesn’t love a little extra eye candy while we wait?
Prompt: redhead pirate girl

StyleV7 Extreme fastv7 Quality
Photo Hdr
MRE Undead Art
Misc Grunge

Link to v7 Extreme fast Gallery (same as in the previous paragraph)
Link to v7 Quality Gallery

V2 Style bias

Here’s the comparison I was really looking forward to – how Fooocus V2 style expansion affects the overall output.
This dynamic feature is a game-changer, allowing us to combine various styles more effectively!
However, it does come with its own set of biases. In some cases enabling Fooocus V2 seemed to tone down the original styles and steer images towards a more generic professional look.
This was quite noticeable when using very specific styles such as Quilling, Kirigami, or one-line art. Prompt: dancing dog in bow tie

StyleWith Fooocus V2Plain style
Papercraft Paper QuillingPapercraft_Paper_QuillingPapercraft_Paper_Quilling_no_fv2
Papercraft KirigamiPapercraft_KirigamiPapercraft_Kirigami_no_fv2
Mk One Line ArtMk_One_Line_ArtMk_One_Line_Art_no_fv2

A funny observation I made while testing this feature is that when Fooocus V2 is enabled,
the “unreal” part of the prompt seemed to be overshadowed by the bias completely.
Altho it’s mostly not present with the plain models either it’s more prevalent with the V2 on, but it can be solved with more fine-tuning on the prompt
For example, a prompt like “dog dancing” would lose that playful aspect and instead focus solely on the close up photo elements. It’s an interesting phenomenon to explore!

StyleWith Fooocus V2Plain style
SAI Comic BookSAI_Comic_BookSAI_Comic_Book_no_fv2
SAI AnimeSAI_AnimeSAI_Anime_no_fv2

Link to Fooocus v2 on Gallery
Link to Fooocus v2 off Gallery

Be Creative

Now that we’ve explored various scenarios and comparisons, let’s see how creative Fooocus can be!
Note: this test is more about the capabilities of the underlying model rather than the software itself. However, since inner optimizations are still in play, it presents an excellent opportunity to push this tool to its limits.
Prompt1: busy city street in the 1600’s
Prompt2: colorful scary toy lama creative
Prompt3: poster of a new marvel movie about a lizard wizard, creative

StyleHistoric streetToy lamaLizard Wizard
Mk Tlingit Art
Mk Ukiyo E

As mentioned earlier, one interesting observation was that the enabled Fooocus V2 (by default) seemed to have a strong bias towards 3D/photorealistic representations of the main subject.
This creative challenge allowed to test the model’s flexibility and adaptability.

Link to Historic street Gallery
Link to Toy lama Gallery
Link to Lizard WizardGallery


In conclusion, the Fooocus V2 style expansion has a noticeable impact on the overall look and feel of images generated by this software. While it may not be perfect for every scenario, having a more realistic appearance can be advantageous in certain cases. Of course, our comparisons haven’t touched upon proper prompt engineering or experimentation with different LoRA models – something that could help break free from the inherent biases.

But what truly stands out is how easy Fooocus is to set up and use while still delivering impressive results. It’s heartwarming to see such high-quality open-source tools thrive in an environment where countless paid AI options emerge daily.

© 2024 Bence Stumpf