Fooocus FreeU
FreeU is another possibility for image variation. FreeU is not really an open university but instead its full title is "FreeU: Free Lunch in Diffusion U-Net". Like the Fooocus Vary function, FreeU operates at the denoising level of Stable Diffusion, as detailed here.
As shown in the image above, there are two pairs of settings for FreeU. Backbone Stages 1 and 2 (B1 and B2) adjust composition and colour while Skip Connections Stages 1 and 2 (S1 and S2) adjust texture and detail. There are no settings that are best for all images, and choosing the best - or using FreeU at all - is often a matter of taste. However, selection is made somewhat easier by a set of eight presets selected from the dropdown menu at the top left of the interface.
The settings for any one of these presets can be adjusted manually with the sliders, but if you wish to return to the default values the Restore Preset Values button at the top right will do so.
Most importantly, ensure that the Enabled checkbox is activated if you wish to use FreeU. The FreeU feature will apply to text-to-image (normal generation mode) as well as to all the Input Image functions.
The first set of test images use the following settings:
- Prompt: a Congolese woman walks away towards a mystic mountain portal on a Dark and Stormy Night. Sharp background, deep depth of field
- Negative Prompt: blurry background
- Styles: Fooocus V2, Fooocus Semi Realistic
- Substyle: Spellcasting Druid
- LoRA: blur_control_xl_v1.safetensors, strength 1
- Preset: Magick
The first image shows the result without using FreeU, then the results for each FreeU preset are shown in alphabetical order:
All of the images satisfied the requirements of the very basic prompt, "a Congolese woman walks away towards a mystic mountain portal on a Dark and Stormy Night", but there were quite different levels of fantasy and detail between the images.
In the non-FreeU image there was just the hint of a castle at the mountain peak. All FreeU images expanded on the castle theme with F: Original Fooocus - Subtle providing a nicely proportional building and a more convincing storm. Landscape work is typically not a strength for B: Realistic Portraits but if did produce an interesting tower. And C: Minimal produced a surprisingly detailed castle built into the mountain that went right down to its base.
Next, D: FooocusPlus Default - Realism and G: GreyScope Default both produced well formed archways for the portal interpretation, as appropriate for their typically realistic bias.
A: Automatic1111 Default and E: Creator's Default - Fantasy have gone more strongly into the fantasy realm, where the portal is a great illuminated gash in the castle/mountain.
And lastly, H: SD1.5 Default, which is really designed for Stable Diffusion 1.5 base models, is probably the winner in terms of detail with a fully formed castle and a well built portal archway.
The overall trend for this set is that all the FreeU processed images provided a higher level of detail than the non-FreeU rendering. The choice between any of them - including the original - is a matter of creative expression.
This set of images use the following settings:
- Prompt: Full body portrait: Faroese woman, gentle smile, (sunlight:1.4), clearly defined facial features, perfect hands. Sharp background, deep depth of field
- Negative Prompt: blurry background
- Styles: Fooocus V2, Fooocus Enhance
- Substyle: Unlit
- Preset: Default
These tests give more opportunity to evaluate the accuracy of prompt interpretation. As before, the first image shows the result without using FreeU, then the results for each FreeU preset are shown in alphabetical order:
The prompt for these portraits was more specific than for the fantasy landscapes: "Full body portrait: Faroese woman, gentle smile, (sunlight:1.4), clearly defined facial features, perfect hands. Sharp background, deep depth of field". This provides an opportunity to measure prompt fidelity when using FreeU.
First, unlike almost all the FreeU examples, the non-FreeU image did not depict the "perfect hands" described in the prompt. Also the sun rays seemed to be emerging from inside a rocky bluff and the ocean lacked detail.
A: Automatic1111 Default had nicely formed hands with convincing sun rays and a more details ocean.
B: Realistic Portraits had a nicely detailed face and good hands, but the ocean on the right side was a bit anemic.
C: Minimal had good hands and the interesting addition of some white birds, but I found the sun somewhat bland.
D: FooocusPlus Default - Realism had good hands and more detailed birds but, like the non-FreeU image, the sun rays were emerging from the cliff.
E: Creator's Default - Fantasy is perhaps my favourite of the set, with good hands, dramatic sun rays, ocean detail and an attractive use of colour in the grassy background.
F: Original Fooocus - Subtle had the faults of the non-FreeU image: no hands, a mushy ocean and sun rays emerging from the cliffside.
G: GreyScope Default had good hands and it was the only one with sheep! The sun however was lacking in impact.
And lastly, H: SD1.5 Default had distorted hands and I found the background rather bland.
In conclusion, I felt that A: Automatic1111 Default, B: Realistic Portraits and E: Creator's Default - Fantasy were an improvement over the non-FreeU version, for prompt following, enhanced detail and creative impact.
I found that Rodney's FreeU video at the Kleebz Tech AI YouTube site was a great resource for learning about FreeU. The settings for the B: Realistic Portraits, C: Minimal and D: FooocusPlus Default - Realism presets all came from there. The G: GreyScope Default settings are sourced from Greyscope's article at reddit.
When watching the video, top left = C: Minimal, top right = B: Realistic Portraits, bottom left = E: Creator's Default - Fantasy and bottom right = D: FooocusPlus Default - Realism.
To close this article, I added an Outpainted version of my favourite image from the Fantasy Landscape set, H: SD1.5 Default. This image was extended to the Bottom by 78.8% to make it square, with FreeU H: SD1.5 Default active.
