Midjourney Style Reference Case Study

Midjourney dropped a new feature call Style Reference. I decided to test it and compare the results using an image prompt and a style reference, while making changes to the prompt and the parameters.

Observations and results below.

I started with an image promote of a yellow origami boat and a style reference of a blond woman in saturated neon type colors.

The prompts for these two images:

yellow origami boat on a sea of computer code –ar 16:9 –v 6

woman with long blond hair, saturated pink, purple, and black, anime style –ar 16:9 –v 6

Observations: – “yellow” is a strong influence from the boat image.

Run 2:

Prompt elements from both images pink origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6 –sref [url]

Observations: – boats in the images lost some crispness – due to added words in the prompt? (that is the only change).

Run 3:

non-specific boat color. I removed the “pink” color for the boat. origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6 –sref [url].

Observations:

• Sharpness is back on the boat
• interesting

Do we have a sinking ship

Run 4:

Slight prompt adjustment.

Added: anime style (from style ref. prompt) origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors, anime style –ar 16:9 –v 6 –sref [url].

Observations:

• slight changes in the boat, geometric.

Run 5:

Midjourney Stylize parameter tests.

With simplified boat prompt.

origami boat in a sea of computer code –ar 16:9 –v 6 –sref [url] –s [add value]

Below:
–s 50
–s 250

Observations:

• no significant differences

–s 500
–s 750
–s 1000

Observations:
– quite difficult to see significant differences, may be slightly sharper going to higher stylize value

On occasion I still got “fuzzy boats” too but less than in the lower stylize values.

Observations:

• all boat are crisp and sharp (occasional loss of sharpness using a style reference?)

Run 6: no style reference, yes image prompt.

Combined text prompts from boat and girl (the colors).

origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6.

I wanted to see if/how much the style reference affects the image.

Observations:

• all boat are crisp and sharp (occasional loss of sharpness using a style reference?)

Run 7: same as run 6 with addition of anime style.

origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6

Observations:

• no significant differences
• some outputs included a vertical lights type background (see 1st image below)

Run 8: Tests with style weights –sw.

This time I want to see what difference if any, adding style weights does.

All previous runs has used the default –sw of 100

origami boat in a sea of computer code –ar 16:9 –sref [url] –sw [value] –v 6

This batch:
–sw 0

Observations:

• wow – sharp and crips, but! style weight is non-existing – meaning do I then even need to use it?

tyle weights continued…

same prompts as above

–sw 250
–sw 500

Observations:
– boats look fuzzier in the about 50%+ outputs

–sw 750
–sw 1000

Observations:
– more fuzzy boats, otherwise crisp
– no significant differences

Just combined text prompts of the two images. Left out the color of the boat. origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6

Run 10: no style reference, no image prompt

Added “yellow” to the prompt

yellow origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors –ar 16:9 –v 6

Observations:

• another wow (totally subjective haha)

no reference, no image prompt

added “anime style”

yellow origami boat in a sea of computer code, vibrant, saturated pink, purple and black colors, anime style –ar 16:9 –v 6

Summary

Ok I could keep going on using image weights etc., but that’s for another time.

Let’s wrap it up.

Notes:

1. Adding stylize or style weights didn’t make a large, significant difference
2. Higher style weight values increased “fuzzy boat” outputs, about 50% of the time.
3. Zero style weight resulted in crisp image that look very similar to original image reference.
4. Using full prompts, vs. shortened prompts didn’t result in significant changes.
5. Similar outputs are possible even without using the style reference and/or the image references.
6. Adding stylize did not result in significantly different image outputs.

So what should you do? It depends. Do you have a destination? Are you just playing around?

Experiment. Observe. Evaluate.