Swirly Bokeh

Makes the depth of field "bokeh" take on a swirly pattern. This effect was perhaps most notably seen with certain Helios lenses. However, the Voigtlander 35mm f1.4 MC was another with similar characteristics. Then the newer TTArtisan 75mm F1.5 lens was made specifically to reproduce this effect, in a perhaps more dramatic way.

You can adjust the model strength of this LoRA to get a range of swirl to your bokeh.

Suggested model strength: 0.5-0.65

Model strength workable range: -1.0 - 0.85

Suggested samplers/schedulers:

• seeds_3 / normal (or beta)

• dpm_2 / ddim_uniform (or normal)

• dpm_2_ancestral / normal

I've included two files in this first release (two versions since Civitai won't allow multiple files in a version). One that was rank 64 at 1,000 steps and the other was rank 32 at 2,800 steps. What's the difference?? Aside from file size? Well, the rank 64 has more contrast, punch, and holds on to finer details better. The rank 32 can lose some of the finer details, but I felt like it was more forgiving to the sampler/schedule. It seemed to do better with a greater range of combinations. If you're using the euler or euler ancestral sampler, I'd probably suggest the rank 32 version. Note that while I did not put euler on my list of suggested samplers/schedulers, euler and euler_ancestral do work.

However, the rank 32 version has this "shrinking" effect. It seems to make subjects smaller, or at a distance. They aren't unrealistic in terms of size. They are just farther away, unless your prompt has stronger language like "close up" in it.

Both start to really cap out at 0.85 model strength. You can get some results at 1.0, but it will really be hit or miss.

I may train another version yet since I was still a bit undecided. Similar to my experiences training a glitch style LoRA, Z-Image Turbo likes to fight against styles that distort images.

Give both a try, share your results, and let me know what you think! It'll really help guide me for the next version.