Short Experiment: How StableDiffusion Paints Hatsune Miku

作者コメント：

I was wondering about the best number of steps for StableDiffusion AI image generator to explore through random seeds and create the finest image as the filnal result.
In this video I used "Euler" as the sampling method and I set the random subseed strength to 0.1.
Looking at the result, what I feel is this.
- In step 1 the iteration starts from noise and it takes about 20 steps to fix the basic structure of the image. When trying various prompts with some tens of images, and when doing a rough exploration with some hundreds of images after fixing the prompt, I think 20-30 is enough for the number of steps.
- It takes until step 30-50 to converge to a fairly good result. After selecting some good seeds from the result of rough exploration, I would generate those images again with 50 steps to narrow down further. Generating some variations by using random subseeds with a small weight value to find an image with better details may be a good idea.
- After step 50, and especially after step 100, it looks like the image starts to degrade. To me the result of step 50 is a much better result compared to step 400. Thus "larger number of steps" doesn't always mean "better results".
In this case I felt step 44 (which is the thumbnail of this video) gave an overall best result.
So what I do is I find a good text prompt first, do a rough exploration through random seeds, select some good seeds and generate from the same seeds again with steps through 30-100, and finally combine the results from different steps manually using an image editor to get the best looking result.