Prevent burns
The following images are created by adjusting the generation parameters.
Burning can be avoided by adjusting parameters. At present, we believe that although burns may occur more easily, they are not fatal. If the original image is slightly burnt, it will be burnt more, and this can be covered by adjusting the output parameters. In some cases, it may be better to adjust using Lora's weight.
Reference model for SD1.5
In the first test, we used SD7 as a base, but this may not be the best one. In theory, it would be best if there was a model that was similar to all other models.
Perhaps the official model can be used as the standard model for SD2.0 and SDXL (something to be confirmed). However, this is absolutely impossible with SD1.5.
SD1.5 requires a suitable reference model other than the official model.
It may be difficult to decide whether it is better to set several reference models for each genre in order to increase accuracy, or to create one reference model that mixes models from various genres.
Personally, I think it would be better to create one reference model. Although the designs were quite different between the SD7 and KA2, we were able to achieve a good reproduction. Maybe you don't need to be so particular about similar patterns. It would be much easier to handle if there was only one standard model.
As a provisional standard model, I created an equal mix of SD7, BRA5, epiCRealism V3 ((c)epinikion), and Hassaku V12, judging from the number of downloads for trained ckpt on CIVITAI. Why trained? That's because the trained model is very important. So it's no problem if I forget to remove the search filter flag.
I could have used the latest version, but I chose a older version considering the possibility that it was used in various models in the merge. I think the balance between 2.5D, Photo (East girls), Photo (West girls), and 2D anime is not bad. It should reduce model bias.
m1=BRA5+SD7 (50%)
m2=epiC3+H12 (50%)
RM15 = m1 + m2 (50%)
I checked on 1girl to see what kind of pictures this standard model itself produces.
I thought it might still be biased because the output is more photo-oriented, but when I checked the similarity with KA2, I found that RM15 is closer than SD7.
Below is an example of Lora's output created using this interim reference model.
I thought it might still be biased because the output is more photo-oriented, but when I checked the similarity with KA2, I found that RM15(right) is closer than SD7(left).
Below is an example of Lora's output created using this interim reference model.
However, although the design is similar, it feels more like "not my girl" than the DS7.
This may be something that only the creator of the model can understand.
I felt this way with DS7 as well, but since the age groups corresponding to "1girl" are different, there are some pictures that won't appear if you don't specify the age. You probably know that the interpretation of age itself is different depending on the model.
Well, this could be compensated for with prompts or Lora, but here I decided to add 20% of CuteYukiMix (cuteyukimixAdorable_echodimension). I didn't have much material for anime pictures, so it wasn't a bad choice in terms of balance.
The following is the result of remaking the RMc15 that was created in this way.
This is considered to be acceptable for most images.
Q. Aren't the girls in 1girl a little too young?
A. Please deal with it by specifying the age or using the Age Slider.
Generally, adding "anime" to the prompt will make the girl look a little younger and have bigger eyes. So, you can try "anime:-1".
However, although the design is roughly the same, to be honest, the usability is slightly different. There is also some deterioration. I felt that the drawings had become hard enough that I decided to try EulerA. There are some small changes that have advantages and disadvantages, and cannot necessarily be called disadvantages.
Since stable diffusion is a system that is almost impossible to control as desired, I don't think there will be any problems if I continue to respond to the results as I have done in the past.
to be continued part 3...