creator: windir19
Introduction
SakuraRain is a 50/50 merge of two of my favorite checkpoints: RainPonyXL and SakuraSushi XL. (Go try out both models, they are both fantastic) I wanted to get something that had the the coloring, good anatomy (fingers especially) and interesting backgrounds of RainPonyXL and the bigger more cutesy anime eyes of SakuraSushi XL. The intention of the checkpoint merge is anime character focused but from RainPonyXL there is the ability to generate interesting landscapes. During preliminary testing I got some promising results that I liked. I hope to keep testing out the merge and sharing my own creations here. I am also happy to see whatever you try to make!
Example images above did not use inpainting, ADetailer or are edited and pushed through img2img in anyway. That said, I did use Hires. fix on them. (Details on process below)
Use Guide
For suggestions on use, these are my preferences for a starting point if you are using Automatic1111/ForgeUI. Again, if there is a way you have been using SDXL/PonyXL checkpoints, you can experiment with your own settings.
Generation Settings
Size Resolution: I prefer 832X1216 for portrait or 1216X832 for landscape. r_ain uses 896X 1152 for portrait and 1152X768 for landscape. 1024X1024 is good for 1:1 since this is SDXL based checkpoint.
Sampling method: DPM++ 2M Karras, euler a or Restart. It really depends on my mood.
CFG: 7, but anywhere from 4-7 can work.
Sampling steps: 25-35. Usually 25, 28 and 30 are my lucky numbers. If you want to see which one yields the best result on a seed, try running it through X/Y/Z plot to see what looks best.
Positive Prompts
As this is PonyXL based positive prompts should start like this:
score_9, score_8_up, score_7_up, score_6_up, score_5_up
That said, I usually start like this for my positive prompts so that I am not wasting tokens:
score_9, score_8_up, score_7_up
I would also say that prompts for quality like "masterpiece" and "high quality" are not super useful since the scores are prompting for quality. When using language for describing what you want, PonyXL is trained on booru tags. I would refer to a list of booru tags or install something like SD WebUI Tag Autocomplete to assist. That said, it is still SDXL based so for example I have found that if you say "chinese village" as opposed to "east asian architecture", it can get the idea of what you are saying. You can be creative and see what works, but it's good to keep in mind this model as well as the models it was based on work best with booru tags.
Generally I start my postive prompt as:
score_9, score_8_up, score_7_up, source_anime, {prompt tags describing lo}
Negative Prompts
There is a lot of debate as to whether the lower score tags (5 and lower) should go into the negative prompt. I personally don't include them. In fact I try to keep the negative prompt as small as possible. This is inspired by how @FallenIncursio suggests using his checkpoints and it works for me in my prompts. For my base negative prompt I start with:
3d, monochrome, simple background
You can leave out "simple background" but I find this prompt alone does a lot to make the environment more interesting and is something that r_ain does with a lot of their prompts.
That said, negative prompts do have an impact but you shouldn't be starting by including what you don't want in an image like "bad hands, extra fingers, extra toes, etc." I would generate an image with a seed wit the positive prompts and if you see issues, try adding tags in the negative to get rid of those items. Sometimes not even adding a negative and adding or changing something in the positive prompt such as placement of keywords or adding weight to them can change the direction of your image as well. Therefore my suggestion is to add negatives iteratively.
Hires. Fix
This is a controversial opinion since some people say that you should upscale for details later in Img2Img but I never figured it out in a way that has worked out for me. (If you have a better way please tell me) What I like to do is generate a bunch of images without Hires. fix and look through the ones that have promise by going back to that seed. If needed I will try to make changes to the positive or negative prompts to add/remove keywords until I have something I will result in something promising. Then these are the settings I use:
Upscale by: 1.5x Any higher and you tend to get unexpected results like extra fingers, long torsos, ect.
Upscaler: for anime I use 4x_BooruGan_650k
Denoising strength: 0.5-0.6 I start at 0.6 and if I am getting added details I don't like I go down to 0.5. Sometimes I will try 0.7 and get success but at this point or higher you are going to get extra fingers, hands, toes as well as other things like navels shifting, etc. If you go below 0.5 you tend to not get better details.
Hires steps: 0
ADetailer
I would use ADetailer to fix details in the image. Here are my suggestions:
For faces use face_yolove8n.pt and add only the positive prompts that describe face features such as eye color, eye features, blush or anything that might be on the face like piercing are certain bodily fluids. I also find that adding the LoRA embed to the ADetailer positive prompt can help with coloring tone for the eyes for some characters.
For nipples I suggest getting the ADetailer Nipples model (Anime) for Adetailer to find nipples and use the Nipple LORA for ADetailer to fix them. Read on how to use the LoRA because you need to include the LoRA embed in the ADetailer positive prompt as well as the word "nipples". Also make sure you set the detection model confidence threshold to 0.7 other wise you will get nipples in weird places like elbows.
Using ADetailer for hands is pretty useless at this point. Ignore it.
Conclusion
I hope you enjoy this merge! I look forward to seeing what you post. Also I will be posting as many examples here that I can to showcase the possibilities below so be sure to check the images to get ideas for prompts. I would like to thank both @r_ain and @Random992 for letting me post the merge of their checkpoints as well as offering words of encouragement and advice. I would also like to thank everyone else that had a part in the genealogy of this merge.
