I wish I could rate it 4.5, because it deserves more than 4, but I can't quite call it a 5. The main reason for that is that after trying both versions, neither of them worked well with my preferred model. Despite that, nearly all other LoRA that I have tried do work well with it. (Edit: I'm going to call it a 5 anyway. The problems I describe are just as likely to be with my model as with the lora itself.)

That said, I switched over to Anything 3.0 and made some gens that I think are some of the best I've ever made!

The first 3 swim brief gens are made with hires fix. All the remainders are done first at 512x768 and then put through img2img at 1024x1536. (Basically, a manual hires fix.) For the larger versions, the lines became more like rough sketches instead of a final version. So when you look at my examples, keep in mind that the 512x768 versions have more solid, "final version" lines. I don't know what happened to cause them to do that in the i2i, but I don't blame the LoRA for it.

So... Here are some of my favorite gens! I made three scenes to share here: Haru at the pool, Haru shopping for clothing in casual clothes (for some reason he looks like the middle school version of Haru, but that's okay too!), and Haru cooking Rin and Makoto some mackerel for breakfast after the three spent a wild night in bed. (lol) Hope you enjoy them!

Note that I also used Koolchh's Husbando LoCon v3.2 at a lower strength when generating some of these. (I did that hoping to get some penises, which you can tell from the prompt I used, and that particular LoCon almost always generates accurate ones.) I generated these using kohya-ss's additional networks plugin, not the built-in LoRA support. Using the built-in LoRA support may give you different results with the same generation details. I also used the Cutoff plugin, however I learned after the fact that it is not compatible with the kohya-ss plugin, so it was automatically disabled, even though the details for it do show up in the various generation details.

Furthermore, when making the img2img enlarged versions, I used the smaller (512x768) image as the base image for img2img but I also used the same image in Controlnet with OpenPose, just to ensure that SD would adhere to the same image as much as possible. However, just thinking about it, this may be the reason the lines turned out sketchy instead of smooth. Or maybe not. 🤷

All generation info can be found under the Negative Prompt, including denoising strength, etc.

The model I used for all of these is Anything 3.0. I personally don't use Anything 4.5. It's been a long time since I used Anything 3.0, but I'm happy with the results I've posted.

I'm happy to answer any questions.