Follow-up

I noticed in your RInne LoRA that you're getting the same kind of blurriness, even though you're using a different base model (Hassaku instead of Counterfeit). So the problem isn't the base model.

Then I saw that you have one image in your Rinne gallery that is different. It could still use a bit of improvement, but overall it's good. Not nearly as blurry as the rest.

Long story short, you're setting your CFG and step count too high for the DPM++ 2M Karras sampler. 2M Karras was made to work well at lower step counts (approximately 10 to 30). It's also well known that setting a CFG too high will cause 2M Karras to produce images that appear overcooked or distorted in some other way. Setting it too high with any sampler will cause distortion, but the 2M Karras sampler seems to be especially sensitive to it.

CFG is basically like a multiplier for the entire prompt. When you use high emphasis in the prompt, like "(naughty smile:1.8)" along with a high CFG, it has a similar effect to using "(naughty smile:3.5)" with a low CFG. If you enjoy getting your music to sound just right, you can think of the tags within the prompt as the various different bands of an equalizer and the emphasis is the amount your raise or lower that band, and the CFG as a pre-amp: When you increase the CFG, you're boosting all the bands of the equalizer by the same amount.

I'd suggest either:

* Set your step count to around 20 to 25 and your CFG around 5 to start. Try your CFG as low as 2 and as high as 9. You could run an XYZ grid, with step count on one axis and CFG on the other to determine the sweet spot for that image. The CFG especially will vary depending on the prompt you use, so you'll need to do that more or less every time you try a new prompt. You'll get a good feel for it before long, though, and can make a fairly accurate guess, which will let you reduce the size of the XYZ grid and speed up the process.

* Use a different sampler. 2M Karras is a good general-purpose sampler, but it does degrade the image if you set those values too high. If you want to use a high step count and CFG, you'll want to use a sampler that can handle that better. You might try one of the ancestral samplers, for example, like DPM++ 2S a Karras or Euler a. Ancestral samplers will keep making changes to the image as long as you keep increasing the step count, with essentially no limit. Most other samplers will "converge": They settle down into a final image, and going above the number of steps it takes to get there will give you at most very small changes but take longer. "Diminishing returns," as they say.

You might want to try the UniPC sampler. It's the newest, and it was made to work in as few as 3 steps. You'll probably get somewhat better results with it around 5 to 7 steps, though. My system does not like UniPC, so most of the time I use DDIM, which is the fastest for me. I generally start around 15-20 steps and a CFG between 5 and 7, then make adjustments as necessary.

TL;DR

I think if you either change samplers or reduce your CFG and step count, your images will be a lot crisper and look better overall. You might have to play with it a bit to get exactly the image you want, since lowering the CFG is going to reduce the effect the prompt has on the final image, but despite that, I think you'll be pleased with the results.

I hope you find this message helpful, and please don't think of me as being bossy or anything. I just like to help people get better results from their AI generations when I am able to do so. Plus, I'm an Ensemble Stars fan, so I want to see the boys looking their best! 😉

Happy creating!

@AI_Kengkador I'm glad you feel that way about my comments. Yes, I believe if you can stay closer to 20 steps and cfg of 7 or lower, you'll get results that look sharper and more refined. The image composition may not look quite like you want it, but in that case, I'd try using a different sampler. With enough steps most samplers will give you roughly the same image.

So if you're finding you need 40 steps with 2M Karras to get the composition you want, you could try one of the other samplers, although you may need to move your step count up or down a bit. Just avoid the 'a' samplers because they will continue to change as you increase the step count, instead of settling down into a final image. So Euler, but not Euler a.

I apologize that I'm so wordy. I try to keep things short, but I just can't seem to do it. I usually just end up making it longer if I try to shorten it. 😅

@AI_Kengkador Glad to know that it was helpful for you!