@Chumple Ok! that sounds promising (especially given the context of what I know of your standards from your previous checkpoints...)
Now I have to try it out for myself; see if I agree ^_^

And thx for the explanation
(I think I need to learn French, you're not the only one on this site suddenly pulling out Frenchesian terminology, and anyway I'll definitely sound more like an Artiste (TM)

@Chumple A pleasure to be of service to Her Highness the Comtesse in any way.

As long as my part can be in English :P

I'm so sad.

At first, they said if I knew some basic English words, I could start drawing, and I was studying really hard. Now, French? What's next, Chinese?

@vibhisana Yeah, I feel for you.

I have it even worse, I was hoping if I become one of those devil-may-care Artiste types I wouldn't even need to know English...

(the biggest change of the "AI Art revolution" is that now artistry works thru language and you can't flunk basic schooling anymore) :D

@Chumple Hm. Got it, makes sense.

Why do you prefer EulerA when you're trying for changes?
(I guess I'd think yeah, it's more predictable so more reliably stick to the base image composition while editing it. Or maybe does it handles changes better? or needs less steps or smth?)

@Chumple Oh! very useful pointer, I really need to be doing my hires fixing like this too.

Do you have a preferred denoise level for when you like the image and are only trying to polish it without changes? I start at 0.40, but really prefer lower (0.25 thereabouts) to keep the original look. But I've always used EulerA, and anyway i'm a noob.

Here are my two cents. I saw you asked before how to maintain composition when upscaling, but I was busy testing something at the time (which failed), so I didn't reply.

First, it seems you're using the online generator, which is quite different from ComfyUI, which I use. So ultimately, this probably won't help at all.

Regarding the steps, after 30 steps, there's limited changes for details; after 40 steps, it's almost meaningless. But I've heard that dpm_2(a) works very well even up to 60 steps; you could give it a try.

In short, regarding composition, low resolution->high denoise. I usually use 0.75 denoise on 1.50x upscaling (832*1216->1248*1824), combined with a dpmpp series sampler, to obtain rich details and a clear, sharp output. Recently, I tried 1.75x (1456*2128) and 2x (1664*2432) upscaling and found that 0.60 denoise is the optimal choice.

There was a blurring issue at low denoise, possibly because I used different samplers when upscaling, but I'm currently happy with the settings. Believe it or not, in many cases, 0.75 denoise helpes to correct extra/missing fingers.

Let me give you an interesting example: See the long black thing above the white-haired girl with the red mask? That's actually a masked head. I forgot to use "solo" in the prompt, so the AI ​​drew two girls and three heads. I loved the composition so much that I first tried high denoise, then gradually lowered it, and finally removed the extra head when 0.50 denoise.

I don't use upscale models, and in my 20+ tests, they gave different results and were not better. I would like to hear your opinions.

Regarding the test I mentioned earlier, I tried replacing the 2.00x upscaling with 1.15x upscaling * 5 (2.011) combined with low deniose, hoping to obtain rich details while maintaining the composition, but it failed. The problem isn't that "if the distance between each step is short enough, the AI ​​can walk steadily," but rather that "if the total distance is long enough, the AI ​​will eventually fall." In short, the AI ​​doesn't know how to correctly draw at high resolution, and always ends up with extra limbs.

You mentioned five articles, which I'd love to read. I've always enjoyed manipulating different parameters; it's like gambling—you never know which combination will win you the lottery.

@vibhisana Thanks! your answer is packed with lots of useful info.

Despite the civit on-site generator having different settings than Comfy, I think a lot of it is relevant. I'll need to spend time experimenting before I can provide a more solid response.

You say you enjoy playing around with gen parameters it's like gambling,and yeah me too -- this type of experimentation is a major part of what I love with AI gen. But IMO it's much cooler than gambling, it's really like engineering/architecture, except the fractal nature of AI token parsing makes the engineering much less predictable. Infinite fun.

@ravemry9 Yes, it's way cooler than gambling. I was addicted to the Diablo series and Path of Exile for years, and those gear-farming games were essentially gambling. Last year, when I started drawing, it suddenly dawned on me that in games I killed/destroyed things for fun, while in drawing I created beautiful things. txt2img has truly made me a better person, and I've never looked back. 

Sometimes, less predictable is good, like this and this. In my prompt, the boy should hold flowers, front view, at middle of frame, girl at left side of frame. The result is totally irrelevant, but what can I say? They're happy! God of AI works in mysterious ways. AI understands what beauty is; I'm simply working with it.

@vibhisana Oh, well put about the fun of AI being in creation, and in beauty.

"The God of AI" , lol. It's definitely true AI has a mind of its own and some of our best results are unplanned.
I don't rmemeber which ones now, but at least twice i got amazing gens that didn't replicate when i realized i had misspelled important parts of the prompt. Fixing the typos took all the magic out of the image.

Those images you link are beautiful works of art. Hearing the story behind how you genned them is really interesting.

@ravemry9  , Let me show you another miracle created by AI. Once I wanted to draw one girl, play piano on beach, ocean, whale jumping in background. If all goes well, it sounds good, right? I cut and pasted some words to adjust the sequence, didn't realize I missed the beach and ocean parts, and ran the generator directly. This is the result. AI turned the impossible into beauty, I lack that creativity, that vision, and that taste. This is the 24th pics we create together, and since then, I've known I'm merely a humble explorer, searching for undiscovered beauty created by AI.

@Chumple  , I'm thinking about a lazy solution recently: just do 1.50x upscaling, then 4x upscaling with model (4992*7296), LoL.

@vibhisana I know I'm late to replying this comment... but hey I've seen your awesome whale-piano post and loved it; it's great to hear the backstory now :D

@Chumple wow those two are great examples, 2 very different ways AI gen mess-ups can make the results so magnificent.
(It really does work parallel to gambling, vibhisana's def right. But... somehow, when I find these happy accidents and then look at the rest of the creator's gallery, even the non-messups are good. Hmmm, I wonder what that means... ^_^)

@ravemry9  , I tried heun+beta recently, 5 steps is enough to "peek" the composition, it has much clearer output than euler, and usually more details. You should give it a try.

@vibhisana Will do. Thx.

@Chumple ngl one of the best IL checkpoint, its my main thx very much!

@Chumple whats the recommend resolution for base generating?

@Chumple like literally i like prompting long using 60 steps with 7cfg, i put 10 loras at the same time, and this model can still handle it pretty well, like i tried to generate same seed same lora chain but its just not as creative as this one does, looking forward in the future. Been comparing it to others such as wai illustrious, one0bession, amix from phinjo, etc... this one handle the details without losing the detail significantly and it does work with a bunch of loras which is few tuned models sometimes din't pop it out as much as this...