The weird Renard
Hello everyone,
while working on my new “Renard (Redo of Healer)” version, I noticed something strange.
After generating around 100 images, I suddenly asked myself:
“Did Renard always have this deep blonde hair?”
To answer that immediately: No, he didn’t.
(And he never will, because this 💩 is ☠️.)
It was honestly a bit frustrating — after generating 100 images and even creating an extra model to possibly produce him/her without bandages… and now I have to start from scratch and find the culprit.
So I generated a picture with only hires enabled.
Then I used the same prompt with Adetailer enabled — a slight change in the eyes, expected and normal.
Turning on FreeU:
And suddenly the hair turns deep blonde. Gotcha 🎉!
But wait — that wasn’t the only issue.
When I compared this new image to the other 100, the new one looked less detailed.
But… in a good way?
Anime characters are usually fairly simple, and those 100 images actually looked better than the original Renard:
IMG: the "original" Renard from the anime
BUT the goal is to create a Renard model, not a subjectively prettier version of him/her 🤔
So I needed to figure out which extensions I should use, and which ones I should avoid.
Spoiler:
I ended up using only Hires and Adetailer 😅🤣
The other extensions are great — with two exceptions: CFG-Rescale Integrated and Detail Daemon (I’ll explain later).
But many of them do almost the same thing, and using all of them together can distort the trained model.
What the extensions did in my tests
Most of the time, I used extensions “out of the box” — meaning: I installed them and left everything on default settings.
I always did it like that (except for FreeU), and that’s probably the biggest problem here.
CFG-Rescale Integrated
This was actually the reason for an issue I’ve had for a long time:
sometimes my images had weird bright edges — and CFG-Rescale seems to be the cause.
That doesn’t mean the extension is useless — I just probably never configured it correctly.
Detail Daemon
It basically didn’t do anything for me — again, likely because I didn’t set it up right.
CD Tuner
This one I didn’t use on default, because I read a bit about it.
You have to set d1 to 2 and leave the rest.
And honestly: it worked really well —
but a bit too well.
Like I said earlier, I don’t want to create a “better” version of a character.
But CD Tuner kind of forces that 🙈
FreeU
For SDXL I had to change the settings.
Again, it works great, but it goes against my goal of making an accurate model.
It strengthens colors and increases depth.
SDXL
Also changes colors and details, and corrects small anatomical issues.
Again, great — but difficult to use when the goal is a model that stays very close to the original.
CFG-Fix
At first, this was the most promising extension: it improved the image without changing it too much.
It adjusts shadows, hairlines, body lines, adds small details, and slightly changes anatomy.
But unfortunately, it also causes errors or even reduces quality by introducing issues (for example, wrong nose positioning, random eyelashes, etc.) from time to time.
SelfAttentionGuidance
This changed anatomy and pose, and increased depth a little.
First, it’s redundant, and second, the changes are still too strong.
Conclusion
I realized that some of the extensions I usually recommend at the end of my model posts are actually redundant or unnecessary.😒
You probably need to test which extensions work best for your setup.
For pure image generation, I’d probably use SDXL Styles and CD Tuner — the first is a great all-rounder, and the latter is great for detail.
What are you experiencing with the above extension? I'm grateful for every tip I can get.
I’ll have to make a new non-bandage model of Renard for my Version 2.0 now. 😏
Bye bye!