@Ikena hah! thanks, totally missed it. keep on rockin it! this checkpoint is amazing. 👌

You can check danbooru tag list danbooru.donmai.us/wiki_pages/tag_groups
It show everything sorted by catergory!

@Rkkzne Thanks, I'll read them.

From what I can see., the images here are 512*1024 or 768*1024 I think

@Rkkzne Many thanks!

@Ikena thanks a lot! 

https://danbooru.donmai.us/wiki_pages/tag_groups

If you click on the little "i" you can see the prompt that generated the sample images.

I'm using It with 768×1024 and it produces very good images. So I think You can go to 768 or even 1024.

@Ikena Hi, i have --no-half-vae but I still get black images when I set vae to "none" in SD settings. Any ideas?

edit: vae-ft-mse-840000-ema-pruned vae works though when I switch it back to this one.

@Ikena Thanks a lot! Hats off to you.

名前がすごいことになってます。

@Ikena i dont know why but restarting fixed the issue but thanks

The most important thing is to use the :R-ESRGAN 4x+ Anime6B" upscaler option for illustration-type images, and R-ESRGAN 4x+ for realistic images. Never use "Restore faces", especially for illustration-type images.

Denoising strength is a matter of taste, since the higher the number, the further the result will be from what you're choosing to upscale, which could be good or bad, but above 0.4 you're basically getting a different image. I've never really played with Hires steps and I don't know the advantages or disadvantages of using more or fewer steps in upscaling than in the original sampling. I generally leave it blank, which means it defaults to the same number of steps, and it seems to be fine.

And be sure you're using the ADetailer extension, since that runs after the upscale, which means more attention paid to faces, including people in the background that would otherwise just be smudges. I think a number of anime checkpoints assume you'll be using ADetailer and put less effort into faces because of it, which is going to be a lot more obvious after the upscale.

You can ignore all the settings and keep the defaults, except for the choice of ADetailer model itself, which can provide slightly different takes on the same image. I find face_yolov8n is usually most reliable, but it can be worth checking with face_yolov8s and mediapipe_face_full for a second opinion if you don't like the first one. Again, it's a matter of taste. (I've personally never gotten good results from the hands or person models, which is a shame.)

I hope this helps!

@A1322 I find your comment quite helpful so thank you.

@A1322 For me I use about 7 for hires steps, and that's what I found to be the best balance between quality and time

Well the only logical explanation is that model is not trained enough on it or not at all + might also be related to deep throat prompts.

@Ikena how about the old preview images on the grapefruit V1 model? Like the girl wearing a kimono holding a grapefruit, do you still have that one?

@Ikena Or do you still remember the merge recipe for grapefruit V1?

@Ikena ah alright, was kinda missing the more 3D and desaturated look of your original V1 model haha

@FIFTTP I might find them. I'll let you know

@IAmSteve Tried finding it on huggingface, almost found it on web archive but the login screen prevented me from accessing the model :<

@IAmSteve It's alright lol, no need to waste your time finding a lost model for some stranger on the internet haha

@FIFTTP I found it on my google drive, feel free to download it (it's the 1.4 version though): V4 - MISREAD THE TITLE, SORRY

@Ikena Yeah you are right... my bad, I misread the version number. Sorry